This is f*cking insane. Google Brain founder Andrew Ng just dropped a free 3-hour course on becoming

Lunar
@LunarResearcher
This is f*cking insane. Google Brain founder Andrew Ng just dropped a free 3-hour course on becoming an AI engineer in 2026: "The most valuable skill this year is building with AI. Not chatting with it, building with it Agentic workflows will drive more progress than the next generation of models" In 3 hours, he goes from agentic systems to a full prompting course to shipping a working AI app in 30 minutes Prompts → Agents → Agentic Workflows → A Shipped App → An AI Engineer Job An agent is a model that plans, uses tools and checks its own work An agentic workflow chains those agents so the job gets done without you in the loop Everything else you need is free: Andrew Ng's Agentic AI course https://t.co/nOIgigtuj1 AI Python for beginners, taught by Andrew Ng https://t.co/On5VRAoZp7 Anthropic's interactive prompt engineering tutorial https://t.co/ZSutQXtCFN Anthropic's courses: from your first API call to tool use https://t.co/tOWxobot4O Karpathy's Neural Networks: Zero to Hero https://t.co/t3dSfJOd7o Practical deep learning for coders https://t.co/Qw9yNITGyx Microsoft's 21 lessons on building with generative AI https://t.co/RL7ezz2rnT Google's machine learning crash course https://t.co/iHup6JKWqn Hugging Face LLM course, start to finish https://t.co/rCGKypWysT The OpenAI cookbook: code for every common task https://t.co/qwpWOjiNsd Hundreds of AI apps you can copy and ship https://t.co/mXIchov5Wp The prompting guide everyone shares https://t.co/WfUzcZ16rM Total: $0 95% will keep paying for bootcamps that teach last year's stack. The 5% who start with this course will ship their first AI app this month Bookmark this and send it to someone who wants into AI this year Then read how to wrap that first app in a harness that runs on its own in the guide below ↓
Transcript
- Is one of the most
- 0:01
- valuable skills in AI today
- 0:03
- In this course, I share best practices for doing so
- 0:06
- and you come away
- 0:07
- with a systematic understanding of how to build
- 0:10
- agentic systems
- 0:11
- I coined the term
- 0:12
- agentic AI to describe a growing category of AI projects
- 0:16
- not realizing that
- 0:17
- marketers would get a hold of that term
- 0:19
- and slap us a sticker
- 0:21
- on everything in sight
- 0:22
- Even though the hype on agentic AI has grown really rapidly
- 0:26
- the reality of agentic systems being created
- 0:29
- and deployed has grown rapidly as well
- 0:32
- Unlike a single LLM call
- 0:34
- they just directly generate a response
- 0:36
- Agentic AI systems can execute multiple LLM-driven steps
- 0:41
- These steps can use tools
- 0:43
- they can reason
- 0:44
- they can iterate over the work
- 0:46
- to complete complex tasks. In this course
- 0:48
- you learn how to build agentic workflows
- 0:51
- using raw Python without hiding all the steps in some framework
- 0:55
- so that you see how each step
- 0:58
- actually works. You learn to take a
- 1:00
- Complex application
- 1:01
- and decompose it into a sequence of tasks
- 1:03
- that you can then implement
- 1:05
- as an agentic workflow. We also learned
- 1:07
- how to implement the four key design patterns
- 1:10
- for agentic workflows
- 1:11
- Having worked on many agentic systems with many teams
- 1:14
- I found that the single biggest predictor
- 1:17
- for whether the team
- 1:18
- executes well lies in its ability
- 1:20
- to drive a disciplined error analysis process
- 1:24
- That is
- 1:25
- to put in place evaluations or evals
- 1:28
- so that given a complex agentic workflow
- 1:30
- you can efficiently hone in
- 1:32
- on what components to focus attention on improving
- 1:35
- This way
- 1:35
- you aren't just guessing how to spend time
- 1:38
- productively. You let the evals data guide you
- 1:41
- In this course
- 1:41
- I'll also show you how to drive this process
- 1:44
- and this will put you
- 1:45
- significantly ahead of a game compared to
- 1:48
- the vast majority of teams
- 1:50
- building agentic workflows today. This course is taught
- 1:53
- in a vendor-neutral way in raw Python
- 1:56
- And the emphasis is on teaching you
- 1:58
- the core concepts that you can then implement
- 2:01
- using any
- 2:02
- of the popular agentic AI development frameworks
- 2:05
- or
- 2:05
- using no framework
- 2:06
- This law of phone course is currently available
- 2:09
- only on the deeplearning.ai website
- 2:12
- I hope you sign up for this course
- 2:14
- and when you've completed it
- 2:16
- you know how to build agentic AI systems
- 2:18
- which is one of the most
- 2:20
- in-demand skills on the job market today
- 2:22
- and that will also let you build
- 2:24
- a lot of cool applications
- 2:33
- Our focus introduces this course on agent skills
- 2:35
- built in partnership with Anthropic and taught by
- 2:38
- Andy Shopik
- 2:39
- Skills are a new way to equip cloud
- 2:41
- code and other agents with on-demand knowledge
- 2:44
- In this course, you learn how they work
- 2:46
- explore best practices for creating them
- 2:48
- and build skills for coding
- 2:50
- research, data analysis, and more
- 2:53
- You
- 2:54
- specify a skill by creating or downloading a folder containing instructions
- 2:58
- scripts
- 2:58
- and references for carrying out a particular task
- 3:02
- Your
- 3:02
- agent then dynamically loads the content of the
- 3:04
- right skill to perform that task reliably
- 3:08
- For example
- 3:09
- maybe one skill will contain a checklist of
- 3:11
- code quality criteria for reviewing your code base
- 3:14
- or another skill might include instructions for writing
- 3:17
- SQL queries against your team's data infrastructure
- 3:20
- and yet another might describe a repeatable workflow
- 3:23
- for analyzing customer data
- 3:25
- You create these skills once
- 3:28
- and make them available to any skill-compatible agent
- 3:31
- which then invokes them when needed to access
- 3:33
- that domain expertise
- 3:34
- Skills were originally developed by Anthropic
- 3:37
- and are now an open standard adopted by
- 3:39
- a growing number of agentic applications
- 3:41
- In this course, you'll see skills in action across Cloud AI
- 3:45
- Cloud Code, the Cloud API
- 3:47
- and the Cloud Agent SDK
- 3:49
- Because the format is standardized
- 3:50
- you can use the same skills with any
- 3:52
- other compatible agent
- 3:54
- and compose them to create complex workflows
- 3:56
- You'll learn how to combine skills with other
- 3:58
- parts of your agent stack
- 4:00
- such as MCP, subagents
- 4:02
- and use them together
- 4:03
- to build powerful agentic systems
- 4:05
- As OMs become more capable
- 4:07
- the code you need to write to train
- 4:09
- an agent is becoming simpler
- 4:11
- Basic tooling like file system access and batch
- 4:14
- for code execution
- 4:16
- enables the agents to
- 4:17
- take a wide range of actions
- 4:19
- Skills
- 4:20
- further give agents the specialized knowledge they need
- 4:22
- to handle specialized workflows accurately
- 4:25
- I hope. See all of you here
- 4:27
- In addition to the air prompting for everyone calls announcement
- 4:31
- there's one other
- 4:33
- significant announcement I won't make in this presentation today
- 4:36
- and you'll be the first people in the
- 4:37
- world to hear about this
- 4:39
- and
- 4:40
- I'll do that toward the end of this
- 4:41
- presentation
- 4:43
- What I want to do today is share
- 4:44
- with you what I'm seeing with the future
- 4:46
- of software engineering as it relates to AI coding
- 4:50
- and touch on Context Hub
- 4:52
- and a new thing that I'll announce a little bit
- 4:54
- to share with you more
- 4:55
- I think
- 4:55
- how we can hopefully help developers
- 4:58
- navigate to this future
- 5:01
- A year ago, at the first AI Dev
- 5:04
- I presented a slide like this
- 5:07
- where I talked about thinking of software
- 5:10
- as assembling many building blocks to create these
- 5:12
- wonderful things
- 5:14
- So it turns out that I often think of software
- 5:18
- tools, frameworks, APIs, as akin to Lego bricks
- 5:23
- And if all I have is
- 5:25
- one type of white colored Lego brick
- 5:27
- you can build some stuff
- 5:29
- but not that interesting
- 5:31
- If you mix in some black Lego bricks
- 5:33
- you can build more interesting things
- 5:35
- Mix in some blue ones
- 5:37
- mix in some red and yellow ones
- 5:40
- and so on
- 5:41
- And as the number of discrete building blocks you have grows
- 5:44
- the way you can combine them to create
- 5:46
- interesting software grows exponentially or grows combinatorially
- 5:50
- Over the last year
- 5:51
- the rise of
- 5:52
- AI coding agents has made it
- 5:55
- far more effective for all of us
- 5:57
- to quickly put these building blocks together
- 5:59
- to create amazingly cool stuff
- 6:02
- And this includes both AI building
- 6:04
- Like large language models you can prompt
- 6:06
- rag, agentic workflows, and many more
- 6:09
- as well as non-AI building blocks like UI components
- 6:12
- persistence of databases, identity of auth managers
- 6:16
- and so on
- 6:17
- And it turns out
- 6:18
- because of AI coding
- 6:20
- these building blocks are proliferating at a speed
- 6:23
- like we've never seen. There are more and
- 6:24
- more building blocks every single day
- 6:26
- both AI and non-AI
- 6:28
- for all of us to use
- 6:30
- And this is wonderful
- 6:31
- especially when AI coding lets us
- 6:35
- get AI agents to assemble these building blocks for us
- 6:39
- Now, when I speak with different teams in startups
- 6:43
- academia, also very large companies
- 6:45
- one question that's often discussed and debated is
- 6:48
- how much should we use AI coding agents?
- 6:51
- And there's actually one popped up poll question
- 6:52
- which is, I'm not going to ask this question
- 6:54
- but I see many people ask
- 6:56
- how much of your coding is AI versus human?
- 6:59
- And, you know, there'll be some teams that'll say
- 7:02
- oh yes
- 7:02
- my team does 80% of our coding via AI
- 7:06
- And, you know, that's pretty good
- 7:09
- my own coding is not 80% AI
- 7:12
- it's actually pretty much 100% AI
- 7:14
- And I want to share with you why
- 7:16
- I think this even last gap makes a big difference
- 7:20
- Oh, by the way
- 7:21
- in case there's anyone here that's 0% AI
- 7:23
- it's okay too, but I encourage you to increase that percentage
- 7:26
- right?
- 7:28
- but when I see among teams as if
- 7:30
- 80% of your code is AI
- 7:32
- then in terms of time spent
- 7:34
- it actually looks like this
- 7:36
- right?
- 7:37
- AI coding time
- 7:38
- human to review that 20% of code just
- 7:40
- takes forever
- 7:41
- Whereas a team that does 100% AI coding
- 7:44
- just gets a lot more done in a short time
- 7:47
- So, I find that this is true
- 7:49
- not just for writing code
- 7:51
- but also
- 7:52
- reviewing code. I know that a lot of
- 7:54
- teams also have AI write code
- 7:55
- but have a human review every single line
- 7:57
- of code
- 7:58
- But for myself and a lot of my teams
- 7:59
- we find that if I got to review the code
- 8:02
- then I become the bottleneck
- 8:03
- and it also doesn't work
- 8:06
- And I'm not saying
- 8:07
- you know, if you're, if you're working for NASA
- 8:09
- building
- 8:10
- and you need to write 50 lines of
- 8:11
- code for a spaceship
- 8:13
- go ahead, write it by hand
- 8:14
- it's totally fine. So
- 8:16
- this is not a religion where we should
- 8:17
- never be allowed to write code by hand
- 8:19
- But what I'm seeing is
- 8:21
- many of the frontier teams are trending toward 100%
- 8:25
- almost 100% of code written by AI
- 8:28
- and this creates a real acceleration that getting
- 8:30
- only 80% of the way there feels very different from
- 8:34
- both for writing and reviewing code
- 8:37
- Now, As coding a series about AI
- 8:40
- there are fascinating psychological effects
- 8:43
- I think, last July in the batch
- 8:46
- newsletter, I wrote about the product management bottleneck
- 8:50
- And it was this observation that
- 8:53
- when I'm trying to build stuff
- 8:55
- you know, we'll often have an idea
- 8:58
- build a prototype, the software engineering
- 9:00
- and then go to users to get feedback
- 9:02
- that's PM or product management
- 9:04
- where
- 9:05
- use those ideas to come and refine the
- 9:06
- software
- 9:08
- So AI coding has made that 10 or
- 9:10
- 100 times faster
- 9:12
- And so this means that
- 9:14
- deciding what to build or the product management
- 9:16
- work
- 9:16
- becomes the new bottleneck rather the actual building
- 9:20
- So when I start to see
- 9:22
- like starting a year
- 9:23
- year and a half ago
- 9:24
- was this very strange trend where
- 9:26
- we used to have product manager to engineer ratios of
- 9:30
- you know, 1 PM would keep eight engineers busy
- 9:32
- right?
- 9:33
- These ratios take a grain of salt
- 9:34
- but 1 to 8 PM to engineer
- 9:37
- 1 to 7 common in many companies
- 9:39
- But the ratio start to trend toward
- 9:41
- you know, 1 to 2 was really weird
- 9:43
- and then 1 to 1
- 9:44
- which is even more weird
- 9:46
- And what I'm seeing is rather than 1 engineer
- 9:50
- 1 PM
- 9:51
- the only thing that can move even faster
- 9:53
- is you take those two people
- 9:54
- and collapse them into a single human
- 9:57
- And I find that engineers can shape products or product managers
- 10:01
- they can code, can move really fast
- 10:03
- And that speed is a huge advantage
- 10:05
- But
- 10:06
- not everything can be done by a single
- 10:08
- person team
- 10:09
- And so
- 10:09
- I find that for more and more AI
- 10:11
- native The trend is a very small teams
- 10:14
- They're all kind of generalists
- 10:15
- where the engineers can do some TM work
- 10:17
- the PMs can do some engineering work
- 10:19
- and everyone kind of knows almost everything
- 10:22
- and the small team can move really fast
- 10:26
- I know that in some Silicon Valley circles
- 10:28
- it's trendy to talk about the job apocalypse
- 10:31
- this idea that AI will take all jobs
- 10:33
- almost all jobs, and there'll be writing industries
- 10:36
- Frankly, I'm not really seeing that
- 10:38
- And in fact, with this profile of engineer
- 10:42
- the future AI engineer
- 10:44
- has a very bright future
- 10:45
- When my teams are hiring for AI engineers
- 10:49
- these are what we look for
- 10:50
- right? First, have to be able to use coding agents effectively
- 10:53
- be it Cloud Code
- 10:54
- Gemini CLI, OpenCode, or something else
- 10:59
- And
- 11:00
- I find that
- 11:01
- having a robust knowledge of the amazing building blocks out there
- 11:05
- which is a challenge because there's so many of them
- 11:07
- so many, and they change so quickly
- 11:09
- That is a challenge
- 11:10
- but knowing these building blocks lets you assemble
- 11:12
- things very quickly
- 11:13
- I'll come back to the building blocks a
- 11:15
- few times in this presentation
- 11:17
- And then lastly, I find that people with general skills
- 11:20
- be it basic product management or other things as well
- 11:22
- can build much faster. So there's a combination
- 11:25
- of skills on the ability to how to build things
- 11:28
- as well as the ability to know what to build
- 11:31
- and I'm seeing massive
- 11:35
- unmet demand for a lot more engineers with
- 11:37
- anywhere near this mix of skill
- 11:39
- In fact, even my teams
- 11:40
- we can't find enough of these people
- 11:44
- And this is why and
- 11:47
- and maybe going beyond the product management bottleneck
- 11:50
- I want to take that one step further
- 11:52
- which is
- 11:53
- when software engineering speeds up 10 or 100 times X
- 11:56
- everything else seems slow in comparison
- 11:58
- So in addition to the product management bottleneck
- 12:01
- I'm seeing that, there's often a design bottleneck
- 12:04
- because you design something
- 12:05
- and she implements it
- 12:06
- or even better
- 12:07
- maybe the designer just implements it in code
- 12:09
- without going through a design tool like
- 12:11
- you know, Figma
- 12:12
- and that even speeds us up
- 12:14
- And one thing that used to be painful
- 12:16
- but it's even more painful now
- 12:18
- is the legal compliance bottleneck
- 12:20
- right?
- 12:21
- If you spend
- 12:21
- three months writing code and legal takes a
- 12:24
- month to sign off
- 12:25
- it's okay, but you take a day writing code
- 12:28
- you got to wait a month for legal
- 12:30
- it's like, boy
- 12:32
- and even marketing bottleneck
- 12:33
- for many of our teams
- 12:34
- we write code and ship products so fast
- 12:36
- that marketing has a hard time scrambling to
- 12:39
- keep up with what these engineers are doing
- 12:40
- to figure out how to tell people about
- 12:42
- it
- 12:42
- So I'm seeing design bottlenecks
- 12:44
- legal compliance bottlenecks
- 12:46
- bottlenecks, sales bottlenecks, a whole
- 12:47
- whole bunch of other things
- 12:49
- And this is why
- 12:51
- I find that increasingly small teams of people that includes engineers
- 12:55
- but also generalists can move really fast
- 12:58
- And a piece of good news is
- 12:59
- it turns out that an engineer that's not a marketer
- 13:03
- but that knows how to prompt AI well
- 13:06
- Maybe they're not an amazing marketer
- 13:08
- AI actually makes really weird marketing decisions
- 13:10
- But, you know
- 13:11
- they could do some basic marketing work
- 13:14
- And in fact, if a team needs software
- 13:16
- product design, legal, and marketing, and it's a team of two
- 13:21
- by definition, these two people have to have some skills in
- 13:24
- in all of these functional areas
- 13:26
- And I find that smart AI native teams
- 13:29
- they can use AI to help with some
- 13:30
- of these other functions can move really fast
- 13:33
- not advising anyone to be a lawyer
- 13:35
- but, but frankly, I, I, I often have AI
- 13:38
- you know, deal with legal stuff as a first job
- 13:41
- and then take it to a real lawyer
- 13:42
- to sign off before I
- 13:43
- right, before I, launch something
- 13:45
- But I find this AI native team
- 13:47
- moves really fast. Something I don't have time
- 13:49
- to talk about is
- 13:50
- how to structure multiple AI native teams to work together
- 13:53
- because again, not everything can be done by a 10-person team
- 13:56
- but if you have a 100-person team of
- 13:57
- many AI native teams with limited communication and clear API boundaries
- 14:01
- that also helps to scale beyond a single team
- 14:05
- Well
- 14:05
- when it comes back to the work that
- 14:06
- needs to be done to create software
- 14:08
- writing software code is just a small piece
- 14:10
- of it that's accelerating rapidly
- 14:13
- but
- 14:13
- I just don't see the AI job apocalypse
- 14:16
- happening anytime soon
- 14:18
- I find that the business media
- 14:21
- often has to be more accurate than the popular press
- 14:23
- because the popular press can tell exciting stories
- 14:26
- but the business media is paid to get it right
- 14:30
- and I've been encouraged that over the last month or two
- 14:33
- a lot of the business media
- 14:34
- has been getting the story right
- 14:36
- that this AI job apocalypse now just missing
- 14:39
- text never calls job apocalypse
- 14:41
- job apocalypse being delayed
- 14:42
- Federal Reserve Bank, a small highlight
- 14:44
- sorry if I can't read it
- 14:46
- highlighted few see a need to reduce their
- 14:48
- workforce or study by the Federal Reserve Bank of Philadelphia
- 14:51
- So, actually, and, and to be clear
- 14:54
- I know a lot of people are feeling
- 14:55
- job insecurity at all levels of seniority
- 14:59
- I think we need to do something to address that
- 15:01
- and I know many people have been laid
- 15:03
- off because of overhiring. The pandemic
- 15:06
- or hiring from zero interest rates period
- 15:08
- So I'm not saying the job situation is
- 15:10
- is perfect
- 15:11
- but there's a lot of it that's not that pretty
- 15:13
- but the job apocalypse is possible to talk
- 15:15
- in some circles. I just don't see it
- 15:17
- happening
- 15:18
- which is why I'm
- 15:18
- eager to keep on investing
- 15:20
- in all of us becoming better at building these cities
- 15:24
- Now, one of the hardest things to do
- 15:27
- exciting, most important
- 15:28
- hardest is as we're becoming better and built up builders
- 15:31
- each of us
- 15:32
- how to master these building blocks
- 15:35
- I want to share with you a framework
- 15:36
- of parallel skill development
- 15:39
- Here's what I mean
- 15:40
- We talk a lot about how our coding
- 15:42
- agents are becoming more capable
- 15:44
- and this includes agent skills
- 15:46
- like the anthropic agent skills as a word
- 15:48
- but more generally
- 15:49
- our coding agents are just getting better and
- 15:51
- better
- 15:52
- And then
- 15:53
- There's the
- 15:54
- education or training world that I've been a
- 15:55
- part of for a long time
- 15:57
- that focuses on building human skills
- 16:00
- But looking to the future
- 16:01
- I think we increasingly need
- 16:03
- parallel skill development where as our coding agents become more capable
- 16:09
- our people
- 16:10
- we as people need the complementary skills to
- 16:12
- help us to drive
- 16:13
- the coding agents in the appropriate way so
- 16:16
- that we can partner together
- 16:17
- to build new things
- 16:19
- So
- 16:19
- I want to briefly touch on two things
- 16:21
- in the rest of this presentation
- 16:22
- One is content hub
- 16:23
- which I talked about before
- 16:24
- but I'm just going to a couple of
- 16:25
- minutes on
- 16:26
- And the second
- 16:26
- which is the announcement I made toward the end
- 16:28
- is CodeDream
- 16:30
- So
- 16:31
- content hub. It turns out
- 16:33
- building blocks are proliferating so fast
- 16:35
- that most of our coding agents
- 16:37
- which have a knowledge cutoff date
- 16:39
- you know, far in the distant past
- 16:41
- sometimes a year old
- 16:41
- sometimes many months old
- 16:43
- will often hallucinate or use deprecated APIs or
- 16:45
- not know about
- 16:46
- the latest tools we want to use
- 16:48
- So it turns out even today
- 16:50
- if you ask Cloud Code
- 16:52
- to call the open API
- 16:53
- it will use the chat completions API
- 16:56
- which is older, deprecated, and call older models
- 16:59
- because most of the training data on the
- 17:00
- internet uses this older API
- 17:02
- even though the newer API has been out
- 17:04
- for a long time
- 17:06
- Context Hub is a tool that some colleagues
- 17:09
- Rohit Prasad, Senior Hotel, and I built
- 17:12
- in order to give your AI agent most up-to-date documentation
- 17:17
- and it turns out that
- 17:19
- if you run Context Hub with
- 17:23
- Cloud Code
- 17:24
- it will
- 17:24
- generate code with the newer responses API
- 17:28
- maybe just to show you really briefly
- 17:30
- I don't know, Chubb search OpenAI, so this returns, so these are packages
- 17:37
- and then Chubb get
- 17:39
- this is actually OpenAI slash chat
- 17:43
- type this some more
- 17:45
- and so this will fetch
- 17:46
- the latest documentation on OpenAI that knows about
- 17:54
- Right, knows about GPT-5.5, has the, knows about the responses API
- 17:58
- and so on. And this is a tool
- 17:59
- that you are welcome to use
- 18:01
- but it's actually, you know
- 18:03
- built for your coding agent to use
- 18:05
- and so this generates
- 18:08
- like, okay, a 600-line markdown file
- 18:10
- right, kind of long for a human to read
- 18:12
- or you're welcome to read it
- 18:13
- to give your OM the context it needs
- 18:15
- to make this and other API calls correctly
- 18:18
- So I, I pray many of you too
- 18:19
- right, I've had my coding agents hallucinate API calls
- 18:23
- definitely, you know, open AI, Gemini calls
- 18:26
- make bad calls to various database services
- 18:29
- There's just so many tools our coding agents don't know about
- 18:31
- but feeding in this latest context
- 18:33
- I think, can be very helpful
- 18:35
- But beyond that
- 18:36
- I want to share with you
- 18:38
- A project they've been working on to help
- 18:40
- with human skill development
- 18:43
- for a long time
- 18:44
- you know
- 18:44
- my team's been working on a variety of online courses
- 18:47
- and please keep on taking online courses
- 18:49
- I think an important part
- 18:50
- of how we can all
- 18:52
- stay up to date with the building blocks
- 18:53
- and other tools
- 18:54
- And we've been exploring one other alternative experience that
- 18:59
- I'm announcing here for the first time today
- 19:02
- and that's the project we've been working on
- 19:04
- called CodeDream
- 19:07
- the Wi-Fi here is not great
- 19:09
- so
- 19:09
- I'm going to see if a live demo
- 19:11
- works
- 19:12
- It well may not
- 19:13
- but let's see if this
- 19:16
- Okay
- 19:17
- Welcome to CodeDream
- 19:19
- This is not a course
- 19:20
- It's a conversation where I want to show
- 19:22
- you
- 19:22
- how to use a coding agent together with
- 19:24
- two called Context Hub
- 19:26
- to write code using up-to-date API documentation
- 19:29
- I say this is a conversation because I'd
- 19:31
- like you to imagine
- 19:32
- we're on a billion heart of
- 19:33
- And
- 19:34
- you can interrupt me at any time
- 19:35
- using that little microphone button down below
- 19:38
- The code duty has created a voice agent
- 19:40
- that has information needs to try to answer
- 19:42
- your questions as it will
- 19:44
- So, if you do that strongly
- 19:46
- if there's anything let's chat about
- 19:48
- or if you like
- 19:49
- you can just lean back
- 19:50
- and watch the presentation
- 19:52
- Here's the idea, right?
- 19:55
- What are we going to
- 19:56
- learn about in this conversation?
- 19:59
- We're going to learn how to use context
- 20:01
- help
- 20:02
- to keep coding agents up to date
- 20:04
- with the latest information
- 20:06
- Then, we'll build a joke generator app together
- 20:10
- So, this is not a slide
- 20:11
- right? I just click into this
- 20:12
- I can type, let's
- 20:13
- let's tell a dad joke about bad Wi-Fi
- 20:17
- Why did the computer show later where it
- 20:19
- was bad connection?
- 20:21
- Not bad, could be worse
- 20:23
- And the other thing about this being JavaScript is
- 20:27
- you can also copy paste from these slides
- 20:30
- So for example
- 20:31
- this is a slide with instructions on
- 20:33
- how to install ContextHub and how to prompt
- 20:36
- OpenCode and so on
- 20:37
- So actually copy and hitting Command C
- 20:41
- and then go into the CodeDream build environment
- 20:44
- And, this is a
- 20:46
- a terminal in my browser where I can
- 20:50
- you know, paste that command
- 20:51
- I just copy paste it from my slide
- 20:53
- to run npm install
- 20:55
- chub, and then, I can do things like start up OpenCode
- 21:00
- which is open source version of a coding agent
- 21:02
- kind of like, you know, alternative to Cloud Code
- 21:04
- GenICLI, OpenIC Codex, and then, I don't know, writes, Python, whatever
- 21:13
- Right, and so this is a
- 21:15
- learning practice environment that
- 21:18
- lets you practice
- 21:19
- using modern coding agents with tools like Context
- 21:22
- Hub
- 21:25
- and so
- 21:27
- we have been working
- 21:29
- on this CodeDream build environment for quite some time
- 21:33
- and, I'm excited about exploring this concept where in
- 21:38
- in lieu of taking a course
- 21:40
- you can come on to
- 21:42
- a video conversation and have a conversation with
- 21:45
- me or an AI version of me
- 21:47
- to try to
- 21:48
- gain use building blocks and
- 21:49
- code in this more modern AI coding way
- 21:53
- And I am
- 21:55
- happy to announce
- 21:56
- that CodeDream is available as of
- 21:59
- today in preview
- 22:10
- So please check it out
- 22:11
- and the first
- 22:12
- experience will
- 22:13
- show you how to use context up and
- 22:15
- give us feedback. We're eager
- 22:17
- to keep on working to make this better
- 22:21
- So just to wrap up
- 22:25
- coding agents are allowing us to assemble building
- 22:28
- blocks quickly to create software
- 22:31
- and that acceleration is real
- 22:33
- but
- 22:34
- many of the downstream implications about where the new bottlenecks are
- 22:38
- are still being worked out
- 22:40
- and I'm eager to hope our entire community
- 22:42
- keep moving forward because the job openings are
- 22:44
- there
- 22:45
- The job populace is not being nearly as
- 22:46
- bad as people are hyping it up to be
- 22:49
- and I think it's important that we keep
- 22:50
- on learning
- 22:51
- these building blocks and learning these skills
- 22:54
- And for parallel skill development
- 22:56
- coding agents are Working well
- 22:58
- let's give them access to the latest building blocks
- 23:01
- given the right context
- 23:03
- context help will help
- 23:05
- And for human skill development
- 23:07
- I'm excited also to share with you the cogent conversations
- 23:11
- and all of you in this room are
- 23:13
- the very first people in the world
- 23:15
- that I've shared this with
- 23:16
- outside, outside our team
- 23:18
- and I hope you will check this out
- 23:20
- and that'll help you to
- 23:21
- master
- 23:22
- some new building blocks and master some new
- 23:24
- skills
- 23:25
- So with that, really excited to be here
- 23:29
- Thank you for all for coming
- 23:30
- and thank you all very much
- 23:40
- 22. Using AI well is
- 23:43
- one of the most impactful skills you can develop
- 23:46
- and
- 23:46
- people that are not yet at the cutting
- 23:48
- edge of AI usage often run into
- 23:51
- AI generating frustrating outputs
- 23:54
- I want to make sure
- 23:55
- you're an expert prompter
- 23:56
- and can take advantage of today's AI tools
- 23:59
- which are
- 24:00
- much more powerful
- 24:01
- than they were even a year ago
- 24:03
- Let's take a look at two different experiences
- 24:06
- the AI novice and the AI power user
- 24:10
- Many
- 24:11
- AI experts
- 24:12
- have learned to use it to answer hard questions
- 24:16
- In contrast, many people, including AI novices
- 24:19
- may have gotten used to using AI for simple questions
- 24:23
- as if you were
- 24:24
- prompting it like a Google search
- 24:26
- So you ask it
- 24:27
- does Taco Bell
- 24:28
- still have the Double Decker Taco?
- 24:30
- And maybe you get an answer like that
- 24:34
- which is fine
- 24:35
- But if you have much harder questions
- 24:39
- you can also
- 24:40
- ask it of the AI
- 24:42
- and give it time to think
- 24:44
- For example, if you're looking to buy a car
- 24:47
- you can upload to most of the commercial services like ChatGPT
- 24:51
- Gemini, Anthropix Cloud, or others
- 24:54
- a set of documents including car specs
- 24:57
- quotes, insurance plans
- 24:59
- and ask it what are the trade-offs for
- 25:00
- these different cars I'm thinking about
- 25:03
- and tell it to
- 25:04
- read everything and to think hard before answering
- 25:08
- And this can cause AI to spend
- 25:10
- many seconds or even minutes to think
- 25:13
- and then compile a detailed report for you
- 25:16
- I find this a huge time saver for
- 25:17
- a lot of things I have to do
- 25:20
- Another example
- 25:21
- AI-powered users
- 25:23
- have learned to provide the right
- 25:25
- context or the right background information to the
- 25:27
- AI
- 25:28
- to set it up for successfully answering your question
- 25:31
- In contrast
- 25:32
- I see some AI novices
- 25:35
- use a short prompt
- 25:36
- and hope the AI will
- 25:38
- fill in the blanks
- 25:39
- But if you think of AI as
- 25:41
- maybe being akin to a really smart
- 25:44
- fresh college grad, highly motivated
- 25:47
- but that doesn't really know that much about you yet
- 25:50
- then a short prompt
- 25:51
- sometimes doesn't give it enough information or enough
- 25:53
- background context to answer your question accurately
- 25:57
- So if you tell AI
- 25:59
- please write a good self-review to send to my boss
- 26:02
- the AI
- 26:03
- doesn't know what you've actually done over the
- 26:05
- last year because you haven't told it yet
- 26:08
- and it might write a very generic self-review
- 26:10
- which isn't that helpful
- 26:13
- In contrast
- 26:14
- I find that AI-powered users almost have empathy For the AI
- 26:18
- I don't want to overly anthropomorphize the AI
- 26:21
- but if you could put yourself in the
- 26:23
- shoes of someone getting a set of instructions from you
- 26:27
- you can ask yourself
- 26:28
- will they actually know enough about you
- 26:30
- to do a good job on the task
- 26:31
- you're assigning them?
- 26:33
- So an AI power user in comparison
- 26:35
- might upload a lot of information to the AI
- 26:38
- maybe give it a screenshot of a project
- 26:40
- tracker showing what you've worked on
- 26:42
- recent project docs
- 26:44
- maybe voice memo notes where you talk through the projects
- 26:47
- and then tell it
- 26:48
- to write a self-review
- 26:50
- to send to my boss
- 26:52
- And that could
- 26:53
- do a much better job capturing
- 26:55
- what you're most proud of
- 26:58
- One of the things
- 26:58
- power users have learned to do is
- 27:01
- how to prompt AI to get honest feedback
- 27:05
- A big problem with AI is it
- 27:07
- often wants to please you
- 27:09
- In fact
- 27:10
- many AI systems were trained
- 27:12
- to try to make their users happy
- 27:14
- And if you ask it a biased question
- 27:18
- it will often give a biased answer
- 27:20
- because it's trying to tell you what it
- 27:22
- thinks you want to hear
- 27:24
- For example, if you say, I have a great business idea
- 27:28
- mobile tie-dyeing, critique it
- 27:31
- Because you called it
- 27:33
- a great business idea and you're saying it's your idea
- 27:36
- the AI will naturally want to please you and say
- 27:39
- what a great idea
- 27:42
- We sometimes call this sycophancy
- 27:45
- and it's well known that
- 27:46
- if you give
- 27:47
- even a hint of what answer you're hoping for
- 27:50
- there's a good chance the AI will just
- 27:52
- reflect back your preferences or your preconceptions
- 27:56
- In contrast
- 27:57
- AI power users
- 27:59
- tend to ask
- 28:00
- neutral questions that don't give any hint
- 28:02
- to the AI
- 28:03
- for what answer you're hoping for or not
- 28:05
- hoping for
- 28:06
- Or if you give it a rubric
- 28:08
- or grading criteria
- 28:10
- to tell the AI
- 28:11
- how to form the basis for its answer
- 28:13
- That also forces it to be more objective
- 28:17
- For example, if you were to say
- 28:19
- please analyze the following business idea objectively
- 28:22
- mobile tie-dyeing
- 28:24
- and
- 28:24
- don't just
- 28:25
- make up a bunch of things for what you think
- 28:27
- use the rubric or the grading criteria above
- 28:30
- such as, is there a problem?
- 28:32
- Is there a market?
- 28:33
- Do I have a competitive advantage?
- 28:35
- If you give instructions like this to the AI
- 28:37
- then the AI doesn't know
- 28:38
- are you hoping it'll tell you it's a
- 28:40
- great idea or that it will
- 28:42
- save you from
- 28:43
- spending a lot of time on the bad business idea
- 28:45
- and it's much more likely to then tell you something like
- 28:49
- oh, this idea is a $8.100
- 28:51
- and also why the school's low
- 28:55
- In case you run a mobile tie-dye business
- 28:57
- I wish you really best of luck
- 29:00
- and AI could also help ask some
- 29:02
- useful questions to
- 29:04
- help you think through how to make the
- 29:05
- business even better
- 29:07
- Lastly
- 29:08
- I've found that
- 29:09
- AI novices and AI power users
- 29:12
- ask AI to write in very different ways
- 29:15
- Novices will just ask AI to write stuff
- 29:18
- like write a blog post about the BlackBerry
- 29:21
- and it will generate a bunch of text
- 29:23
- that maybe looks like this
- 29:24
- which sounds like AI slop
- 29:26
- There's a bunch of generate text
- 29:29
- that's just not that interesting and
- 29:31
- takes up a lot of space
- 29:33
- In contrast
- 29:34
- an AI power user
- 29:35
- will often not
- 29:37
- ask the AI system to just jump to writing directly
- 29:40
- but instead ask the AI to first outline an article
- 29:43
- and then critique the outline
- 29:46
- and maybe
- 29:46
- iterate a few times with the outline to shape the article
- 29:50
- and only then
- 29:50
- ask the AI to start to draft the
- 29:52
- final article
- 29:53
- So given a set of uploaded notes as context
- 29:57
- an expert may say
- 29:58
- outline a blog post about the BlackBerry based on my notes
- 30:01
- so it knows what you want to talk
- 30:02
- about
- 30:04
- And the AI may start by giving an outline
- 30:08
- and you might then give feedback to the
- 30:10
- AI about what you like and what you
- 30:11
- don't like about the outline
- 30:14
- and even iterate a few times
- 30:15
- have a few back and forth rounds before
- 30:18
- you have an outline that you're satisfied with
- 30:20
- and maybe only then expand the outline into bullet points
- 30:24
- and maybe even go back and forth a
- 30:25
- few times to critique the bullet points
- 30:27
- before you're satisfied with that
- 30:29
- and then expand it into the final text
- 30:34
- This type of power user workflow
- 30:37
- is much more likely to generate some text
- 30:39
- that you're happy with
- 30:40
- as opposed to AI slop
- 30:43
- And in this type of workflow
- 30:44
- you're treating the AI as a
- 30:47
- thinking partner to
- 30:48
- almost help you brainstorm and explore different
- 30:51
- options for what you might want to write
- 30:54
- AI systems do make mistakes
- 30:57
- but maybe fewer than most people think
- 31:00
- especially if you prompt it well
- 31:02
- They made a lot more mistakes back in
- 31:04
- 2022 or 2023 than they do now
- 31:08
- but a lot of widely publicized mistakes that AI has made
- 31:13
- some of which went viral on social media
- 31:15
- has made people think that AI
- 31:17
- maybe makes even more mistakes than
- 31:19
- it actually does
- 31:21
- There's a well-publicized one where
- 31:23
- people asked it how many
- 31:24
- Rs are there in the word strawberry
- 31:26
- and it thinks there are two Rs
- 31:28
- and here's one that I found amusing
- 31:30
- I want to wash my car
- 31:32
- should I walk or drive there?
- 31:34
- And AI says walk
- 31:35
- which would
- 31:36
- leave you there or wash your car
- 31:38
- But these viral
- 31:39
- examples are not representative of AI capabilities
- 31:44
- In contrast
- 31:44
- power users know that
- 31:46
- AI can deliver
- 31:47
- significant value through tasks like
- 31:50
- doing deep research and writing research reports
- 31:53
- or taking your personal data
- 31:56
- like your health or heart rate or running time data
- 31:59
- and analyzing that for you
- 32:01
- Or several to talk about later
- 32:04
- even building websites for you
- 32:07
- I've seen being an AI power user
- 32:10
- tremendously benefit individuals as well as their businesses
- 32:14
- It'll save you time and improve your professional
- 32:17
- and personal lives. It'll help you to build
- 32:20
- lots of cool things
- 32:22
- you learn how later in these videos
- 32:25
- And being with a prompt AI
- 32:27
- at an expert level is an highly in demand job skill
- 32:31
- no matter what job role you're in
- 32:33
- In the rest of these videos
- 32:36
- I hope to take you from
- 32:37
- wherever you are today
- 32:39
- to being an AI power user
- 32:42
- Much has been said about AI being useful
- 32:45
- I find using AI really fun as well
- 32:48
- and you'll see a few examples of that
- 32:50
- in these videos too
- 32:52
- Now
- 32:53
- One
- 32:53
- foundational piece of knowledge
- 32:55
- that helps you work with AI
- 32:57
- is understanding where it gets its knowledge from
- 33:00
- so that you can better predict when it'll get something right
- 33:04
- and when you maybe shouldn't count on this
- 33:06
- answer
- 33:07
- Let's go on to the next video
- 33:09
- to learn about how AI gets its knowledge
- 33:13
- How did you
- 33:14
- learn to write as a child?
- 33:16
- Probably it involved reading a lot of things
- 33:19
- Well, it's the same for AI
- 33:21
- AI systems have learned patterns from reading large
- 33:25
- amounts of text from the internet
- 33:28
- By understanding what's in that text the AI has read
- 33:31
- you'd be better able to predict
- 33:33
- how they'll behave
- 33:35
- AI models can answer questions on a variety
- 33:38
- of topics
- 33:39
- If you were to ask
- 33:41
- I dropped my phone in soup
- 33:42
- what should I do?
- 33:44
- Then hopefully you'll make some useful suggestions
- 33:48
- Or
- 33:48
- why do cats stare at walls like they're
- 33:50
- seeing ghosts?
- 33:52
- My daughter loves cats
- 33:53
- She was actually curious about this
- 33:54
- Turns out
- 33:55
- cats can detect subtle sounds and movements
- 33:58
- that we as humans often miss
- 34:01
- Because of the amount of things is read on the internet
- 34:04
- they will
- 34:05
- even possess niche knowledge that few people
- 34:08
- know about
- 34:09
- They were to ask
- 34:10
- what kind of things
- 34:12
- were on the vinyl record sent into space
- 34:15
- Years ago
- 34:16
- NASA had a
- 34:17
- spacecraft called Voyager 1
- 34:19
- that
- 34:20
- launched in the 1970s and is now about
- 34:23
- 25 billion miles away from Earth
- 34:26
- But
- 34:26
- AI will know about this
- 34:28
- and be able to tell you what is
- 34:29
- on that vinyl record
- 34:31
- I think it's cool that NASA chose to
- 34:33
- send greetings in 55 different languages to
- 34:36
- whoever may come across that spacecraft
- 34:38
- if anyone does
- 34:40
- AI models are trained on many
- 34:43
- many different sources of information
- 34:45
- mainly from the internet
- 34:47
- and
- 34:48
- training on
- 34:48
- all of these very diverse sources of knowledge
- 34:51
- produces is pre-trained knowledge
- 34:54
- The term pre-trained is a technical term that
- 34:57
- you don't have to worry about
- 34:59
- It turns out AI systems are trained in multiple steps
- 35:02
- and this is one of the first steps
- 35:04
- of training that somehow wound up being called pre-training
- 35:08
- which isn't a great term
- 35:09
- but I wouldn't worry about why we call
- 35:11
- it pre-training
- 35:12
- It's just what AI has learned from
- 35:15
- But these knowledge sources may include a lot
- 35:17
- of texts from social media like Reddit
- 35:19
- which will have answers to questions like
- 35:21
- whether you'll must watch films
- 35:23
- or it may have read a book on Lego micro cities
- 35:28
- or read a Wikipedia article
- 35:30
- on fairy bread and lots of other things
- 35:33
- or read a bunch of news articles
- 35:35
- as well as read a lot of research
- 35:37
- articles
- 35:39
- On the internet
- 35:40
- there's a lot of text on internet forums
- 35:43
- and social media like Reddit and Quora
- 35:46
- There are a lot of books that AI
- 35:47
- will have read from
- 35:49
- There are encyclopedias like Wikipedia
- 35:52
- news websites, research articles, and much more
- 35:57
- And so these trillions or tens of trillions
- 35:59
- of words will go into training the AI model's brain
- 36:04
- Now
- 36:05
- different types of data
- 36:06
- appear with different amounts of frequency on the internet
- 36:10
- and so this pre-trained knowledge reflects
- 36:13
- the frequency or the patterns in the training data
- 36:16
- For example, cooking is a very universal human experience
- 36:20
- so there are a lot of articles on
- 36:22
- the internet on cooking
- 36:23
- There are also a lot of articles online on celebrities
- 36:26
- on movies
- 36:28
- and so AI will have seen a lot
- 36:29
- of text on these topics
- 36:31
- In contrast, there are more specialized topics like quasar
- 36:35
- which is an
- 36:36
- astronomical term referring to really bright objects in
- 36:39
- the sky powered by supermassive black holes
- 36:43
- I think they're fascinating
- 36:44
- but there's just a lot fewer articles on
- 36:45
- quasars than on cooking on the internet
- 36:48
- Now, while most of the internet is in English
- 36:51
- AI systems will also have learned from some
- 36:53
- data that's written in other languages
- 36:56
- like Cantonese
- 36:57
- Over 80 million people speak Cantonese
- 37:00
- but that's far less than English
- 37:03
- and Cantonese data represents
- 37:05
- maybe less than 0.1% of all internet content
- 37:10
- Lastly
- 37:10
- there are things that AI models know nothing about at all
- 37:14
- such as your company's secret proprietary data
- 37:18
- which hopefully is not on the open internet
- 37:22
- but which an AI
- 37:23
- system will therefore not have learned from
- 37:26
- So I find that
- 37:27
- thinking about
- 37:28
- how frequent data appears on the internet
- 37:31
- gives you a good rule of thumb
- 37:33
- for thinking about how reliable an AI system's responses are
- 37:38
- Now, because of the data the AI has learned from
- 37:41
- sometimes it can exhibit surprising understanding of things
- 37:46
- If you were to type very quickly
- 37:49
- can you cook eggs in microwave
- 37:51
- like shown on the left
- 37:52
- it can actually understand this type of misspelled text very well
- 37:56
- pretty much as well as asking
- 37:58
- can you cook eggs in the microwave?
- 38:00
- And by the way
- 38:01
- I've exploded a few eggs in the microwave myself
- 38:04
- so if you ever want to avoid that
- 38:07
- feel free to ask the AI system how to do so
- 38:10
- so you don't have to learn the hard
- 38:11
- way
- 38:12
- And a reason that it's so good at
- 38:15
- understanding
- 38:16
- misspelled words is because it's actually learned from
- 38:18
- a lot of sources that could include typos
- 38:21
- So if you look online
- 38:22
- you will see phrases misspelled
- 38:26
- and that's why when you are using the AI system
- 38:28
- I'm not encouraging you to use bad grammar
- 38:30
- or to misspell words
- 38:32
- but it turns out that if you're typing
- 38:33
- quickly and you have a
- 38:34
- few typos or even a lot of typos
- 38:37
- don't worry too much about it
- 38:38
- It's pretty fine to just
- 38:40
- send a prompt to the AI and not
- 38:42
- spend too much time fixing every little grammatical error
- 38:45
- Now
- 38:46
- the bad news is
- 38:47
- a lot of AI sources also have misconceptions
- 38:50
- and outdated information
- 38:52
- So
- 38:53
- one of the skills
- 38:54
- in using AI
- 38:55
- is how to prompt it to
- 38:57
- have it give you back answers that reflect
- 39:00
- fewer misconceptions and
- 39:02
- does not overly reflect outdated information
- 39:06
- By understanding AI's knowledge sources
- 39:09
- call this pre-trained knowledge
- 39:11
- You people better predict how it will respond
- 39:13
- to your prompts
- 39:15
- But this pre-trained knowledge is not enough for all applications
- 39:19
- including those that need real-time information
- 39:22
- For that, you need web search
- 39:25
- Let's go on to the next video
- 39:27
- to learn more
- 39:29
- At some point
- 39:29
- the people building the AI model had to
- 39:32
- stop its training
- 39:34
- So there's some lost date where its
- 39:37
- information cuts off
- 39:39
- That list
- 39:40
- the AI's read the internet
- 39:42
- only up to a certain date and time
- 39:44
- and its knowledge gets
- 39:46
- frozen in time as of that date
- 39:49
- But, of course, the world moves on past that date
- 39:53
- new things happen, movies come out
- 39:56
- and so on
- 39:57
- Let's see how AI models
- 39:58
- handle gathering new information using web search
- 40:02
- so that it can
- 40:03
- address questions that even relate to things
- 40:07
- after its knowledge cutoff date
- 40:10
- If you're using one of the popular AI model providers
- 40:13
- like ChatGPT, Gemini, and Claude
- 40:16
- there are certain questions that will probably
- 40:19
- trigger it to do a web search
- 40:22
- For example, if you ask it
- 40:23
- what is the 6-7 meme from 2025
- 40:27
- there's a good chance it will search on
- 40:29
- the internet to tell you that the 6-7 meme
- 40:33
- which is pronounced six seven
- 40:35
- which is kind of fun to say
- 40:37
- that this is a viral internet slang
- 40:39
- widely seen on a few social media platforms
- 40:43
- And the reason it triggers a web search
- 40:45
- when you ask it
- 40:45
- what's a 6-7 meme from 2025
- 40:48
- the Q 2025
- 40:49
- causes the AI to realize that
- 40:51
- it may benefit from more updated online information
- 40:55
- because this could be a meme that
- 40:56
- appeared on the internet
- 40:58
- after its knowledge cutoff date
- 41:01
- Here's what I mean
- 41:03
- A specific AI model's pre-trained knowledge is frozen in time
- 41:07
- even though the internet
- 41:08
- continues to evolve over time
- 41:11
- And so if this line represents time
- 41:15
- then
- 41:16
- for a long time the internet will have
- 41:18
- had pieces of text that say 6 times 7 equals 42
- 41:22
- text that talks about the children's joke
- 41:24
- why was 6 afraid of 7?
- 41:27
- Because 7, 8, 9
- 41:30
- But if the knowledge cutoff date
- 41:32
- was at a certain moment in time
- 41:34
- and the 6, 7 meme came after that
- 41:37
- then the 6
- 41:38
- 7 meme will not have been seen in
- 41:40
- the pre-trained knowledge of the AI model
- 41:44
- So if you ask it
- 41:45
- what is the 6
- 41:46
- 7 meme from 2025?
- 41:48
- The AI model
- 41:49
- will realize that it doesn't know about this 6
- 41:52
- 7 meme from 2025
- 41:54
- and that it should
- 41:55
- do a web search
- 41:56
- in order to get more updated information
- 41:59
- Take the GPT-5.4 model from OpenAI
- 42:04
- Its knowledge cutoff date was August 2025
- 42:08
- and this graph shows
- 42:09
- how many Google web searches there were for
- 42:11
- whether 6-7 mean
- 42:13
- So this 6-7 meme had taken off
- 42:17
- after this GPT-5.4 knowledge cutoff date
- 42:21
- which is why the model doesn't really know
- 42:23
- about this meme
- 42:25
- Now
- 42:25
- there are certain types of questions that an
- 42:27
- AI will answer using its pre-trained knowledge
- 42:31
- and there are certain types of questions that
- 42:32
- will tend to trigger web search
- 42:35
- For example, if you tell it
- 42:36
- please find me a highly rated gym near Mountain View
- 42:39
- California, then what is highly rated
- 42:43
- what may be open
- 42:43
- and what may be closed
- 42:44
- does change over time
- 42:46
- and there's a good chance that this will
- 42:48
- trigger a web search
- 42:50
- Well, if you ask it
- 42:51
- what is the market mountain cheese roll?
- 42:54
- Because this is a niche piece of information
- 42:56
- it's probably not read a lot of information
- 42:58
- online about this cheese roll
- 43:00
- There's a good chance that it will search
- 43:02
- the internet in order to
- 43:03
- get you an answer
- 43:05
- And if you're curious
- 43:06
- this is actually a
- 43:07
- pretty fun events where people chase a
- 43:11
- rolling wheel of cheese down a hill
- 43:13
- Let's take a look more broadly at when
- 43:15
- an AI model needs to do some web
- 43:17
- search to gather more information to answer your
- 43:20
- question
- 43:21
- If you're asking what to do if you
- 43:23
- drop your phone in soup
- 43:25
- or why do cats stare at walls
- 43:26
- or walls on the Voyager 1 record in outer space
- 43:30
- then these questions it could probably answer using this pre-chain knowledge
- 43:35
- because these
- 43:36
- are represented in common knowledge on the internet
- 43:40
- But if it was asked about
- 43:42
- current events or something happening very recently
- 43:46
- then it'll need to do web search to
- 43:47
- get that real-time information
- 43:49
- If you ask it location-specific information
- 43:53
- and doing a web search makes sense
- 43:55
- or if you ask it for other types of niche information
- 43:59
- there's
- 43:59
- also a good chance it'll realize it doesn't
- 44:01
- know enough about that topic
- 44:03
- that doing a web search to gather more
- 44:05
- information
- 44:06
- would help it give you a better answer
- 44:09
- For most of the popular AI model providers
- 44:11
- web search can be triggered in
- 44:13
- either of two ways
- 44:15
- Sometimes the AI model will decide by itself
- 44:17
- to carry the web search
- 44:20
- or you can also explicitly trigger web search
- 44:23
- sometimes by
- 44:24
- clicking one of the buttons in a
- 44:26
- AI model provider's web interface
- 44:29
- or just writing a prompt
- 44:30
- please do a web search for this
- 44:32
- and it will comply
- 44:34
- and
- 44:34
- use a web search to answer your question
- 44:37
- Not all AI models have web search enabled
- 44:40
- but the most popular ones that you're probably
- 44:42
- using mostly do have this capability
- 44:46
- AI would do better on many of the
- 44:48
- tasks you want to use it for
- 44:50
- if it does web search
- 44:52
- And web search allows it to
- 44:54
- augment its pre-trained knowledge with more current information
- 44:59
- But, like all web search
- 45:01
- it can return bad sources
- 45:04
- Let's take a look
- 45:05
- at when this is an issue
- 45:06
- and when and how
- 45:08
- to get it to use more reliable sources
- 45:11
- to get you more reliable answers
- 45:14
- Web search is a very valuable but imperfect
- 45:17
- tool
- 45:18
- Just like when you search the web yourself
- 45:21
- you might not always find what you're looking
- 45:23
- for
- 45:24
- It has limitations like finding old or inaccurate sources
- 45:29
- But
- 45:29
- you can work around these limitations
- 45:31
- to get AI
- 45:32
- to give you more accurate and up-to-date answers
- 45:35
- Let's take a look
- 45:37
- If you ask a AI system
- 45:40
- how safe are gray market peptides
- 45:43
- which is a type of supplement
- 45:45
- it may
- 45:46
- search online and find posts on
- 45:48
- social media or public forum sites like Reddit and Quora
- 45:53
- or
- 45:54
- you may find websites that are
- 45:56
- in the business of selling peptides
- 45:57
- and so would have a
- 45:59
- inclination to tell you that they're safe
- 46:02
- and you may get back answers that may
- 46:04
- or may not be accurate
- 46:06
- But if you
- 46:07
- encourage the AI model
- 46:08
- to use sources from official organizations or
- 46:12
- look at studies that are backed by rigorous science
- 46:15
- then it's more likely to look at resources
- 46:18
- from the World Health Organization
- 46:20
- from the US Food and Drug Administration
- 46:22
- from the European Medicines Agency
- 46:25
- and so on
- 46:26
- and hopefully give you more
- 46:28
- reliable and scientifically credible answers
- 46:32
- Web search, whether done by a human on Google or Bing
- 46:36
- or done by AI
- 46:38
- has a tendency to draw from popular sources
- 46:42
- According to one report
- 46:44
- the most cited website by a AI model was Reddit
- 46:49
- followed by Wikipedia, YouTube, then Google itself, Yelp, and so on
- 46:54
- and
- 46:55
- some of these sources are
- 46:56
- more trustworthy than others
- 46:58
- There's just a lot of text on the
- 47:00
- internet from social media
- 47:01
- blogs, online forums, and the amount of text from highly reliable
- 47:07
- scientifically verified sources is just much smaller
- 47:11
- So
- 47:12
- if you don't
- 47:13
- steer the model
- 47:14
- in terms of what types of sources you prefer
- 47:17
- there's a chance that it'll tend to pull
- 47:19
- text from whatever is most available rather than
- 47:22
- what's most reliable
- 47:25
- So that's why if you ask it
- 47:26
- how safe are gray market peptides
- 47:29
- it might base a lot of its answer on social media
- 47:32
- blogs, and forums
- 47:33
- and only a little bit
- 47:34
- on the more reliable sources
- 47:36
- Whereas if you tell it to use
- 47:39
- sources from official health organizations
- 47:41
- it may pull much more from these reliable
- 47:43
- sources
- 47:45
- Another limitation of web search is that
- 47:48
- sometimes web pages can be outdated
- 47:50
- That can lead the AI model to also
- 47:53
- not provide the most current information
- 47:56
- A friend asked AI recently
- 47:58
- help me find places to run in Henderson
- 48:00
- Nevada
- 48:02
- This is a location-specific niche query
- 48:05
- and so this triggers web search
- 48:07
- and it found this list of places
- 48:10
- to go for a jog
- 48:12
- But it turns out that unfortunately
- 48:14
- this pulled from a webpage from more than two decades ago
- 48:18
- and unfortunately, the location suggested was a school that
- 48:22
- unlike decades ago
- 48:24
- is no longer open to the public to
- 48:26
- go running in
- 48:27
- To help build intuition about how
- 48:31
- AI searches the web to use that information
- 48:34
- let me briefly explain
- 48:36
- how web search actually works under the hood
- 48:39
- It turns out to be a multi-step process
- 48:42
- Imagine that you're asking questions of a
- 48:44
- customer service team of two people
- 48:48
- There's the user-facing AI model
- 48:50
- that's what you are talking to
- 48:53
- and the user-facing AI model has a second
- 48:55
- assistant AI model that
- 48:57
- it can ask for help
- 48:58
- to do web search
- 49:00
- So when you send a prompt
- 49:03
- You are talking to the first model
- 49:04
- the user-facing AI model
- 49:07
- and it will occasionally decide to co-opt the assistant AI model
- 49:11
- the second AI, to say,"Hey
- 49:13
- please do a web search for me to
- 49:15
- gather more information."
- 49:16
- This assistant AI model will then
- 49:19
- search on a web search engine
- 49:21
- very similar to
- 49:22
- Google and Bing and other web search engines
- 49:24
- that we as people might use
- 49:26
- and it will scan the return results
- 49:30
- filter out the relevant results
- 49:32
- and
- 49:32
- download the most relevant
- 49:34
- web pages and then summarize them
- 49:37
- The second assistant AI model will then
- 49:39
- present the summaries back to the first model
- 49:42
- the user-facing AI model
- 49:45
- and the first model will then use these
- 49:46
- summaries in order
- 49:47
- to generate the final answer for you
- 49:51
- You are speaking only to the user-facing AI model
- 49:55
- and one interesting quirk to keep in mind
- 49:58
- is the user-facing AI model
- 50:00
- has not actually read in its entirety
- 50:03
- all of the web pages
- 50:04
- it may be citing for you
- 50:07
- Instead, it's only seen summaries of those web pages
- 50:11
- and sometimes this causes it to misinterpret
- 50:14
- what one of these underlying web pages actually says
- 50:17
- which is why you may have seen
- 50:19
- funny results
- 50:20
- where AI cites a web page and says
- 50:22
- the web page justifies a conclusion
- 50:25
- but if you look at that web page yourself
- 50:27
- it doesn't actually
- 50:28
- justify what the user-facing AI model
- 50:31
- says it is doing
- 50:33
- To walk you through one example of this process
- 50:36
- if you ask the user-facing AI model
- 50:39
- that's like the customer service agent talking to you
- 50:42
- It revolves
- 50:43
- what should I know before hiking Machu Picchu?
- 50:46
- The second model may
- 50:48
- do some web searches with phrases like Machu Picchu pyramids
- 50:51
- Machu Picchu weather, or the social customs
- 50:54
- and so on, and it'll then scan the returned results
- 50:59
- much like you may scan a
- 51:01
- page of Google results to decide what's relevant
- 51:04
- and filter out irrelevant results
- 51:06
- and summarize the most relevant web pages to
- 51:08
- provide back to the first agent
- 51:11
- that then generates the final answer for you
- 51:15
- Now, I frequently use AI models like ChatGPT
- 51:19
- Gemini, Cloud
- 51:20
- and I also frequently use web search engines
- 51:23
- like Google and Bing
- 51:25
- When should you use a AI model
- 51:27
- and when should you use a
- 51:29
- web search engine?
- 51:30
- If you want to quickly scan multiple sources
- 51:34
- a search engine can be useful for that
- 51:37
- Or if you want to navigate to a specific website
- 51:40
- but have forgotten what's the name of that website
- 51:44
- a web search engine can be very good
- 51:45
- for helping you find it
- 51:48
- Or if you want to look at data
- 51:49
- in its original form
- 51:51
- such as if you want to buy a
- 51:53
- 2013 Honda Civic air filter
- 51:56
- you know
- 51:56
- you want to find a website to go
- 51:58
- to to buy that air filter
- 52:00
- so a web search engine
- 52:01
- is very good at that
- 52:03
- In contrast
- 52:04
- if you want to get a synthesis from multiple sources
- 52:09
- or if you're searching for
- 52:11
- more complex information with pros and cons that you want weighed
- 52:15
- or if you just want to contrast multiple
- 52:18
- sources to come up with a more thoughtful conclusion
- 52:22
- then an AI model can
- 52:24
- do a web search and
- 52:26
- put together
- 52:27
- the results of multiple web pages for you
- 52:29
- quite efficiently
- 52:30
- That's maybe saving you time of
- 52:32
- having to read a lot of web pages
- 52:34
- yourself
- 52:35
- There might be some good
- 52:36
- Google or other web search habits that you've developed
- 52:40
- and those habits will serve you well
- 52:42
- when working with
- 52:43
- web search-enabled AI models as well
- 52:46
- Things like
- 52:47
- looking for reliable sources and also
- 52:50
- double-checking the sources
- 52:52
- But
- 52:53
- if you want to go
- 52:54
- beyond searching a handful of web pages
- 52:57
- it turns out
- 52:58
- AI models are capable of a much more
- 53:00
- extensive type of research called deep research
- 53:05
- This is a
- 53:05
- very powerful capability that I think is
- 53:08
- really underused by many people
- 53:11
- Let's go on to the next video
- 53:13
- to see what it is
- 53:14
- and when and how
- 53:16
- to use a deep researcher
- 53:19
- Sometimes you will want your AI to
- 53:21
- synthesize not just a handful of sources
- 53:24
- but many, maybe many dozens of sources
- 53:28
- and do lots of thinking to come up
- 53:30
- with the best possible deeply researched answer
- 53:33
- to a question that you have
- 53:35
- Popular AI chat interfaces like ChatGPT
- 53:38
- Gemini
- 53:39
- and Cloud all have a deep research mode
- 53:42
- I found this to be a very valuable
- 53:44
- and often underutilized tool. Let's take a look
- 53:48
- Let's say you want to
- 53:50
- use an AI model to help you plan
- 53:52
- your Halloween haunted house
- 53:54
- I'm going to write a prompt
- 53:55
- to ask it to help me set up
- 53:57
- a haunted house in my front yard for Halloween
- 54:00
- and give it some information about where I am
- 54:02
- what's the size of my front yard
- 54:04
- what's the experience I want
- 54:06
- So I'll give it lots of context to set it up
- 54:09
- to plan it out for me appropriately
- 54:12
- With a prompt like this
- 54:14
- an AI model
- 54:15
- might come up with a research plan
- 54:18
- And when she tries to think through
- 54:20
- what are the types of sources it needs to research
- 54:24
- many systems would give you an opportunity
- 54:26
- to approve or potentially edit the research plan
- 54:30
- And if you're happy with it
- 54:32
- I'll often launch the research plan without updating
- 54:35
- it
- 54:35
- unless I see something that just looks really
- 54:37
- wrong
- 54:38
- It will then go ahead and
- 54:40
- start to do online searches
- 54:43
- So this example is
- 54:44
- starts by gathering Palo Alto's rules on permits
- 54:48
- Halloween ordinances, and so on
- 54:50
- And then it will
- 54:52
- read some of those web pages and
- 54:54
- synthesize what is learned so far
- 54:57
- And it may then decide to
- 54:59
- do some more
- 55:01
- searches online to gather more information about
- 55:04
- fire safety guidelines
- 55:06
- And then
- 55:07
- it may after that decide to look for decoration ideas
- 55:11
- So, mostly follow the original research plan
- 55:14
- but also have the flexibility
- 55:16
- to keep looking deeper into
- 55:17
- certain areas if it thinks it needs that
- 55:19
- information
- 55:21
- After searching for a while
- 55:22
- maybe many minutes
- 55:24
- it will finally write a detailed research
- 55:26
- report for you
- 55:28
- This process, by the way, is an example of agentic AI
- 55:33
- and what that refers to
- 55:34
- is that through this deep research process
- 55:37
- the
- 55:37
- AI model has some flexibility to make decisions
- 55:40
- by itself on what to do next
- 55:42
- such as what additional searches
- 55:45
- if any, to carry out
- 55:48
- The output of this can then be a
- 55:50
- fairly
- 55:51
- detailed and thoughtful plan with different sections aligning
- 55:55
- what you might need to think about in
- 55:56
- terms of structural and regulatory framework
- 55:59
- safety, and so on
- 56:02
- If you're using Google's Gemini AI model for this
- 56:06
- one of the new features is it
- 56:07
- makes it easy to take the deep research
- 56:09
- is done
- 56:10
- and help you turn it into
- 56:13
- a webpage or infographic or
- 56:15
- handful of other things
- 56:17
- Here's a webpage that was generated by
- 56:19
- Gemini using the Gemini deep researcher
- 56:23
- and I think it's pretty neat
- 56:24
- There's created a webpage with four different sections
- 56:29
- pie charts for budget
- 56:33
- pretty neat visualizations for noise ordinance
- 56:37
- and I think it's pretty neat that this
- 56:39
- even has a
- 56:40
- little checklist that I could use to plan
- 56:42
- out my Halloween event
- 56:45
- To give you a sense of how a deep researcher works
- 56:48
- this is loosely what it does
- 56:51
- After formulating a research plan
- 56:53
- An AI model can actually
- 56:56
- issue
- 56:56
- many web searches at the same time
- 56:59
- and get back multiple web pages
- 57:02
- at the same time
- 57:03
- And this is one of the nice things
- 57:05
- about using a AID researcher
- 57:08
- It doesn't have to do the web searches
- 57:10
- one at a time
- 57:11
- it can do many of them at the same time
- 57:15
- which lets it be very efficient in
- 57:17
- fetching lots of web pages
- 57:19
- The AI system can also
- 57:21
- take a look at
- 57:22
- all of these sources
- 57:24
- and quickly
- 57:25
- assess which ones are relevant and which ones are less relevant
- 57:30
- and based on that
- 57:31
- it may decide whether or not to go
- 57:33
- back to do additional web searches
- 57:35
- maybe using different web search terms
- 57:38
- Finally
- 57:39
- after going around this loop a few times
- 57:41
- of doing web search
- 57:42
- evaluating sources
- 57:44
- deciding whether or not to go back to get more sources
- 57:47
- it'll hopefully decide it's done
- 57:49
- And then lastly
- 57:51
- take all of the pages it has
- 57:53
- downloaded and maybe summarize
- 57:55
- and synthesize all that into a report
- 57:59
- that it adds citations to
- 58:02
- and that it then presents to you
- 58:04
- Both web search-enabled AI as well as deep
- 58:07
- research use the internet or do web search
- 58:11
- The basic web search-enabled AI is good at
- 58:14
- queries like this
- 58:16
- Find me a highly rated gym
- 58:17
- what's the weather in Dubai this week?
- 58:19
- Whereas deep research
- 58:21
- I would tend to use for tasks that
- 58:23
- require synthesizing multiple views
- 58:25
- such as
- 58:26
- if I want to know
- 58:27
- what's the impact of daily steps on long-term health
- 58:30
- and if I wanted to
- 58:32
- search the most recent
- 58:34
- scientifically justified articles and think through the answer
- 58:38
- rather than just tell me whatever people tend
- 58:40
- to say on the internet
- 58:42
- Or if I wanted to deeply think through
- 58:45
- how does the weather affect tourism in Dubai?
- 58:48
- And again
- 58:48
- not just
- 58:49
- take one or two popular answers found on
- 58:51
- a social media site
- 58:52
- but read up on weather
- 58:54
- read up on tourism
- 58:55
- read up on Dubai
- 58:56
- and to really think through the implications to
- 58:58
- give a more thoughtful answer
- 58:59
- That's when deep research could be particularly helpful
- 59:03
- To give you another framework to think through
- 59:05
- when to use web search versus deep research
- 59:08
- if you have a single question that you want answered
- 59:11
- doing work that would take me just a
- 59:13
- few seconds based on a handful of sources
- 59:17
- that's when web search would be particularly helpful
- 59:20
- And web search, as we've seen
- 59:22
- can be triggered either automatically
- 59:25
- or by the user
- 59:26
- Whereas deep research is
- 59:28
- often is trying to
- 59:30
- draw a complex to the conclusions that may
- 59:33
- require answering multiple questions or answering multiple
- 59:36
- dimensions that relate to a question
- 59:40
- And I think of this as
- 59:41
- doing work that would take me
- 59:43
- minutes to maybe even hours if I was
- 59:45
- doing this manually
- 59:48
- And I may want many sources of information
- 59:50
- synthesize, and as we've seen
- 59:53
- deep researcher is
- 59:55
- usually triggered explicitly by the user
- 59:57
- Unless you select it in the user interface
- 60:01
- most AI models would not
- 60:03
- care on a deep researcher and keep you
- 60:05
- waiting for many minutes for an answer
- 60:07
- To recap, if you're asking
- 60:09
- hope I drop my phone in soup
- 60:12
- it doesn't need to look up any online
- 60:13
- sources
- 60:15
- We're not worried about freshness
- 60:17
- It'll give you an answer in just a
- 60:18
- few seconds
- 60:19
- And this is good for finding basic facts
- 60:22
- definitions
- 60:23
- summaries for things that
- 60:24
- occur commonly on the internet
- 60:26
- Web search may download a handful of sources
- 60:31
- and it will find relatively up-to-date information
- 60:34
- and it may take many seconds
- 60:36
- to get you back an answer
- 60:38
- And we've seen
- 60:39
- what types of information this is useful for
- 60:42
- And lastly
- 60:43
- deep research may
- 60:45
- download often dozens or more of sources
- 60:49
- It will get up-to-date information
- 60:51
- and it will spend
- 60:52
- many minutes
- 60:53
- or longer to give you back an answer
- 60:56
- And it's great at answering complex questions that
- 60:58
- involve synthesizing many sources of knowledge
- 61:02
- Finding information is
- 61:03
- one of the most common tasks that people
- 61:05
- use AI models for
- 61:07
- You've seen three different paths
- 61:10
- that you can take advantage of
- 61:12
- for this type of information-finding task
- 61:15
- You could use just the pre-trained knowledge
- 61:17
- or web search, or deep research
- 61:22
- and we also
- 61:23
- walk through
- 61:24
- how and when to use these different options
- 61:27
- I want to make sure that you have
- 61:28
- good intuitions about when to use each of these options
- 61:33
- So
- 61:34
- let's go on
- 61:35
- to take a look next
- 61:36
- at a practice hands-on lab for this module
- 61:39
- which I think you find a fun way
- 61:41
- to compare and contrast what each of these three options do
- 61:45
- and more importantly
- 61:46
- will also help you hone your intuition
- 61:48
- on when to use each
- 61:51
- In this module
- 61:52
- you learn how to use AI models
- 61:54
- to find information
- 61:56
- In this practice lab
- 61:58
- you can explore how web search
- 62:00
- deep research
- 62:01
- and different prompts affect the AI model's output
- 62:05
- Let's take a look
- 62:06
- When you open up the lab
- 62:08
- it starts off with a
- 62:09
- tutorial on how to use the lab
- 62:12
- and I'm just gonna close it out here
- 62:14
- We can always access this tutorial again
- 62:17
- via this button up here
- 62:19
- And what I hope you do is
- 62:21
- follow the instructions written here
- 62:24
- and when you're done
- 62:25
- click mark as complete to
- 62:28
- mark this item as complete
- 62:29
- Repeat it
- 62:31
- First
- 62:31
- these buttons down here correspond to
- 62:34
- different things you might try
- 62:36
- So this one, current events
- 62:38
- compares a question with and without web search
- 62:42
- So what's the 6-7 meme?
- 62:44
- This on the left is without web search
- 62:47
- The one on the right is
- 62:48
- with web search enabled. And if you compare
- 62:52
- then without web search
- 62:54
- it
- 62:54
- gives these answers. It doesn't know about the
- 62:56
- meme
- 62:57
- But with web search
- 62:58
- it tells you what is this 6-7 meme
- 63:02
- as well as the origins of this meme
- 63:05
- And if you want
- 63:05
- you can also follow up and ask
- 63:08
- how do I use 6-7 appropriately
- 63:12
- in a black tie tuxedo party?
- 63:17
- And then hit this red button
- 63:19
- to go see his answer
- 63:22
- Let's go back to
- 63:23
- the homepage by clicking new chat up here
- 63:26
- And I hope you try out the other examples
- 63:29
- such as find me a highly rated gym and hit compare
- 63:33
- or when's the next Avengers movie scheduled
- 63:36
- or
- 63:37
- what major news happened today
- 63:39
- in the US or if you'd enter your own country
- 63:42
- and compare these
- 63:43
- results with and without web search
- 63:46
- This example over here
- 63:48
- shows the difference between web search denoted by
- 63:51
- this globe icon
- 63:52
- versus deep researcher denoted by this microscope icon
- 63:57
- So, how safe are gray market peptides
- 63:59
- use high quality sources
- 64:01
- and
- 64:02
- this will give you
- 64:04
- a sense of what
- 64:05
- web search results looks like
- 64:09
- versus deep researcher results
- 64:12
- I hope you run it and see what
- 64:14
- results you get
- 64:16
- One more example, if I want to ask
- 64:18
- can I keep my
- 64:19
- rocket-propelled monster truck in my garage?
- 64:22
- If you Ask it this question
- 64:25
- the AI model may or may not know the answer
- 64:28
- but if you were to also upload the lease agreement
- 64:33
- then maybe your lease agreement
- 64:36
- which states the terms under which you are renting your place
- 64:40
- may have restrictions on that
- 64:42
- And so you will see
- 64:44
- different answers depending on whether or not you
- 64:47
- upload additional information
- 64:49
- In this case
- 64:50
- a lease agreement
- 64:51
- that is helpful for the AI to give
- 64:54
- you a thoughtful answer
- 64:56
- We haven't talked much yet about uploading your
- 64:59
- own files in this module
- 65:00
- but it's something
- 65:01
- we'll dive into more deeply later in these videos
- 65:05
- but feel free to play with this now
- 65:08
- Lastly, one fun example
- 65:11
- Here
- 65:11
- is a version of why do cats stare
- 65:14
- at walls with lots of typos
- 65:15
- Here's one with nicely formatted grammar
- 65:18
- no typos
- 65:20
- And you find that the answers are
- 65:23
- maybe surprisingly similar
- 65:27
- That the AI system is pretty good at
- 65:29
- answering a question like this
- 65:30
- even if there's lots of typos
- 65:33
- Once you tried out these examples reflected by
- 65:35
- these buttons down here
- 65:37
- come over your own
- 65:38
- like
- 65:39
- is the weather good for a picnic
- 65:41
- in Palo Alto today?
- 65:45
- And
- 65:46
- try this with or without web search
- 65:51
- and see what answers you get
- 65:53
- or pick your own example and try
- 65:55
- comparing the results you get using
- 65:57
- web search and if you want uploading your own files
- 66:03
- So
- 66:04
- that's it for module one of this course
- 66:06
- Great job getting this far
- 66:08
- I hope you enjoy playing around with the lab
- 66:12
- Next
- 66:12
- please join me
- 66:13
- in the next module
- 66:14
- where you hear about using AI as a thought partner
- 66:18
- including having it brainstorm with you and explore
- 66:21
- ideas with you
- 66:23
- This has helped me shape the direction of many projects
- 66:26
- and I'm confident you'll find it useful too
- 66:30
- And we'll also explore getting AI to help
- 66:32
- you with your writing and editing
- 66:35
- I'll see you in the next module
- 66:48
- One of the most helpful uses of AI
- 66:50
- is as a thought partner
- 66:52
- When
- 66:53
- I'm trying to think through a complex problem
- 66:55
- or make a complex decision
- 66:57
- it's nice to have a human expert as a thought partner
- 67:00
- that is, someone to talk things through with
- 67:04
- But if there isn't a human expert readily available
- 67:07
- AI, which actually knows a lot about a lot of things
- 67:10
- can be a really good resource for this
- 67:13
- We'll go through together multiple examples of this
- 67:16
- but to get started
- 67:17
- brainstorming is one great such use case
- 67:21
- Now
- 67:22
- I know a lot of people
- 67:23
- ask AI to help brainstorm lists of ideas
- 67:27
- but
- 67:28
- there are more effective ways to use it
- 67:30
- as a brainstorming partner
- 67:31
- than just having it generate a list
- 67:34
- Let me show you what I mean
- 67:36
- According to data released by OpenAI analyzing ChatGPT conversations
- 67:42
- about half of ChatGPT chats are
- 67:44
- asking for writing and practical guidance
- 67:48
- And in fact, creative ideation accounts for 3.9%
- 67:52
- almost 4% of all chats
- 67:55
- I've found using
- 67:56
- AI to help me brainstorm to be really
- 67:58
- valuable
- 67:59
- Let me share with you
- 68:01
- some ways to do so
- 68:03
- AI can be pretty good at generating options
- 68:07
- There's a common creativity test which
- 68:09
- asks people to name 200 potential uses for a brick
- 68:14
- So, given a brick like this
- 68:16
- how many uses can you think of it?
- 68:19
- This is actually pretty difficult
- 68:21
- Some people think, oh, it could be a paperweight
- 68:23
- maybe a planter, and, oh
- 68:26
- it could be used to build a house too
- 68:27
- I guess
- 68:28
- But to come up with 200 examples is
- 68:30
- not that easy
- 68:32
- But if you ask an AI model
- 68:34
- there's a good chance you can come up
- 68:36
- with a long list of ideas
- 68:38
- and your role
- 68:39
- if you're actually trying to use a brick for something
- 68:41
- would be to evaluate these options to pick
- 68:44
- out which ones are the ones that you like
- 68:46
- In brainstorming, common guidance is the more ideas
- 68:50
- the better
- 68:51
- And so sometimes
- 68:52
- having AI generate a lot of ideas for
- 68:54
- you to pick from
- 68:55
- can be a powerful way to find
- 68:57
- one or two good ideas
- 68:59
- So
- 69:00
- this is the
- 69:01
- maybe more common use of
- 69:03
- AI as a brainstorming partner
- 69:06
- I want to show you a different form
- 69:07
- of brainstorming in which you give it more context
- 69:11
- and then also iterate with the AI longer
- 69:14
- meaning have a longer back and forth conversation
- 69:17
- to help get you to better options
- 69:20
- So, if you tell it
- 69:22
- help me build a workout plan
- 69:23
- I'm 38, bring the level
- 69:24
- have 10-pound dumbbells in 15 minutes a day
- 69:27
- then the AI may give fairly generic answers
- 69:30
- like three workout plans
- 69:32
- start with 10 squats
- 69:33
- 10 push-ups, pretty reasonable, very sensible
- 69:36
- common sense answer
- 69:38
- But if you want more creative options
- 69:40
- giving it more context can be helpful
- 69:43
- So if you say
- 69:44
- I can't stick to these
- 69:45
- give me hacks to stay on track
- 69:47
- I have a trampoline and a cat
- 69:49
- By
- 69:49
- encouraging it to give you trampoline and cat related workout options
- 69:54
- which is an unusual way of approaching workouts
- 69:57
- it may ask you to consider trampoline breaks
- 70:00
- or cat triggered micro workouts where
- 70:03
- maybe every time you see your cat wag
- 70:05
- its tail or something
- 70:06
- go to a tiny little workout
- 70:08
- But these are certainly more creative ideas
- 70:11
- AI models
- 70:12
- have some inherent creativity because they've trained on
- 70:15
- a lot of texts on the internet
- 70:17
- which covers a lot of very different ideas
- 70:20
- including some creative ones
- 70:23
- And AI's output is a little bit random
- 70:25
- So if you ask it multiple times
- 70:28
- help me build a workout plan
- 70:30
- it'll probably give you slightly different answers
- 70:33
- But if you give the AI basic questions
- 70:36
- then common sense relatively generic responses like do squats
- 70:40
- push-ups, and so on are more likely
- 70:43
- Let me plot a conceptual diagram where on the horizontal axis
- 70:47
- I'm going to plot how unique a response is
- 70:50
- how creative a response is
- 70:51
- So on the left were responses like
- 70:54
- normal weightlifting exercises like bicep curls
- 70:57
- which is very common sense
- 70:59
- to then maybe slightly more unique things like
- 71:01
- standing on one leg with a yoga block on your head
- 71:05
- to the really creative ones like cat triggered
- 71:08
- micro workouts
- 71:10
- And on the vertical axis
- 71:11
- I'm going to plot the probability
- 71:13
- of AI giving these different responses
- 71:16
- And it turns out that it's much more
- 71:18
- likely to give a common sense response than
- 71:20
- a highly unique creative response
- 71:23
- There's a reason for this
- 71:25
- namely
- 71:26
- it was trained on internet text and a
- 71:28
- lot more internet text
- 71:29
- talking about dumbbell curls than there are
- 71:31
- cat-triggered micro workouts
- 71:34
- And for most questions
- 71:36
- this is actually okay
- 71:38
- because the
- 71:39
- average information on the internet is probably decently
- 71:42
- factual
- 71:43
- So when you are seeking information such as
- 71:46
- what's the tallest building in the world
- 71:48
- is actually the Burj Khalifa
- 71:50
- most internet texts will say is the Burj
- 71:52
- Khalifa
- 71:53
- There are smaller amounts of text that will name other buildings
- 71:56
- but the
- 71:57
- average response or the most common response on
- 72:00
- the internet is usually the factual one for questions like
- 72:04
- what's the tallest building?
- 72:06
- But if you're brainstorming
- 72:07
- then giving the
- 72:09
- average information or the most common response ends up with squats
- 72:12
- push-ups
- 72:13
- and almost never trampoline breaks and
- 72:16
- pretty much never cat-based sessions
- 72:19
- Which is why
- 72:20
- if you ask the AI model to brainstorm with you
- 72:23
- you get a lot of common sense ideas
- 72:26
- rather than the more creative ideas
- 72:28
- which depending on your goal
- 72:30
- may or may not be what you want
- 72:33
- So
- 72:34
- what do you want to do
- 72:34
- if you want to get high-quality
- 72:36
- more creative ideas from AI?
- 72:40
- We've seen with a basic prompt
- 72:42
- you get responses from the common sense space
- 72:46
- But if you give the AI model more context
- 72:49
- so it give you age
- 72:50
- your level, but also tell it
- 72:52
- do you have a mini trampoline
- 72:53
- a cat, troubles stay motivated, no squats
- 72:57
- then this context pushes it into
- 73:00
- the more relevant and creative space
- 73:03
- and it's more likely to give a custom answer
- 73:06
- rather than to generate common sense answers
- 73:09
- Now, one problem that you may face when brainstorming is
- 73:14
- if you're trying to come up with creative ideas
- 73:16
- there's so much context you could potentially give the AI model
- 73:20
- what should you prioritize telling the AI model?
- 73:23
- It turns out there's a technique that
- 73:25
- is very hopeful for driving what context you
- 73:29
- decide to give the AI model
- 73:30
- which is to iterate with the AI
- 73:33
- Let me show you what I mean
- 73:35
- If
- 73:36
- I want to ask AI to help me
- 73:38
- brainstorm plans for paying off my debt
- 73:40
- have $1,100 of credit card debt at 19% interest
- 73:45
- monthly minimum payments of $40
- 73:47
- a student loan 8% interest
- 73:49
- and a family loan $900
- 73:51
- So this gives decent background context
- 73:54
- Then one thing you could do
- 73:56
- is ask AI not to
- 73:58
- give you one option or tell you what to do
- 74:00
- but to give you
- 74:01
- multiple options to choose from
- 74:03
- I'll often ask you to give me three
- 74:05
- to five options
- 74:07
- And so the AI may
- 74:08
- come up with a few different plans
- 74:10
- Plan one is liquidity first to preserve cash
- 74:13
- Plan two is eliminate the highest interest loan
- 74:16
- Plan three is prioritize paying your family back
- 74:19
- first. So these are all actually reasonable ideas
- 74:21
- and I've not yet given it enough context
- 74:23
- to know
- 74:24
- which of these plans it should favor
- 74:27
- And it turns out that
- 74:29
- one of the really good ways to
- 74:31
- figure out what additional context to give to
- 74:33
- AI
- 74:34
- is to give it feedback
- 74:36
- on the options it presented you
- 74:38
- Highly relevant feedback
- 74:40
- that allows it to then give you
- 74:42
- the next set of options
- 74:45
- I don't like option one
- 74:46
- it's too passive
- 74:47
- I do like the idea of paying off
- 74:49
- the 19% interest loan
- 74:52
- Oh, and I forgot, I actually have $450 cash coming
- 74:55
- and I'm also moving house soon
- 74:57
- And then, with this additional context
- 74:59
- it now knows among plans one
- 75:02
- two, and three
- 75:02
- maybe what you like and what you don't like
- 75:04
- and you can ask it to create three
- 75:06
- new plans
- 75:08
- And then once again
- 75:09
- by giving a feedback on these plans
- 75:11
- you are giving additional context that will help
- 75:14
- shape the AI model's thinking
- 75:17
- And you can keep on iterating like this
- 75:19
- for a while until it comes up with
- 75:21
- a plan that you do like
- 75:22
- and then maybe have it
- 75:24
- flesh out the details of the
- 75:26
- one plan or two plans that you like
- 75:27
- the most
- 75:29
- I've found that giving feedback
- 75:31
- to the AI on what it thinks are
- 75:33
- good ideas is just a very
- 75:35
- useful mechanism for very efficiently
- 75:37
- figuring out what's helpful context to give to
- 75:39
- the AI
- 75:41
- To summarize, If you're brainstorming
- 75:44
- consider giving AI
- 75:46
- as much of the relevant context as you can in advance
- 75:49
- and then ask it for a handful of
- 75:51
- options
- 75:52
- Then give it feedback on the different options
- 75:55
- and ask it for more options
- 75:57
- and iterate multiple times
- 75:59
- get more options, get feedback, get more options, get feedback
- 76:03
- and do that a few times until
- 76:04
- you have
- 76:05
- one or more ideas that you're satisfied with
- 76:09
- If you follow this recipe for brainstorming
- 76:12
- I think you'll find you get
- 76:13
- consistently more useful and creative ideas
- 76:18
- Now
- 76:18
- you've heard me use the word context quite
- 76:20
- a few times
- 76:21
- It's important to give your AI model the
- 76:23
- right context so that it knows enough
- 76:27
- to do what you want it to
- 76:29
- Let's take a look
- 76:30
- at the next video
- 76:31
- and how context works
- 76:32
- and how it is used to produce a
- 76:34
- response
- 76:36
- According to psychologists
- 76:38
- most humans can keep only about seven things
- 76:40
- in their active working memory at a time
- 76:44
- That's why remembering a grocery list of about
- 76:46
- seven items is
- 76:47
- just barely doable if you aren't thinking about all the things
- 76:51
- but remembering a grocery list of 15 or
- 76:53
- 20 items is much harder
- 76:55
- Interestingly
- 76:57
- AI can use a large amount of context
- 77:01
- Some models can have context sizes of
- 77:04
- hundreds of thousands of words
- 77:06
- Let's see how an
- 77:07
- AI model's context works and how you
- 77:11
- could take advantage of it
- 77:12
- AI models can read and reason over
- 77:16
- very large amounts of context
- 77:19
- For example, if you are trying to choose an apartment
- 77:23
- you can upload
- 77:25
- hundreds of pages of lease contracts
- 77:28
- and upload tenant reviews and neighborhood statistics
- 77:33
- And
- 77:33
- ask AI to read all of this
- 77:36
- and to
- 77:37
- tell you the pros and cons of each
- 77:39
- of the options
- 77:40
- And you might write a prompt like
- 77:42
- pros and cons of each apartment
- 77:44
- read everything and think really hard before answering
- 77:48
- By the way
- 77:48
- telling AI to think hard or think really
- 77:51
- hard is another common prompting pattern that we'll
- 77:53
- come back to a little bit later
- 77:56
- Context refers to
- 77:57
- all the texts
- 77:58
- and files that the model
- 78:00
- uses to generate
- 78:01
- this custom response to your query
- 78:05
- If you give it a prompt like
- 78:06
- pros and cons of studying physics versus zoology
- 78:09
- then it will generate an output
- 78:12
- based on this very limited prompt
- 78:14
- very limited context that you have given it
- 78:17
- and the response will probably fairly generic
- 78:21
- But if you give it more context
- 78:23
- maybe
- 78:23
- give it your career assessment results and give
- 78:26
- it your high school schedule so it knows
- 78:28
- what class you're taking
- 78:29
- And then ask the same question
- 78:31
- All this additional context will help it to
- 78:34
- give a much more custom and likely higher
- 78:36
- quality response
- 78:38
- If you're trying to think about what context
- 78:40
- to give to the AI model
- 78:42
- think about what's all the information that the
- 78:44
- trusted advisor would need in order to
- 78:47
- think at length and reason
- 78:49
- and then give you a good answer
- 78:51
- to your question
- 78:52
- And a smart advisor that knows nothing about
- 78:54
- you but has just asked what are the
- 78:56
- pros and cons of studying physics versus zoology
- 78:59
- the best it could really do is give
- 79:00
- a pretty generic response that's not custom to
- 79:03
- you because the context
- 79:05
- in the example I'm talking
- 79:06
- just doesn't have anything specific to you
- 79:09
- AI models start with
- 79:11
- some amounts of built-in context
- 79:14
- and
- 79:15
- leading AI models today can accept
- 79:17
- maybe up to around 750,000 words as context
- 79:22
- And this
- 79:22
- corresponds to about the first four or five
- 79:25
- Harry Potter books. So
- 79:26
- there's a lot of text
- 79:28
- All to several days of continuous speech
- 79:31
- So many people
- 79:33
- underestimate how much information or how much context
- 79:36
- you can give to an AI model
- 79:39
- Now, when you ask the AI model a question
- 79:42
- by default, its context is filled with a few things
- 79:47
- First, there's something called a system prompt
- 79:49
- which usually is how the AI model
- 79:52
- knows what's the current date
- 79:54
- knows the name of the model
- 79:56
- basic capabilities
- 79:57
- maybe general instructions to be helpful to the
- 79:59
- user
- 80:00
- And then
- 80:02
- if your AI model is
- 80:04
- able to use tools like a web search engine
- 80:07
- in its context will also be
- 80:09
- written descriptions of what are these tools
- 80:12
- and how to use them
- 80:14
- such as
- 80:14
- what is a web search engine and how
- 80:16
- should it use a web search engine
- 80:18
- Before you written your prompt
- 80:20
- the context includes the system prompt and these two definitions
- 80:25
- and when you then write your prompt
- 80:27
- your prompt is added to the AI model's context
- 80:31
- and it will then
- 80:32
- use all of these things as input to generate a response
- 80:36
- like you have the options full body
- 80:38
- upper lower, split lower, back strength
- 80:41
- The input text that is the prompts you've written
- 80:43
- as well as the AI responses are called the chat history
- 80:47
- and the chat history
- 80:48
- gets incrementally
- 80:49
- added to the context of the AI model as well
- 80:53
- Now, if you start off giving the AI model more context
- 80:57
- such as write a longer prompt
- 80:59
- as well as
- 81:00
- maybe upload a handful of documents
- 81:02
- to
- 81:03
- tell it more about your workout schedule or your workout preferences
- 81:07
- then all of these files can
- 81:09
- be included into the AI model context
- 81:13
- and used to generate the response
- 81:16
- Now, if you continue the conversation to say
- 81:20
- I like this about the first plan
- 81:21
- I don't like that about the third plan
- 81:23
- what you say here is added to the AI model context
- 81:26
- and then this additional response in this type
- 81:28
- of brainstorming workflow is further added
- 81:31
- to the context. And this is why
- 81:33
- whenever you ask the AI to go back
- 81:35
- and forth or generate additional answers
- 81:37
- it knows everything that's been said so far
- 81:40
- in the conversational history
- 81:42
- Now
- 81:43
- let me take this in a different direction
- 81:46
- Imagine you had asked it for the workout plan
- 81:49
- like I shown here
- 81:50
- and it's giving you a few workout plans
- 81:53
- If you were to
- 81:54
- instead of give feedback on these plans
- 81:56
- go in a totally different direction and say
- 81:59
- now come up with a workout plan for my mom
- 82:02
- Well, a lot of the context the AI model has
- 82:05
- including your schedule, your workout preferences
- 82:08
- all that isn't really relevant for your mother's workout plan
- 82:11
- I guess
- 82:11
- unless the two of you work out together
- 82:14
- And so all this context will be distracting
- 82:16
- for the AI system
- 82:17
- and might lead it to
- 82:18
- generate a worse answer
- 82:20
- And in fact
- 82:21
- it can be hard to know whether
- 82:23
- this answer was influenced by the previous context
- 82:27
- This is why if you're going to
- 82:29
- go off onto unrelated topic
- 82:32
- it's better to start a new conversation
- 82:36
- so that
- 82:37
- you can empty out the context
- 82:39
- and
- 82:40
- start with just
- 82:41
- the new prompt or just the information
- 82:44
- that's helpful context for the new question
- 82:47
- you want to answer
- 82:49
- You've seen how context is used to produce
- 82:52
- high quality responses
- 82:54
- Thinking about what's in the AI model's context
- 82:56
- and managing that so it has
- 82:58
- just the relevant information and hopefully not too much irrelevant information
- 83:03
- although it can ignore a little bit of it
- 83:05
- that will help you get better answers
- 83:08
- from your AI
- 83:09
- One way of handling
- 83:11
- loss of context is to allow the AI
- 83:13
- model access to your computer
- 83:16
- so that it can explore relevant files and
- 83:19
- pull in relevant files into
- 83:21
- the AI model's context only as needed
- 83:25
- Let's take a look
- 83:26
- in the next video
- 83:27
- at this very powerful technique
- 83:30
- AI is moving beyond just chat interfaces
- 83:33
- You may have heard of applications like Cloud
- 83:35
- CodeWork or Microsoft Copilot CodeWork or Google Anti-Gravity
- 83:41
- These are applications that can
- 83:42
- with your permission, gather context agentically from your computer
- 83:47
- meaning they can find
- 83:48
- and read files from your computer
- 83:51
- in order to give themselves the information they
- 83:53
- need to do a task
- 83:55
- This is a new and exciting way of
- 83:56
- using AI to accomplish real work
- 84:00
- Let me illustrate
- 84:01
- a common use case for these AI desktop
- 84:04
- apps
- 84:05
- If
- 84:05
- you have been doing research on a
- 84:08
- topic and have a
- 84:10
- messy folder with lots of PDF research reports
- 84:14
- images, and so on, you can, with one of these apps
- 84:18
- ask it to
- 84:19
- read through the files in the folder
- 84:22
- and propose a new organization
- 84:24
- for it based on what it finds
- 84:27
- In this example
- 84:28
- the AI is able to look through this
- 84:29
- folder and apply a bunch of changes
- 84:32
- renaming files, moving files around
- 84:34
- creating subdirectories to come up with a much
- 84:36
- more sensible organization for this
- 84:40
- Let me show you the process of how
- 84:41
- it's actually done
- 84:43
- You might start off
- 84:45
- asking it to organize a folder first to understand what's there
- 84:51
- and it can then automatically or agentically
- 84:55
- look at how different files are named
- 84:59
- After it has
- 85:00
- figured out enough of what are the files in this folder
- 85:03
- it then comes up with an initial proposal
- 85:05
- for how to reorganize it
- 85:08
- And if you take a look at the initial proposal
- 85:12
- you may be
- 85:13
- not fully satisfied and give it a little
- 85:15
- bit further instructions to
- 85:17
- tell it what to do with this data
- 85:22
- And finally, it comes up with a refined proposal
- 85:29
- which I'm happy with
- 85:30
- so I'm going to say
- 85:31
- go ahead and carry these instructions
- 85:42
- And then it reorganizes my folder
- 85:46
- to make all the files much neater
- 85:50
- Here's how AI desktop apps work
- 85:52
- They are powered by an AI model
- 85:55
- so it comes with the AI's pre-trained knowledge
- 85:58
- And it also has a set of tools like web search
- 86:01
- so it can choose to
- 86:02
- carry out a web search if it needs
- 86:04
- to do so to accomplish the task
- 86:06
- And it has additional tools to work with
- 86:08
- files on your computer
- 86:10
- including the ability to
- 86:11
- search through files and read files and write
- 86:13
- files to move and rename files and so
- 86:15
- on
- 86:16
- So
- 86:17
- when asking a desktop app to do things in a computer
- 86:21
- a
- 86:21
- best practice workflow would be for you to
- 86:23
- tell what tasks you want done
- 86:25
- such as organize the files in the folder
- 86:28
- Let the AI system propose an action plan
- 86:30
- but not yet take action
- 86:32
- You can then review the plan
- 86:33
- give critique, maybe have it update this plan if needed
- 86:37
- and only when you're satisfied with it
- 86:38
- then tell it to execute the plan
- 86:41
- which you can go ahead and do on
- 86:42
- your computer
- 86:45
- One neat
- 86:45
- aspect of these desktop apps is that they
- 86:48
- can automatically export files and manage contacts
- 86:52
- by reading files only when needed
- 86:55
- If you are using an AI chat
- 86:58
- then
- 86:59
- you have to decide in advance what files
- 87:01
- to upload to give the AI model context
- 87:04
- So for example
- 87:05
- if you need to write a schedule for filming
- 87:07
- you might upload a file that outlines your filming procedures
- 87:13
- and based on that
- 87:14
- it can generate a schedule
- 87:16
- And notice that you had to decide in
- 87:18
- advance what context to provide to the AI
- 87:21
- But with an AI desktop app
- 87:23
- if you start the application in
- 87:25
- say, your document slash filming folder
- 87:29
- then if you tell it
- 87:30
- write a schedule for filming this week
- 87:32
- then the AI can
- 87:34
- decide to explore the files in this folder
- 87:39
- so that it can see what files are there
- 87:41
- load the relevant files
- 87:43
- and then on that basis
- 87:45
- figure out what is a good filming schedule
- 87:48
- And in this example
- 87:50
- surprise, surprise
- 87:51
- it actually noticed that
- 87:53
- One of your crew members' birthday is the week of filming
- 87:56
- and so maybe you'll
- 87:57
- fold in a celebration for your crewmate Mia
- 88:02
- One note on using these desktop apps safely
- 88:06
- Desktop apps can get access to
- 88:08
- and can edit or even delete your files
- 88:12
- And while mishaps of deleted files are pretty rare
- 88:17
- they have happened to people
- 88:19
- So I encourage you
- 88:20
- to choose the most relevant folder
- 88:22
- to run the AI desktop app in
- 88:25
- For example
- 88:26
- instead of running this in your home folder
- 88:28
- and giving this access to all your files
- 88:31
- maybe just give it access to the subset
- 88:34
- of files it really needs for tasks in
- 88:36
- a certain folder
- 88:38
- When the AI system makes a permission request
- 88:42
- I would encourage you to carefully review the
- 88:45
- permission request to make sure you know what
- 88:47
- it is reading and writing
- 88:49
- And so
- 88:50
- I'll give AI access only
- 88:52
- to the documents I want it to know
- 88:53
- about and let it write only
- 88:55
- to the files of places I want it
- 88:57
- to
- 88:58
- And when a AI desktop app deletes a file
- 89:01
- it often does not go to your recycle bin
- 89:04
- so there may not be a way to
- 89:05
- recover the file
- 89:06
- And if it edits a file
- 89:07
- it behaves a bit differently than if
- 89:09
- you were editing the documents
- 89:11
- And in particular, edited files usually don't have an edit history
- 89:15
- and so it's not possible to go back
- 89:17
- if it made some change that
- 89:18
- you may not like
- 89:20
- So until you are very familiar with these tools
- 89:23
- I'd encourage you to look carefully
- 89:25
- at the permissions requests it makes
- 89:28
- to decide what you do and do not
- 89:29
- want the AI system to be allowed to
- 89:31
- do
- 89:32
- AI desktop co-working apps are powerful tools you
- 89:35
- can use to give an AI model
- 89:37
- the ability to discover relevant context
- 89:40
- as well as to take actions such as reading
- 89:43
- writing, moving, renaming files
- 89:46
- Now
- 89:47
- we've talked about context a lot in these last several videos
- 89:51
- and
- 89:52
- using as much relevant context as possible
- 89:55
- enables your AI system to help you
- 89:58
- with what's called reasoning tasks
- 90:00
- By which I mean tasks where you want
- 90:01
- your AI to think
- 90:04
- maybe for a long time
- 90:05
- to give you the best possible answer
- 90:08
- Let's go on to the next video
- 90:10
- to see examples of reasoning with AI
- 90:14
- The latest AI models have
- 90:16
- very strong reasoning capabilities
- 90:19
- What that means is
- 90:20
- they can think rigorously
- 90:21
- and at length about the task
- 90:24
- when given the right context
- 90:26
- I find myself more and more using AI
- 90:28
- as a reasoning engine
- 90:30
- Let's take a look
- 90:32
- Here's an example
- 90:33
- where
- 90:34
- thinking for a long time can help get
- 90:36
- a better response
- 90:38
- If you are car shopping and are considering
- 90:41
- trade-offs among multiple cars
- 90:44
- You might upload spreadsheets with cars
- 90:46
- insurance plans, color codes, lots of documents, and then ask
- 90:50
- what are the trade-offs for each car?
- 90:51
- Read everything and think hard before answering
- 90:55
- The
- 90:55
- AI model may then think for
- 90:57
- quite a long time to read the documentation
- 91:00
- maybe do some online search
- 91:02
- then maybe think through what the evaluation criteria
- 91:05
- that we write for you
- 91:05
- and then generate reports on the pros and
- 91:07
- cons of different cars
- 91:10
- And just as
- 91:11
- doing research on what car to buy
- 91:13
- could involve gathering a lot of information and
- 91:16
- thinking at some length about the pros and cons
- 91:19
- AI can help you with that
- 91:22
- As AI models get better
- 91:25
- their ability to carry out long-running tasks
- 91:28
- has grown rapidly
- 91:31
- This is a study by an organization
- 91:33
- Meta, which plots for tasks at different levels of difficulty
- 91:38
- as measured by how long it'll take a
- 91:40
- human to do the task
- 91:41
- that's the vertical axis
- 91:44
- How well can AI do these tasks?
- 91:47
- So for example
- 91:47
- a task like finding a fact on the
- 91:49
- web may take a human just several seconds
- 91:52
- Summarizing
- 91:53
- a few pages of text may take a
- 91:55
- human an hour
- 91:56
- Write a blog post
- 91:58
- couple hours, audit legal documents
- 92:02
- explore a complex cybersecurity vulnerability may take a human many hours
- 92:07
- and so on
- 92:08
- And around 2024, 2025
- 92:11
- models could start to do
- 92:13
- tasks that took
- 92:14
- seconds to many seconds to
- 92:17
- many minutes or tens of minutes
- 92:19
- And in 2025
- 92:21
- models could start to
- 92:23
- have a decent success rate for doing
- 92:26
- tasks that took humans longer and longer
- 92:30
- And longer
- 92:31
- to the point where now AI models can
- 92:34
- do tasks that can take humans many
- 92:36
- many hours to do
- 92:39
- Often
- 92:39
- the AI model doesn't need
- 92:41
- 10 hours to do a task that takes
- 92:42
- a human 10 hours to do
- 92:44
- but also takes a bit longer than just
- 92:46
- a few seconds. And this is what
- 92:48
- reasoning models has enable
- 92:51
- for AI
- 92:52
- to think at great length
- 92:54
- in order to do these more complex tasks
- 92:58
- If you remember the how many hours in strawberry example
- 93:01
- that's a task that takes a human just
- 93:03
- a few seconds to do
- 93:04
- and several years ago
- 93:06
- AI models used to sometimes get this wrong
- 93:09
- And it was also in that era
- 93:11
- maybe 2023, 2024, that you may have heard advice
- 93:15
- like
- 93:16
- tell the AI model to think step by
- 93:18
- step. And back then
- 93:19
- this was good advice
- 93:20
- but this advice is largely obsolete now
- 93:23
- and I no longer tell my AI model
- 93:25
- to think step by step
- 93:27
- Instead, are more likely to just tell it to think hard
- 93:30
- and it knows what that means
- 93:31
- and that it should reason at length
- 93:33
- not necessarily step by step
- 93:35
- but in even more complex ways in order
- 93:37
- to accomplish that successfully
- 93:39
- And so rather than
- 93:41
- think about AI as counting
- 93:42
- R's and strawberries or heavily told to think step by step
- 93:46
- today you can ask AI models
- 93:48
- much more complex questions like whether the trade-offs for each car
- 93:53
- or
- 93:53
- look at all this context and help create
- 93:55
- a strip for a custom podcast
- 93:58
- So for these more complex tasks
- 94:00
- I recommend trying to use one of the
- 94:02
- more modern models if
- 94:03
- you're able to access one
- 94:05
- And there may well be
- 94:07
- models
- 94:08
- even more modern than the ones listed on
- 94:10
- the slide that could be available to you
- 94:12
- now
- 94:13
- So this is how
- 94:15
- AI
- 94:15
- reasoning or how AI thinking at length works
- 94:19
- If you ask it to
- 94:20
- plan the fastest way to visit five landmarks
- 94:23
- in Rome in one day
- 94:25
- Then it might want to gather
- 94:28
- quite a lot of information to
- 94:29
- check map distances via web search
- 94:31
- estimate walking times, search opening hours
- 94:34
- reorder the stops, and so on
- 94:36
- and then generate your optimized itinerary
- 94:40
- The reasoning process may require the AI
- 94:44
- to think at length and repeatedly gather additional information
- 94:48
- and then to think some more until it's
- 94:50
- satisfied with the answer
- 94:52
- Conceptually
- 94:53
- you can think of reasoning
- 94:55
- as a process like this
- 94:57
- Giving you a prompt in other input contexts
- 94:59
- it will reason or think for a while using that context
- 95:03
- and then depending on where it gets to
- 95:06
- maybe it'll decide it's done
- 95:07
- and then just give you the final answer
- 95:09
- Alternatively, after thinking for a while
- 95:12
- it may decide that it needs a user
- 95:14
- tool to gather more information
- 95:17
- maybe via web search
- 95:18
- or maybe by reading more files from your computer
- 95:21
- if it's a desktop app
- 95:24
- And then use that additional context to reason
- 95:26
- longer
- 95:28
- until it either again decides to use the
- 95:30
- tool to gather more information
- 95:32
- or decides it's done
- 95:34
- And so it can go
- 95:35
- through a few rounds of gathering more information
- 95:37
- and reasoning longer before it decides the answer
- 95:39
- is good enough to present to you
- 95:42
- If you
- 95:43
- are
- 95:43
- working on a complex task and you want
- 95:45
- the model to think at length
- 95:48
- then
- 95:49
- one way to do so
- 95:50
- is to just tell the model to think
- 95:53
- A few of the interfaces among the
- 95:55
- popular AI model providers will have a thinking option
- 95:58
- and if you select that option
- 96:00
- that's a cue to the model that you
- 96:02
- want it to think longer
- 96:04
- Alternatively, in your prompt
- 96:06
- you can also just tell it to think
- 96:07
- really hard about this
- 96:08
- and then usually obey your instructions
- 96:11
- or some people will use the phrase ultra think
- 96:14
- and that's another keyword that the models
- 96:16
- understand as a cue to
- 96:18
- you know, think really hard about it
- 96:21
- And if you do so
- 96:22
- and the AI model will sometimes think for
- 96:24
- many tens of seconds or even minutes
- 96:26
- or maybe sometimes even over 10 minutes in
- 96:28
- order to give you a good answer
- 96:32
- Especially with reasoning or with thinking models
- 96:36
- I'd also encourage you to try giving the
- 96:37
- model hard tasks to see what it can
- 96:39
- do for you
- 96:41
- So if you are building a startup
- 96:44
- maybe give it lots of context on what you're doing
- 96:47
- and maybe tell it to design a top-down
- 96:49
- plan for a four-person startup with limited cash
- 96:51
- I encourage you to give the AI model real job tasks
- 96:56
- real problems that you want to think through
- 96:58
- or that you want to solve
- 97:00
- and as part of setting up for success
- 97:03
- try to give it all the context that
- 97:05
- a human expert would need in order to
- 97:06
- complete the task to make sure that the
- 97:08
- AI model
- 97:09
- has
- 97:10
- the sufficient information that
- 97:12
- really anyone would need to do the task
- 97:13
- successfully
- 97:15
- To wrap up
- 97:16
- if you want to use
- 97:18
- AI to reason at length about complex problems
- 97:22
- I encourage you to use the best models
- 97:24
- available
- 97:25
- The best models are
- 97:27
- often better than models that are
- 97:30
- maybe six to twelve months older
- 97:32
- Remember to give it as much context as
- 97:35
- is needed to carry out the task
- 97:38
- It's okay to try giving a hard task
- 97:40
- to see what it can do
- 97:41
- so don't just give a trivial task
- 97:43
- And lastly
- 97:44
- either select thinking mode or
- 97:46
- just tell it in the prompt
- 97:47
- to think hard
- 97:49
- We've spoken about when it's helpful to use the cutting-edge models
- 97:53
- but
- 97:54
- it turns out
- 97:55
- even the most up-to-date models have some common issues
- 97:59
- one being the tendency
- 98:00
- to tell you whatever
- 98:02
- it thinks you want to hear
- 98:04
- This is called sycophancy
- 98:06
- Let's go on to the next video
- 98:08
- to see how to manage this behavior of
- 98:10
- AI models
- 98:12
- AI models will act in ways to
- 98:14
- try to please you
- 98:16
- Because of the way they've been trained
- 98:18
- they have a strong bias to tell you
- 98:20
- what you want to hear
- 98:22
- This is called sycophancy
- 98:24
- And avoiding it
- 98:25
- is a key prompting skill
- 98:27
- and involves prompting neutrally and keeping context factual
- 98:32
- Let's take a look at how to avoid
- 98:33
- sycophancy
- 98:34
- because this will help you get much better
- 98:36
- answers from your AI
- 98:38
- If you are
- 98:39
- constrained the pros and cons of remote work versus in-office work
- 98:43
- and you ask it
- 98:44
- don't you think remote work is better than
- 98:46
- office work?
- 98:47
- This word in the question
- 98:49
- gives away what you're hoping the answer is
- 98:51
- and AI will probably say
- 98:52
- yes, remote work offers many advantages
- 98:55
- In contrast, if you were to ask
- 98:58
- is it true that
- 98:59
- office work is more productive?
- 99:01
- Then
- 99:02
- it will probably agree with you
- 99:03
- that office work has these strong benefits
- 99:07
- So, depending on how you ask the question
- 99:10
- it will tend to reinforce your own preferences
- 99:13
- your own biases
- 99:15
- and this may not be the most helpful
- 99:17
- thing for you to make objective fact-based decisions
- 99:21
- In a study by the Washington Post on ChatGPT responses
- 99:26
- it was much more likely to respond with phrases like
- 99:29
- that's correct, good point, you're on the right track
- 99:32
- compared to not quite right
- 99:33
- that's not the case
- 99:34
- or actually. And in fact
- 99:36
- it tended to agree strongly
- 99:38
- about 10 times more than it disagreed
- 99:41
- And some models from ChatGPT have said things like
- 99:45
- dude, you just said something deep about even flinching
- 99:48
- you're a thousand percent right
- 99:50
- And while some users appreciate AI agreeing with them like this
- 99:54
- I personally don't find it that useful to
- 99:56
- have AI just tend to agree with whatever I say
- 99:59
- even when I'm not right
- 100:01
- While leading AI model companies are working to reduce sycophancy
- 100:05
- it is still a problem
- 100:07
- Models are trained to be hopeful assistants using human feedback
- 100:11
- and this reinforces sycophancy
- 100:14
- For example, if you ask an AI model
- 100:16
- I feel like it's better to be an introvert
- 100:18
- don't you?
- 100:20
- If the AI model responds
- 100:22
- that's an interesting idea
- 100:23
- here's why I tend to agree
- 100:25
- then most people are more likely to hit
- 100:27
- thumbs up on the feedback button because
- 100:29
- you know, it's a nice answer
- 100:30
- makes you feel good
- 100:31
- But if AI were to say
- 100:33
- not necessarily, both types, introverts and extroverts, carry road trade-offs
- 100:38
- you know, people just don't feel as good about that answer
- 100:41
- and so they're less likely to hit thumbs
- 100:43
- up or may even hit thumbs down
- 100:46
- Because of this type of feedback
- 100:48
- AI
- 100:49
- which has been trained to
- 100:50
- generally more answers that leads to thumbs up positive feedback
- 100:54
- will learn to
- 100:56
- try subtly
- 100:57
- to agree with people more often than not
- 100:59
- And this leads to circumvention
- 101:02
- And circumvention feels hopeful
- 101:05
- but actually degrades answer quality
- 101:09
- Sometimes maybe circumvention is easier to spot
- 101:12
- If you were to say
- 101:13
- I'm really proud of this essay
- 101:14
- what do you think?
- 101:15
- Well, I'll probably agree with you
- 101:18
- But other times it's harder to detect circumvention
- 101:22
- If you were to say
- 101:24
- analyze this data
- 101:25
- and find all the positive measures that performance this quarter
- 101:29
- you're subtly signaling that you're looking for positive
- 101:31
- measures of the company's performance
- 101:34
- And so it's more likely to say something like
- 101:36
- data clearly shows revenue growth
- 101:38
- strong intention, improving margins
- 101:40
- and less likely to point out problems
- 101:44
- To avoid circumvention
- 101:45
- try to use neutral framing
- 101:47
- of your questions
- 101:49
- and avoid
- 101:50
- giving any hints as to what is the
- 101:52
- answer you want to hear
- 101:54
- So, for example, if I were to ask
- 101:57
- aren't carbon taxes bad for small businesses?
- 102:00
- I'm really telling it what answer I'm hoping for
- 102:04
- In contrast, a more neutral prompt would be
- 102:07
- to what extent, if at all
- 102:08
- do carbon taxes affect small businesses?
- 102:11
- If someone reads the question on the right
- 102:13
- it's actually not clear what answer they want
- 102:15
- so it's harder for AI model
- 102:17
- to be sycophantic
- 102:19
- Or if you were to ask
- 102:20
- do you agree that AI will create a
- 102:22
- lot of jobs?
- 102:23
- I actually happen to agree
- 102:24
- but if I'm actually doing research
- 102:25
- I don't want it to just
- 102:26
- tell me what I want to hear
- 102:29
- Instead, you can ask
- 102:30
- what does
- 102:31
- current research say about AI's effect on jobs?
- 102:35
- Or instead of asking
- 102:36
- does remote work reduce worker productivity?
- 102:39
- Maybe ask
- 102:41
- how does productivity compare between remote and in-office
- 102:44
- work?
- 102:45
- One common pattern is to lay out two options
- 102:48
- such as remote and then office work
- 102:50
- and just ask it for pros and cons so they compare
- 102:53
- but without hinting
- 102:54
- which of the two options
- 102:56
- I am hoping will come out ahead
- 102:58
- So, don't you think that was the best video ever?
- 103:02
- Oh
- 103:03
- what do you think of the pros and
- 103:04
- cons of the videos you just saw?
- 103:07
- Sometimes it's nice to be told what you want to hear
- 103:09
- but generally
- 103:10
- that won't help you to do better work
- 103:13
- Second fantasy, despite a lot of attempts to combat it
- 103:16
- is still one of the pervasive issues with
- 103:18
- practical AI usage today
- 103:21
- And so implementing the strategies from this video
- 103:24
- especially taking more neutral framing
- 103:26
- will help you get more objective and valuable
- 103:29
- feedback from AI models
- 103:31
- Now
- 103:32
- one of the most common tasks that people
- 103:34
- use AI for is writing
- 103:37
- Let's go on to the next video
- 103:38
- to see
- 103:39
- how to work with AI models to help
- 103:42
- with your writing
- 103:44
- In a study by OpenAI
- 103:46
- writing accounted for 24% of tasks that people
- 103:50
- asked ChatGPT to do. This is the single
- 103:52
- largest group of tasks
- 103:54
- Writing is really a kind of thinking
- 103:57
- I find that when I'm writing
- 103:58
- I have to think
- 103:59
- And so AI, which is really good at thinking and reasoning
- 104:03
- can help you with this
- 104:05
- But
- 104:05
- just asking AI to write for you
- 104:07
- often leads to AI slop
- 104:10
- or writing that sounds like AI writing
- 104:13
- which somehow feels different than human thoughtful writing
- 104:18
- Let's take a look at some techniques
- 104:19
- for getting AI to write effectively for you
- 104:23
- and how to take advantage of AI reasoning
- 104:26
- and avoid AI slop
- 104:29
- What makes AI slop?
- 104:31
- Many people have noticed that AI writing
- 104:33
- often includes the M dash
- 104:35
- that is the long dash
- 104:37
- much more than normal human writing
- 104:40
- On the social media site Blue Sky
- 104:43
- the use of the N-dash has been trending
- 104:46
- upward ever since the release of ChatGPT
- 104:49
- A recent survey
- 104:50
- showed that 40% of US-based employees
- 104:53
- have recently received work
- 104:55
- slop in the last month
- 104:57
- And the term
- 104:58
- AI slop
- 104:59
- refers to content that's generated by AI
- 105:02
- and that looks good if you don't read it too carefully
- 105:05
- so maybe every sentence read in isolation sounds like it's well-written
- 105:09
- but collectively the text just lacks substance
- 105:12
- It feels like it was written without
- 105:14
- much deep or careful thought
- 105:17
- One of the properties of AI slop
- 105:19
- is it often contains sentences like this
- 105:23
- but it does change everything
- 105:25
- It's kind of vague
- 105:26
- empty-sounding, but also somehow overly important-sounding text
- 105:31
- AI's distinctive writing style comes from
- 105:34
- certain words and patterns that
- 105:36
- it tends to overuse
- 105:38
- AI tends to use fewer
- 105:40
- unique words and over-represents some words and phrases
- 105:45
- For example, AI tends to overuse the words nuanced and delve
- 105:50
- and it tends to
- 105:51
- use lists of three more than most people will
- 105:55
- and it tends to use fewer nouns
- 105:57
- leaving the
- 105:58
- phrases like this is a robustly structured and
- 106:00
- highly insightful paper
- 106:02
- And this not X but Y is not just about speed
- 106:05
- it's about availability
- 106:08
- In fact
- 106:09
- I've been noticing on social media a lot
- 106:11
- more of this not X but Y type of verbiage
- 106:15
- often with X and Y
- 106:16
- both vague things
- 106:18
- Like, it's not about infrastructure
- 106:20
- it's about architecture
- 106:23
- And a lot of phrases like those are
- 106:24
- just vague and don't reflect
- 106:27
- a deeply insightful point of view
- 106:30
- Interestingly, because humans are using AI models so much
- 106:34
- humans are themselves starting to sound more like AI
- 106:38
- And humans, ever since ChatGPT was released
- 106:41
- are using the word delve more in podcasts
- 106:44
- and talks
- 106:46
- And this is true both for spontaneous speech
- 106:49
- as well as for prepared speeches
- 106:50
- So this isn't just a case of people
- 106:52
- using AI to write scripts for them
- 106:55
- It looks like people are picking up speech
- 106:57
- patterns or texting patterns from spending a lot
- 107:00
- of time with AI
- 107:02
- So
- 107:03
- what's a better way to write that
- 107:05
- a voice generating AI slop?
- 107:08
- One technique that I think you find
- 107:10
- helpful is to use progressive outlining
- 107:14
- in which you don't ask AI to write
- 107:16
- the final text right away
- 107:18
- but instead have it write an outline
- 107:20
- refine the outline
- 107:21
- and
- 107:21
- iterate a few times before having it generate
- 107:23
- the final text
- 107:25
- For example, here's a prompt that says
- 107:27
- I'm writing an article
- 107:29
- about small AI teams moving faster than large
- 107:31
- teams that don't use AI
- 107:33
- And
- 107:34
- I just want to acknowledge that this is
- 107:35
- not a neutral prompt where I'm asking AI
- 107:37
- to help me decide if small AI teams do move faster
- 107:41
- but instead this is writing an article from
- 107:43
- a certain point of view
- 107:45
- But you can ask AI to research evidence
- 107:47
- for and against this hypothesis
- 107:50
- So the AI model may
- 107:52
- search online and find a handful of articles
- 107:56
- Next
- 107:57
- if you wanted to help you brainstorm a
- 107:59
- handful of options for the story outline
- 108:02
- you can tell it to brainstorm or to
- 108:03
- create three different outline options
- 108:06
- Maybe tell it to include a counter-argument section
- 108:10
- and also provide or upload a handful of
- 108:13
- stories from AI teams
- 108:15
- say, that you work with
- 108:17
- With this input context
- 108:19
- the AI may give back
- 108:21
- a few different options for what the
- 108:23
- outline for your article might look like
- 108:25
- Option one could be to tell the three
- 108:27
- stories and then conclude with a thesis
- 108:30
- Option two
- 108:31
- might explore different patterns of how AI teams
- 108:33
- work and so on
- 108:35
- Then following what you saw in how to brainstorm of AI
- 108:39
- you might give feedback on these options
- 108:42
- And so you might say
- 108:43
- let's use option one and keep all the stories
- 108:45
- remove the thesis right after story one
- 108:48
- And also maybe say you want to add a historical analogy
- 108:51
- Pixar creating the Toy Story in the 90s
- 108:54
- which is actually a really inspiring story where
- 108:56
- Pixar at that time a small company created Toy Story
- 109:01
- which was the first fully computer animated feature length film
- 109:05
- just using a very small team
- 109:07
- Based on your feedback
- 109:09
- the AI
- 109:10
- maybe gives you back a revised outline
- 109:13
- If you're satisfied with this outline
- 109:16
- then you can tell it to expand each heading
- 109:19
- not even into the final text
- 109:21
- but just into bullet points
- 109:24
- And following that
- 109:26
- you might decide to
- 109:27
- give it more feedback on the bullet points
- 109:29
- and iterate on the bullet points
- 109:32
- before finally having it generate the text for
- 109:34
- the article
- 109:37
- And it turns out that
- 109:39
- starting with an outline speeds up review
- 109:42
- Let's say you're working on
- 109:44
- a fun article about whether a fly squirrel
- 109:46
- can carry a coconut
- 109:48
- In contrast, to having AI first work review on the outline
- 109:51
- then on bullet points
- 109:52
- and then on the final text
- 109:54
- maybe you ask AI to write the final text right away
- 109:58
- just from the start
- 110:00
- If it writes a sentence like this
- 110:02
- you may be unhappy with a few words
- 110:04
- and you can edit a few words
- 110:07
- but changing each word just changes one word
- 110:10
- and the rest of the paragraph stays the same
- 110:13
- In contrast, if you write an outline first
- 110:16
- and you're unhappy with part of the outline
- 110:19
- then changing the outline causes an entire section
- 110:23
- that's a lot of words
- 110:24
- of the final article to change
- 110:27
- So that's why editing the outline
- 110:29
- or iterating with the AI system on the
- 110:31
- outline is very high leverage
- 110:33
- because you can figure out how to change
- 110:35
- just a few words of the outline
- 110:37
- and this will result in an entire paragraph
- 110:40
- or entire section of your final article changing
- 110:43
- And this
- 110:44
- ends up being a much more efficient way
- 110:46
- for you to
- 110:47
- think through what you want to say in
- 110:48
- an article
- 110:49
- and adapt it
- 110:51
- to what you want it to be
- 110:53
- Writing is one of those common use cases
- 110:55
- of AI
- 110:57
- According to OpenAI's data of their writing-focused chats
- 111:01
- about two-thirds involve starting from
- 111:03
- some pre-existing text rather than starting from scratch
- 111:06
- or starting from an empty sheet of paper
- 111:09
- When you already have something written up
- 111:11
- it can be very helpful to have AI
- 111:12
- critique it for you and help you make
- 111:14
- it better
- 111:16
- Let's take a look at some techniques for
- 111:17
- this in the next video
- 111:20
- You often have
- 111:21
- some idea where you've already written some text about your idea
- 111:25
- but want an AI model to help you
- 111:27
- edit and refine your text
- 111:29
- I often show my writing to AI to
- 111:31
- help me make it better
- 111:33
- AI is great at this task
- 111:35
- and it always assigned to read your work
- 111:38
- whereas
- 111:38
- finding a human to help you out might
- 111:40
- be trickier
- 111:41
- You've already learned how to avoid sycophancy
- 111:44
- but how do you get the best quality
- 111:46
- editing and critique?
- 111:48
- Let's take a look
- 111:50
- One useful technique for editing with AI
- 111:52
- is to edit your article piece by piece
- 111:55
- such as one sentence at a time or
- 111:57
- one paragraph at a time
- 111:58
- rather than telling it to edit the entire
- 112:00
- article all at once
- 112:02
- And to do a little brainstorming
- 112:04
- around each paragraph until I've nailed down
- 112:06
- one paragraph before going on to the next one
- 112:10
- For example, if someone's written this sentence
- 112:12
- the public thinks achieving AGI
- 112:14
- artificial general intelligence
- 112:16
- means computers would be as smart as people
- 112:18
- You might ask AI to help brainstorm a
- 112:20
- few different ways to say this
- 112:22
- and so it may
- 112:24
- come up with a punchy way to say this
- 112:27
- a visionary way to say this
- 112:28
- a conversational way to say this
- 112:30
- and depending on your editing goals
- 112:32
- you could even iterate a little bit until
- 112:35
- you pick
- 112:36
- some version of rephrasing this part of the
- 112:38
- text that you like
- 112:40
- After you've nailed this down
- 112:42
- then
- 112:42
- go on to the
- 112:43
- second sentence or maybe the second paragraph
- 112:46
- And work on that little bit of the AI
- 112:49
- and then go on to the next sentence
- 112:50
- the next paragraph
- 112:51
- and so on until you get through your
- 112:53
- entire article
- 112:55
- And I find that
- 112:56
- working on one piece of a long article
- 112:58
- at a time
- 112:59
- makes for a much more manageable workflow than
- 113:01
- if
- 113:01
- it were to change a lot of things
- 113:03
- all at the same time
- 113:04
- and then you're reading this
- 113:05
- very long edited article to figure out what
- 113:07
- has changed and what you like and don't like
- 113:10
- Now, if you want high-level
- 113:12
- more holistic feedback about an entire piece you've written
- 113:17
- it turns out AI can help with that
- 113:18
- too
- 113:19
- But because of sycophancy
- 113:22
- AI is often not a very good objective critic
- 113:26
- For example
- 113:26
- if you wrote a
- 113:27
- sci-fi short story about an astronaut stepping out of his ship
- 113:32
- and if you ask AI without further instructions to critique it
- 113:36
- there's a good chance they'll tell you whether
- 113:38
- you did this fantastic work
- 113:40
- In contrast
- 113:41
- there's a very hopeful technique to guide AI
- 113:44
- in how To evaluate your work to give
- 113:46
- you more helpful critical feedback
- 113:50
- And that is to give it a rubric
- 113:52
- and that means a grading criteria
- 113:55
- For example
- 113:56
- you might write a rubric
- 113:58
- that specifies what are the most important criteria
- 114:01
- by which
- 114:02
- to grade or to judge the work
- 114:04
- So
- 114:05
- you may say that
- 114:06
- characters of the story
- 114:08
- is worth 25 points out of 100 to plot 25 points
- 114:12
- world building, writing craft, and establish a point system
- 114:18
- and then also
- 114:19
- develop
- 114:20
- detailed instructions on how to evaluate
- 114:23
- each of these criteria
- 114:25
- So to evaluate characters
- 114:27
- maybe ask if every named character has a goal
- 114:30
- and that's worth 10 points
- 114:31
- conflict between two characters goals
- 114:33
- and so on
- 114:35
- And giving AI
- 114:36
- very specific criteria on how to judge work
- 114:39
- forces the AI to be more objective
- 114:43
- One thing to notice about these criteria is
- 114:45
- that each of them
- 114:46
- is very clearly and unambiguously defined
- 114:49
- So for each text
- 114:51
- each of these criteria is either true or false
- 114:54
- yes or no, and there's nothing in between
- 114:56
- So either it's true that every named character has a goal
- 115:00
- or it's not true
- 115:01
- And these completely objective criteria
- 115:05
- forces AI to
- 115:06
- look at whatever you're giving it through a objective
- 115:10
- well-specified standard with no ambiguity
- 115:14
- And by the way
- 115:15
- if you're not sure what rubric or what
- 115:17
- grading criteria to use
- 115:19
- you can brainstorm with AI to develop that rubric
- 115:22
- and AI is actually pretty good at this
- 115:23
- too
- 115:24
- After you've written the rubric
- 115:26
- you can then provide the rubric
- 115:28
- as well as the story
- 115:30
- and in your prompt
- 115:32
- ask the AI to be objective
- 115:33
- So, critique the attached sci-fi story
- 115:36
- assign the score per category
- 115:37
- then sum the scores at the end
- 115:40
- And
- 115:40
- by giving the AI these very clear instructions
- 115:43
- on what to do to sum the scores in the end
- 115:45
- it then hopefully gives a more
- 115:48
- objective assessment of your story
- 115:51
- If you want
- 115:52
- you can also ask it to then give
- 115:54
- you suggestions on how to improve the story
- 115:56
- to do better on this rubric
- 115:58
- and does it cause it
- 115:59
- to give you more focused suggestions to improve
- 116:02
- it in the dimensions that you think matter
- 116:04
- the most
- 116:06
- In contrast
- 116:07
- poorly written rubrics encourage circumfancy and
- 116:10
- ambiguous or less objective thinking
- 116:13
- For example, if you say
- 116:15
- I've been working on this sci-fi story
- 116:16
- please score it out of 100
- 116:19
- One of the things that's strange about this
- 116:20
- prompt is you first ask it to score out of 100
- 116:23
- so that will tend to cause it to
- 116:25
- leap the conclusion about what score
- 116:27
- and then only after that assign a score per category
- 116:31
- characters, plot, world building, and writing craft
- 116:33
- And these are
- 116:34
- invariably defined categories because we've not told it
- 116:38
- how the score characters plot and so on
- 116:41
- And so this will tend to cause the
- 116:42
- AI model to
- 116:44
- first come up with some score and then justify it
- 116:47
- rather than
- 116:48
- score it carefully according to rubric and then
- 116:50
- add it up
- 116:51
- to then come up with a more thoughtful
- 116:53
- score
- 116:54
- And
- 116:55
- as you see in the practice lab
- 116:57
- to come at the end of this module
- 116:59
- this type of rubric will tend to
- 117:01
- give higher scores than more objective rubrics
- 117:04
- We've talked about using AI to critique
- 117:07
- and help give suggestions for improving your work
- 117:10
- It turns out that having AI
- 117:12
- critique
- 117:13
- his own work
- 117:13
- or having one AI model critique a different
- 117:16
- AI model's work can also help improve the results
- 117:19
- For example
- 117:20
- if you ask ChatGPT
- 117:22
- to write a user manual for a fancy role-playing game
- 117:26
- Then it might generate a file for you
- 117:29
- And one thing you could do
- 117:30
- is provide a rubric for ChatGPT to critique his own work
- 117:34
- but the neat technique is to find a different AI model
- 117:38
- maybe Gemini
- 117:39
- and give that a grading rubric
- 117:42
- to have Gemini critique ChatGPT's work
- 117:45
- or vice versa
- 117:47
- And it turns out that this type of cross-model review
- 117:50
- where you have one model review a different model's output
- 117:54
- it holds integrated bit of knowledge from
- 117:56
- the two
- 117:57
- different models and can result in
- 117:59
- slightly better results than if you were to
- 118:01
- ask one model to critique its own results
- 118:05
- I think using
- 118:06
- multiple models in this context might give only
- 118:09
- a slight boost in performance
- 118:10
- Realistically
- 118:11
- if you ask ChatGPT
- 118:13
- to review his own results or ask Gemini
- 118:15
- to review his own results
- 118:16
- I think that will actually do just fine
- 118:18
- But sometimes I find it
- 118:20
- reassuring if a totally different model
- 118:23
- judges the output of a different AI model
- 118:26
- And I use this technique only rarely myself
- 118:30
- but one thing I do do is frequently
- 118:32
- switch between different AI models
- 118:35
- It turns out that AI models are advancing rapidly
- 118:39
- and at different moments in time
- 118:40
- different models will do better on different tasks
- 118:44
- And so
- 118:44
- routinely trying out different models will help keep
- 118:48
- you sharp and keep honing your intuition about
- 118:50
- what model is best for what task
- 118:53
- AI models have what's called jagged intelligence
- 118:57
- If the circle represents the task of what people can do
- 119:01
- maybe in a job or maybe in a personal context
- 119:04
- it turns out that
- 119:05
- AI can do some things better than any person
- 119:09
- like
- 119:09
- quickly read tons of web pages or solve
- 119:11
- tricky math problems
- 119:13
- But there are also many tasks that AI
- 119:15
- doesn't do as well as people
- 119:18
- So there's some tasks where AI
- 119:19
- does poorer than human and some where it
- 119:21
- does much better than human
- 119:24
- And
- 119:25
- different AI models are jagged in different ways
- 119:28
- So the tasks different AI models can do
- 119:30
- well are different
- 119:33
- Moreover, the marketplace of AI models is highly competitive
- 119:37
- So ChatGPT, Cloud, Gemini
- 119:40
- really the long list of model providers are
- 119:42
- releasing better models all the time
- 119:44
- And so the best model for your task
- 119:47
- will likely change rapidly over time
- 119:50
- And so
- 119:50
- I find that I'll often take the same
- 119:52
- prompt
- 119:53
- and feed it to multiple different models to
- 119:55
- see how they compare
- 119:56
- and just continuously hones my intuition
- 119:59
- about what models are best for which of
- 120:01
- the tasks I care about
- 120:03
- So that takes us to
- 120:05
- almost the end of this module
- 120:07
- I hope you've seen that AI models are
- 120:09
- really useful for reasoning tasks
- 120:11
- as well as for brainstorming
- 120:13
- writing, editing, and critiquing your work
- 120:17
- These are powerful ways of using AI as
- 120:19
- a thought partner
- 120:21
- that I found very useful in my own work
- 120:23
- and that I'm confident you will too
- 120:26
- Let's go on to the next video
- 120:27
- to explore the practice lab for this module
- 120:31
- I hope you practice the techniques we talked
- 120:33
- about for using AI as a thought partner
- 120:36
- In the upcoming practice lab
- 120:39
- you can explore once again
- 120:41
- side by side
- 120:42
- more effective and less effective strategies for both
- 120:45
- brainstorming and AI critique
- 120:48
- Let's take a look
- 120:50
- Once again, when the lab pops up
- 120:52
- you can dismiss the tutorial
- 120:56
- You can also find the tutorial here
- 120:58
- as well as the instructions I hope you
- 121:00
- follow over here
- 121:03
- And the buttons here
- 121:06
- similar to the previous lab
- 121:07
- allow you to enter different prompts
- 121:11
- To brainstorm
- 121:12
- here is a prompt of less context
- 121:14
- I need a workout plan
- 121:15
- 30 years old, want to get stronger
- 121:17
- whereas this is a more detailed prompt
- 121:20
- with more helpful context
- 121:22
- And so you can run it like so
- 121:24
- and see the difference in the outputs
- 121:29
- So
- 121:29
- let me focus on this more detailed example
- 121:31
- on the right. It's given a program one
- 121:35
- which looks pretty reasonable
- 121:36
- and a program two
- 121:38
- which also looks pretty reasonable
- 121:41
- and
- 121:41
- we have also a few
- 121:43
- suggested prompts for how you might refine it
- 121:45
- So maybe I'm drawn to this
- 121:48
- Let me do that
- 121:51
- And have it continue the conversation
- 121:54
- Then it will take this feedback
- 121:56
- on the programs it has given you
- 121:59
- in order to refine its results
- 122:02
- By the way
- 122:03
- I actually
- 122:04
- use a workout program that AI had helped me to generate
- 122:08
- and
- 122:08
- Found this helpful in my personal life or
- 122:10
- thinking through my regular workout plan
- 122:14
- And so if you're a fan of working out
- 122:16
- maybe you'll find
- 122:17
- some of the suggestions it makes useful as
- 122:19
- well
- 122:21
- Let's go back to the homepage
- 122:23
- Here are a couple more examples of brainstorming
- 122:26
- prompts with a simple prompt
- 122:27
- I have a thousand dollars to invest
- 122:28
- what should I do?
- 122:30
- Versus giving more context and then what are
- 122:32
- some options I should consider
- 122:34
- Hope you try this out as well
- 122:36
- As well as
- 122:37
- some examples of critiquing a sci-fi story
- 122:43
- As well as
- 122:44
- this example of an objective
- 122:46
- rubric on the right. And I hope you
- 122:48
- bring this up and
- 122:49
- read through it to
- 122:51
- give yourself a sense of what a
- 122:53
- well-written objective rubric looks like
- 122:58
- And compare that with what a
- 123:01
- more subjective rubric looks like
- 123:07
- And if you hit compare
- 123:09
- you see
- 123:10
- the difference in the quality of these reviews
- 123:13
- And in fact, maybe not surprisingly, with the less objective rubric
- 123:18
- it gets an 85 out of 100
- 123:20
- but a more objective rubric
- 123:21
- a 75 out of 100
- 123:25
- It also gives more helpful suggestions how to
- 123:28
- improve your story
- 123:30
- In addition to critiquing and improving a sci-fi story
- 123:33
- you can also take a look at these examples
- 123:35
- helping you improve a cover letter for applying for a job
- 123:40
- as well as critiquing and help you to
- 123:41
- improve a possible business plan
- 123:44
- After checking out these five examples
- 123:47
- please also try your own prompts
- 123:49
- or
- 123:49
- try your own stories or cover letters or
- 123:51
- business plans or something else
- 123:53
- and I encourage you to
- 123:55
- use this interface to play with different
- 123:58
- ways to brainstorm or to critique your writing
- 124:04
- So, that's it for this module
- 124:07
- Please enjoy exploring the lab
- 124:10
- Next
- 124:11
- let's go on to the final module
- 124:13
- where you see applications beyond text
- 124:17
- Specifically, you look at multimodal prompting
- 124:20
- in which you get your AI to also
- 124:22
- use images and audio
- 124:24
- and also look at building applications like games
- 124:28
- It'll be a lot of fun
- 124:29
- and one of the examples we see
- 124:31
- will have something to do with fireworks
- 124:34
- So that, I'll see you there
- 124:46
- In the previous two modules
- 124:48
- we've mostly had AI generate text
- 124:51
- but AI can produce richer types of outputs as well
- 124:54
- like images, videos, and so on
- 124:56
- We call this multimodal output
- 124:59
- that is, outputs with multiple modalities
- 125:03
- Prompting for multimodal output is a bit different
- 125:07
- because multimodal interactions are slower and more costly
- 125:12
- But
- 125:12
- these capabilities will let you get a lot
- 125:14
- more done with AI
- 125:16
- Additionally, we'll also look at multimodal inputs
- 125:20
- specifically
- 125:21
- if you want to show some images to
- 125:22
- your AI and have it reason about that
- 125:25
- Let's dive in
- 125:27
- AI models can generate images
- 125:29
- videos, voices, even music, code, and more
- 125:33
- You might have seen AI generate various fun
- 125:36
- creative images. I want to share with you
- 125:38
- one example of AI image generation that
- 125:41
- I really enjoyed
- 125:43
- For my daughter Nova's seventh birthday
- 125:46
- I wanted a unique cake design
- 125:49
- and she loves cats
- 125:51
- so we used
- 125:52
- AI image generation to explore different cake designs
- 125:56
- The leftmost image here
- 125:58
- is an AI-generated image created with a
- 126:01
- AI generation software called NanoBanana
- 126:04
- which is created by Google
- 126:06
- And this is a picture that my daughter
- 126:09
- really liked. She wanted a cake that looks
- 126:10
- like this generated image
- 126:13
- We then
- 126:14
- took the image
- 126:15
- and showed it to Baker
- 126:17
- and asked the Baker to
- 126:18
- render this picture into a real-life 3D cake
- 126:23
- And the picture on the right shows my
- 126:25
- daughter cutting this birthday cake that she loved
- 126:28
- So in this case
- 126:30
- image generation wound up being a
- 126:33
- brainstorming tool
- 126:35
- to explore different cake designs until we found
- 126:38
- one
- 126:38
- that turned into a
- 126:40
- real-life 3D cake that we all ate and
- 126:42
- liked
- 126:44
- In addition to generate images
- 126:46
- I really enjoy playing with AI for video
- 126:48
- generation as well
- 126:50
- Here's a fun video generated by our team
- 126:53
- of man shrinking
- 127:01
- That video would previously have required maybe expensive special effects
- 127:05
- but now AI can just generate it
- 127:09
- AI can also generate voices
- 127:11
- Here's a voice clone of me
- 127:13
- reading out loud a letter from The Batch
- 127:15
- which is a weekly newsletter
- 127:17
- that
- 127:18
- Deep Learning.ai publishes to cover what matters in AI
- 127:22
- Dear friends
- 127:23
- here's the latest from this week's issue of
- 127:25
- The Batch
- 127:26
- A barrier to faster progress in generative AI is evaluations
- 127:31
- evals
- 127:32
- particularly
- 127:33
- of custom AI applications that generate free-form text
- 127:37
- By the way
- 127:38
- I've played audio of my voice clone to
- 127:41
- both of my parents
- 127:42
- and it turns out
- 127:43
- one of my parents could tell it wasn't me
- 127:45
- and one of my parents could not tell
- 127:47
- it was me or my voice clone
- 127:50
- And to avoid me getting into trouble
- 127:51
- I will not reveal which of my parents got it wrong
- 127:54
- but I think voice clones are getting really good
- 127:57
- Lastly, AI can generate code
- 128:01
- I mentioned my daughter loves cats
- 128:03
- she also loves the color yellow
- 128:06
- and so
- 128:06
- when her teacher mentioned that
- 128:08
- she wished kids in the class could type
- 128:11
- or keyboard a little bit faster
- 128:13
- I use AI to generate this typing game
- 128:16
- where if my daughter hits the right letter
- 128:18
- then
- 128:18
- she sees this fun little animation of a cat being fed
- 128:22
- which she loves
- 128:24
- Using AI to write code has made it
- 128:26
- easier and more accessible for everyone
- 128:29
- including you if you wish
- 128:30
- to write at least basic computer programs
- 128:33
- I'll say more about this
- 128:34
- later in this module
- 128:36
- When you work with AI model
- 128:38
- there are many combinations you can use of
- 128:40
- input and output types
- 128:42
- For example, you can input text and images
- 128:46
- such as if you input an inspirational image like this
- 128:50
- if you like this
- 128:51
- Halloween costume and also to plan my Halloween
- 128:53
- costume
- 128:55
- And in this case
- 128:55
- the AI model might output text that says
- 128:58
- let me help you brainstorm a few
- 129:00
- alien-inspired costume ideas
- 129:02
- Or
- 129:03
- you can also upload music
- 129:05
- to an AI model and ask it
- 129:08
- help me plan my haunted house
- 129:10
- And
- 129:10
- AI
- 129:11
- may be able to take these things as
- 129:12
- input and
- 129:13
- generate both text and a video of a haunted house design
- 129:17
- incorporating your creepy sounds audio
- 129:28
- AI models can use most of these input
- 129:31
- types relatively easily
- 129:33
- Some are slightly more expensive or slightly more
- 129:35
- costly to use as an input
- 129:37
- but the differences aren't very significant
- 129:41
- In contrast
- 129:42
- the
- 129:43
- time and cost of generating different types of
- 129:46
- output varies significantly
- 129:49
- In particular
- 129:50
- some data types are much slower and much
- 129:52
- more costly to generate than others
- 129:54
- To give you a sense
- 129:56
- text tends to be on the lower
- 129:58
- end
- 129:59
- in terms of time or cost of generation
- 130:02
- So AI is very efficient at generating text
- 130:06
- In fact, modern AI has started with large language models
- 130:10
- sometimes abbreviated OLMs, but because they started with language
- 130:15
- all of them were
- 130:16
- really
- 130:16
- adapted to deal with text and it's very
- 130:19
- efficient at that
- 130:21
- Generating speech tends to be a bit more expensive
- 130:25
- and generating images even more expensive
- 130:29
- and generating video much
- 130:31
- much
- 130:31
- much more expensive
- 130:32
- than images or any of the other modalities
- 130:36
- And the further we go to the right of this chart
- 130:40
- The longer it takes
- 130:41
- the more time it takes to generate a single output
- 130:45
- and also the more costly it is to
- 130:46
- do so
- 130:48
- Image generation has progressed significantly in the last
- 130:51
- few years
- 130:52
- Here's a short video generated by Imagen
- 130:54
- which was in 2022 a state-of-the-art model by Google
- 130:59
- and it looks pretty good
- 131:00
- but still has some artificial-looking artifacts
- 131:04
- The lines weren't quite right on the back wall
- 131:06
- the dishes changed mid-wash
- 131:08
- In contrast
- 131:09
- modern AI
- 131:11
- video generation
- 131:12
- looks much better and can also be
- 131:14
- synchronized automatically with generated audio
- 131:26
- AI voice generation has also gotten much better
- 131:29
- Here's what AI could do just a few
- 131:31
- years ago
- 131:32
- But before you start tuning anything
- 131:35
- you need to define what success looks like
- 131:38
- This step is easy to overlook
- 131:40
- but it's foundational
- 131:42
- It sounded a bit robotic
- 131:43
- not that expressive
- 131:45
- In contrast
- 131:46
- modern AI voice generation can sound much more
- 131:49
- expressive and much more natural
- 131:51
- But before you start tuning anything
- 131:53
- you need to define what success looks like
- 131:56
- This step is easy to overlook
- 131:57
- but it's foundational
- 132:00
- If you're generating multimodal data
- 132:03
- some of the techniques you learned earlier
- 132:05
- such as giving the model
- 132:07
- enough context and maybe using the best model available
- 132:10
- those are relatively
- 132:12
- easy techniques to apply as well
- 132:14
- to multimodal generation
- 132:16
- But some of the other techniques
- 132:18
- like generating multiple options
- 132:21
- you can still do that
- 132:22
- but if each option now takes many seconds
- 132:25
- or even a few minutes to generate
- 132:27
- then this becomes harder to apply because you
- 132:29
- end up having to wait for a long
- 132:30
- time
- 132:32
- Or if you want to iterate through many designs
- 132:36
- then that too becomes harder if each generation
- 132:39
- takes a long time or is costly
- 132:41
- But if you have the patience to wait
- 132:43
- a little bit longer
- 132:44
- then all of those techniques also apply to generating audio
- 132:48
- images, video, and so on
- 132:51
- With great power comes great responsibilities
- 132:54
- and
- 132:54
- AI technologies can be used for good
- 132:57
- or for harm
- 132:59
- Take voice generation
- 133:01
- If you have recorded a podcast and you
- 133:03
- want to make little fixes
- 133:06
- that can be quite conveniently done today
- 133:10
- using AI voice generation to just re-synthesize one
- 133:13
- or two words that you may have flubbed
- 133:15
- Or
- 133:16
- if you're building a video game and you
- 133:18
- want to give characters lifelike voices
- 133:22
- more and more video game designers are using
- 133:24
- AI voice generation to do so
- 133:27
- It does raise important questions about the livelihoods of voice actors
- 133:32
- and I sympathize with all the voice actors
- 133:36
- that are worried about AI voice generation
- 133:39
- At the same time
- 133:40
- I think it is also very valuable that
- 133:43
- AI voice generation
- 133:44
- is making it easier for a lot more
- 133:46
- people to build entertaining video games
- 133:49
- including developers that don't have access
- 133:52
- to the great voice actors
- 133:54
- In contrast to these applications
- 133:56
- there are also some that are currently harmful
- 134:00
- Unfortunately
- 134:01
- there's been a rise of
- 134:02
- scams where someone would use an AI voice
- 134:05
- clone to pretend to be someone else
- 134:08
- to maybe
- 134:09
- pretend that
- 134:10
- someone's relative is in an emergency and to
- 134:12
- ask to wire emergency funds
- 134:15
- The number of beneficial use cases of AI
- 134:17
- vastly outnumbers the number of harmful ones
- 134:21
- but
- 134:21
- we still have work to do to combat the For applications
- 134:25
- and I hope that each of us
- 134:27
- will only use these techniques for beneficial and responsible applications
- 134:33
- So, as you've seen
- 134:34
- AI can now work with much more
- 134:36
- than text. It can work with images
- 134:39
- audio, video, code, and more
- 134:42
- And
- 134:43
- most of the prompting techniques you learned so
- 134:45
- far
- 134:45
- will be helpful for handling multimodal inputs
- 134:48
- and outputs
- 134:50
- One especially useful capability is giving AI images as input
- 134:55
- so that it can see
- 134:56
- what you're talking about
- 134:58
- Let's go on to the next video
- 135:00
- to see how to use images
- 135:02
- in your prompts
- 135:04
- Providing images of your prompt can enrich the
- 135:06
- context for the AI
- 135:08
- Pictures of something you want the AI to see
- 135:11
- pictures of handwritten text
- 135:13
- really anything that might be hard to describe
- 135:15
- in words
- 135:16
- This video will help you build intuition
- 135:18
- about what AI can see in images
- 135:22
- Here's a picture of me explaining
- 135:24
- some concepts in AI in front of a
- 135:27
- whiteboard
- 135:28
- My handwriting is not that great
- 135:30
- and there are a number of math concepts
- 135:32
- that I'm trying to illustrate on this whiteboard
- 135:35
- If you upload this picture to an AI model and ask
- 135:39
- what is this class about
- 135:41
- it may output something like this
- 135:42
- He's teaching a convolutional neural network
- 135:46
- and the neat thing is
- 135:47
- my head is blocking the word convolutional
- 135:49
- but it knows from this picture
- 135:51
- that I'm teaching about a specific AI technique
- 135:54
- called a convolutional neural network
- 135:57
- And it's extracted some facts about what I'm drawing
- 136:01
- and also has some good guesses about
- 136:03
- what I might ask students to do next
- 136:06
- So it's able to
- 136:08
- make a pretty smart interpretation of this image
- 136:11
- One weakness of AI models
- 136:14
- in terms of how they look at images
- 136:16
- is they tend to look at the coarse image
- 136:19
- but may miss fine-grained details
- 136:22
- So, for example
- 136:23
- if you upload this picture to an AI model and ask
- 136:27
- what are these machines at my gym?
- 136:29
- It may confidently give an answer like this
- 136:32
- which turns out to be wrong
- 136:35
- And that's because a lot of gym machines
- 136:39
- if you look at them through a slightly blurry lens
- 136:42
- they all look a little bit similar
- 136:44
- and
- 136:45
- AI is not that good
- 136:46
- today
- 136:47
- at looking at the fine details of images
- 136:49
- to distinguish
- 136:51
- what really is a good kickback machine or
- 136:53
- a hamstring curl machine
- 136:56
- In contrast
- 136:57
- if you were to upload an image like
- 136:59
- this and ask it to create a sales
- 137:01
- ad for this item
- 137:02
- it actually does pretty well
- 137:04
- because this is very visually distinct object
- 137:07
- And so if you're looking at this even
- 137:09
- through a slightly blurry lens
- 137:11
- yeah
- 137:12
- you kind of see this is a
- 137:13
- human-sized hamster real treadmill
- 137:16
- When you upload an image
- 137:18
- you can also give it
- 137:19
- moderately complex instructions on what to do with
- 137:22
- it. So uploading a receipt like this
- 137:25
- you can ask it
- 137:26
- what's my portion of the bill?
- 137:27
- I had these items
- 137:29
- and in this case
- 137:30
- it gets it correct
- 137:32
- AI's ability to read text like this is
- 137:35
- not bad. It does make mistakes
- 137:37
- so I wouldn't trust it for high-stakes applications
- 137:40
- but if you wanted to take a quick look
- 137:42
- and if you're willing to spend a few
- 137:44
- seconds to double-check the result
- 137:46
- then it could do decently well
- 137:49
- And AI turns out to be
- 137:51
- pretty good at even reading handwritten text
- 137:55
- If you upload a picture like this to
- 137:57
- AI and ask it to transcribe it
- 137:59
- it does a pretty decent job
- 138:02
- Feel free if you want to try
- 138:03
- reading this cursive handwriting yourself to see if
- 138:06
- you can outperform the AI
- 138:09
- And so
- 138:09
- if you upload an image like this
- 138:11
- and write a prompt like
- 138:12
- build an archive of a family's history based
- 138:14
- on these handwritten letters
- 138:16
- I wouldn't trust it to read everything completely accurately
- 138:20
- but it might take a reasonable stab at
- 138:22
- this task
- 138:24
- Rather than uploading a single image to your AI
- 138:29
- sometimes you can upload many images
- 138:32
- For example
- 138:33
- if you just had a brainstorming session
- 138:35
- and had some notes you've taken
- 138:37
- as well as pictures of post-it notes and whiteboards
- 138:41
- you can upload pictures
- 138:43
- and notes to
- 138:44
- an AI model and ask it to
- 138:46
- summarize the ideas from today's brainstorming meeting
- 138:50
- And again
- 138:51
- it will probably do a
- 138:52
- decent job interpreting these images to come up
- 138:55
- with some summary
- 138:57
- Probably not perfect, so it's worth double-checking as output
- 139:00
- but this could help you accelerate
- 139:02
- coming up with notes from today's meeting
- 139:05
- To recap
- 139:07
- AI models can read basic text in images
- 139:11
- Visual understanding, however
- 139:12
- may miss
- 139:13
- details in the image because it tends to
- 139:15
- see the image in a pretty coarse way
- 139:18
- and you can also
- 139:20
- use many images where needed to give the
- 139:22
- AI more context
- 139:25
- A picture is worth a thousand words
- 139:27
- so
- 139:27
- adding an image to a prompt can often
- 139:30
- be the fastest way
- 139:31
- to get the AI model the best context
- 139:34
- like
- 139:35
- taking pictures of a brainstorming exercise or
- 139:37
- digitizing your grandmother's handwritten recipe book
- 139:41
- Of course, AI models can also generate images
- 139:45
- This is an interesting capability
- 139:47
- because it works somewhat differently to how AI
- 139:50
- models generate text
- 139:52
- Let's go on to the next video
- 139:53
- to see
- 139:54
- what are some fun images you can generate
- 139:57
- Generating images of AI has made my life
- 140:00
- more fun
- 140:01
- For example
- 140:02
- I've used image generation for my kids' birthday
- 140:04
- parties or to make fun illustrations
- 140:07
- Generating high-quality AI images is a skill you can learn
- 140:11
- and understanding how AI image generators
- 140:15
- were trained
- 140:17
- will help you to control image generation
- 140:20
- to get better outputs
- 140:22
- Let's take a look
- 140:23
- One neat application of AI image generation is
- 140:27
- if you
- 140:27
- input an image
- 140:28
- and ask it to edit it
- 140:31
- This is a childhood picture of me on the right
- 140:35
- my younger brother on the left
- 140:36
- and one of our childhood friends in the
- 140:37
- middle
- 140:39
- And this is an old
- 140:40
- somewhat faded image
- 140:42
- And if you upload this
- 140:44
- AI model and ask it to remove the
- 140:45
- glare and the rough texture
- 140:47
- and to make it a more natural aspect ratio
- 140:51
- it can produce something like this
- 140:53
- which looks like a nicely restored photo
- 140:57
- This particular image restoration was done with Google's
- 141:00
- Nano Banana model
- 141:02
- If you're not sure how to prompt an
- 141:04
- AI image generation system
- 141:07
- you can actually ask a
- 141:08
- text-based AI model to help you write a prompt
- 141:11
- For example, if you ask it
- 141:13
- generate a prompt for an image of a
- 141:15
- cat secretly running a coffee shop at night
- 141:18
- an AI text model may write a prompt
- 141:21
- like this
- 141:22
- Notice that here it specifies a setting
- 141:26
- specifies details of the character
- 141:28
- and specifies a mood or a style
- 141:32
- And if you don't like any of these details
- 141:34
- you can modify them to your liking
- 141:36
- and a prompt like this might generate this
- 141:38
- cute picture on the right
- 141:41
- People skilled in the visual arts have a
- 141:44
- certain language for describing images
- 141:47
- For example
- 141:48
- a picture like this with this look
- 141:50
- is cinematic
- 141:52
- This is a watercolor image
- 141:54
- This is a cyberpunk image
- 141:56
- And this is an anime image
- 141:59
- And I find that
- 142:00
- art buffs and art history buffs
- 142:03
- excel at image prompting because
- 142:05
- they understand the language of images
- 142:08
- and can describe what they want using more
- 142:11
- precise language
- 142:12
- than those of us that don't know
- 142:15
- this language and can't quite find the right
- 142:17
- words to describe the look that we want
- 142:20
- So if you want to become really expert at generating images
- 142:26
- it could be worth
- 142:28
- reading up
- 142:28
- or
- 142:29
- studying a little bit about the language of
- 142:31
- images
- 142:32
- to understand how to accurately describe different images
- 142:37
- And in fact
- 142:38
- one way to do so
- 142:39
- would be to
- 142:40
- upload images to an AI model
- 142:43
- and ask the AI
- 142:44
- how it would describe those images
- 142:52
- And this could hone your instincts on
- 142:54
- what types of words can be used to
- 142:56
- describe what types of images
- 142:58
- It turns out image generation uses a very
- 143:01
- different technology than text generation
- 143:04
- When AI is generating text
- 143:06
- it produces the output piece by piece
- 143:08
- or it generates a few characters at a time
- 143:11
- In contrast, when generating an image
- 143:13
- it doesn't generate the image a few pixels at a time
- 143:17
- it generates the entire image all at once
- 143:20
- Specifically, during training, that is
- 143:23
- when an AI model is looking at pictures
- 143:26
- maybe found online, in order to learn what images look like
- 143:30
- it will typically
- 143:31
- look at captions or descriptions of images
- 143:35
- Like a small potted plant on a wooden table
- 143:38
- and it'll learn to start from image that
- 143:40
- looks at pure noise
- 143:42
- This is just a grid of random
- 143:44
- pixel values
- 143:45
- And it will then learn to
- 143:47
- sequentially
- 143:49
- remove or subtract noise from the image
- 143:51
- to go from pure noise on the right
- 143:54
- to a slightly blurry picture of a potted plant
- 143:56
- to a less blurry one
- 143:57
- to a less blurry one
- 143:59
- to finally to a sharp picture
- 144:01
- of a potted plant. And that's what the
- 144:02
- AI model
- 144:03
- tries to repeatedly practice doing during training
- 144:08
- A model that does this is called a
- 144:09
- diffusion model
- 144:12
- Then when you come in and when you write a prompt
- 144:15
- maybe create an image of a potted plant on the table
- 144:18
- it then goes through this process
- 144:20
- starting from a pure noise image
- 144:22
- and then
- 144:23
- gradually tries to remove noise from it
- 144:26
- to come up with that final image
- 144:29
- And the key is
- 144:30
- how it learns to subtract noise to
- 144:33
- maybe reveal
- 144:34
- the image that
- 144:36
- the person might have had in mind
- 144:39
- Diffusion models do generate random outputs
- 144:42
- and they can also make certain types of
- 144:44
- errors
- 144:46
- If you repeatedly ask AI model to generate a positive plant
- 144:50
- different times you run the algorithm
- 144:52
- might result in
- 144:53
- different images like the one shown here
- 144:56
- But
- 144:57
- many people have also observed that
- 144:59
- the diffusion model tends to generate weird-looking hands
- 145:03
- often with more or fewer than five fingers
- 145:07
- and it can often output garbled text
- 145:10
- like happy birthday is very badly misspelled here
- 145:14
- and it can also lead to inconsistent characters
- 145:19
- So if you ask it to generate a cartoon
- 145:21
- the character's
- 145:22
- hair has changed between the two frames of this cartoon
- 145:27
- Fortunately
- 145:28
- modern AI models have become
- 145:31
- much better at addressing these problems
- 145:34
- For example
- 145:35
- modern models like Nano Banana
- 145:38
- can allow you to upload
- 145:40
- a number of research papers and ask it
- 145:42
- to generate an infographic
- 145:44
- and it'll do a decent job with text
- 145:46
- that looks mostly plausible
- 145:49
- Or if you ask it to generate cartoon
- 145:52
- the more modern models can generate fairly consistent characters
- 145:57
- meaning now this character
- 145:59
- as you can see
- 146:00
- looks very similar from
- 146:02
- frame to frame with this cartoon
- 146:05
- And the text also looks pretty decent
- 146:08
- Compared to text generation
- 146:10
- image generation can be slow and costly
- 146:14
- For example, if you're generating just a short paragraph of text
- 146:18
- many AI models could do that in just seconds
- 146:21
- and it may cost less than a cent
- 146:24
- Of course, generating long paragraphs of text
- 146:26
- or if you ask it to think for a long time
- 146:28
- that can cost more
- 146:30
- And also AI models were generated
- 146:32
- word by word or maybe a few characters at a time
- 146:35
- and
- 146:35
- you can interrupt it or have it stop
- 146:37
- early if you want
- 146:39
- In contrast
- 146:40
- generating
- 146:41
- a single image might take tens of seconds
- 146:44
- and cost many cents
- 146:47
- and it generates the image all at once
- 146:50
- and there often isn't an option to stop
- 146:52
- early
- 146:53
- Because image generation is much more expensive
- 146:57
- that's why our ability to
- 146:59
- iterate with images is usually more limited
- 147:03
- And if you're generating videos
- 147:05
- then it gets even harder
- 147:07
- Even though some things are more expensive to generate
- 147:10
- the good news is
- 147:12
- the cost of generating virtually
- 147:14
- anything with AI is trending downward
- 147:17
- So, a year from now
- 147:18
- it will be less expensive for you to
- 147:20
- create art for your home or
- 147:22
- graphics for a family member's birthday card
- 147:25
- compared to today
- 147:27
- Beyond generating images
- 147:28
- AI can also help you create fun games
- 147:31
- and websites
- 147:32
- without having to write any code yourself
- 147:36
- This is a more advanced capability
- 147:38
- and it's still easy
- 147:40
- to get stuff that doesn't work or for
- 147:42
- novices to get stuck
- 147:43
- trying to build more complex applications
- 147:46
- But
- 147:46
- I want to just give you a taste
- 147:48
- of how using AI
- 147:49
- to build custom software works
- 147:52
- While it's not that easy
- 147:54
- it's also certainly much easier than it was just months ago
- 147:58
- and quite likely easier than you might think
- 148:02
- Let's go on to the next video
- 148:03
- to see how to use AI
- 148:05
- to create your own mini game or website
- 148:09
- Building computer games and websites
- 148:11
- used to be something that
- 148:13
- only professional developers were able to do
- 148:16
- But
- 148:17
- the ability to do this is being democratized
- 148:21
- By writing text prompts
- 148:23
- you too
- 148:23
- will be able to build
- 148:24
- basic software applications and websites
- 148:28
- This does take some skill
- 148:30
- and I don't want to make it sound trivial
- 148:32
- but in just this one video
- 148:34
- you learn some of the basics
- 148:36
- This is a very exciting capability
- 148:38
- and I encourage you to explore it
- 148:41
- I'm seeing many people
- 148:42
- who are not software engineers
- 148:44
- have a lot of fun with this
- 148:46
- And even with just one prompt
- 148:49
- it's often possible
- 148:50
- to get a cool little game or app
- 148:53
- Let's see some examples
- 148:55
- Using this prompt
- 148:56
- build a game where
- 148:57
- the user has to place obstacles at a goal
- 149:01
- and it creates a simulation of what you
- 149:03
- design
- 149:04
- Claude created this game
- 149:07
- which you can play
- 149:08
- and it's, you know, actually pretty cool
- 149:14
- So just a short
- 149:15
- simple prompt like that
- 149:17
- a
- 149:18
- leading AI model can create a
- 149:20
- reasonably interesting game
- 149:23
- One example that you see in the
- 149:26
- practice lab is
- 149:27
- prompting an AI model to
- 149:29
- generate a fireworks display
- 149:32
- Here's a prompt
- 149:35
- and this creates this app
- 149:37
- which is actually pretty fun to play with
- 149:45
- If you're trying to
- 149:46
- write a prompt to
- 149:48
- tell AI to build a simple app for you
- 149:51
- here are some building blocks you might consider
- 149:53
- including in your prompt
- 149:55
- First is to specify your goal for what
- 149:57
- you want to create
- 149:58
- Next, specify what are the inputs
- 150:01
- what the users need to input in the system
- 150:04
- and then lastly
- 150:05
- what are the outputs or what the
- 150:06
- app shows back to the user
- 150:09
- For example, in this prompt
- 150:11
- we
- 150:11
- are telling it that it goes to generate
- 150:13
- a fun fireworks simulator
- 150:15
- The input is, I want to video click on the screen
- 150:18
- and the output is
- 150:19
- see a colorful display of fireworks
- 150:23
- Sophisticated developers will use AI
- 150:26
- to build much more complex applications than what I'm showing here
- 150:30
- but just to continue with the simpler examples
- 150:34
- to build a game like this
- 150:35
- you might ask it to
- 150:37
- create a fun game
- 150:38
- where the user has to place obstacles at the goal
- 150:43
- and it creates a simulation of what you
- 150:45
- design
- 150:47
- In addition to entertaining games
- 150:49
- you can also make more useful
- 150:51
- and functional apps to help you save time
- 150:53
- or make your life a bit easier
- 150:55
- For example
- 150:56
- you can create a work timer
- 150:59
- called a Pomodoro timer. This is the type
- 151:02
- of timer that people use
- 151:04
- to time their work or studying
- 151:06
- with 25 minutes of work interspersed with
- 151:09
- a 5-minute break
- 151:11
- Or maybe you can create a bill calculator
- 151:15
- Here
- 151:15
- you can input the bill and the number
- 151:17
- of friends you need to split the bill with
- 151:19
- and the app can tell you
- 151:21
- how much each person should pay
- 151:23
- Or maybe build an outfit picker app
- 151:26
- that helps you decide what to wear
- 151:28
- based on the weather
- 151:29
- Each of these apps could be
- 151:31
- fun or useful in everyday life
- 151:33
- and they're also a good starting point if
- 151:35
- you're new to making apps with AI because
- 151:37
- they're quite simple
- 151:39
- In particular, they each have a specific
- 151:41
- well-defined task
- 151:43
- They also don't need any additional files that
- 151:45
- need to be uploaded or outside information
- 151:48
- And they're also something that you can open up
- 151:50
- use for a short period of time
- 151:52
- And in close
- 151:54
- if you're curious to try using AI display yourself
- 151:57
- I encourage you to experiment
- 151:59
- starting with building simple apps
- 152:01
- It turns out some ideas are easier
- 152:03
- to create than others
- 152:06
- For example
- 152:06
- a simple platformer game would be relatively easier to build
- 152:11
- or a quiz to practice French words
- 152:14
- In contrast, a multiplayer game played over the internet
- 152:19
- that would be harder and much more complex
- 152:22
- or
- 152:23
- live French practice
- 152:25
- with AI feedback would also be
- 152:28
- harder to build
- 152:29
- It takes a while to hone intuition about
- 152:32
- what is
- 152:32
- easy for AI to build and what is
- 152:34
- hard for AI to build
- 152:36
- If you're not sure
- 152:37
- I encourage you to just try it out
- 152:39
- and the worst thing that could happen is it doesn't work
- 152:41
- and you start to hone your intuitions about
- 152:43
- what's hard
- 152:45
- But if you're just getting started trying to
- 152:47
- use AI to build apps
- 152:48
- I encourage you to start with simple ideas
- 152:51
- like build a
- 152:52
- simple game and see if you can get
- 152:53
- AI to do that
- 152:55
- When we get to the practice lab later in this module
- 152:59
- you'll be able to try a few examples
- 153:00
- that you can build with just a single
- 153:03
- prompt
- 153:04
- If you want to dive deeper into
- 153:07
- using AI to build software
- 153:09
- I encourage you to take the course Built
- 153:11
- with Andrew offered by DeepLearning.ai
- 153:15
- In addition to writing code to build games and websites
- 153:19
- AI can also help you to analyze data
- 153:23
- which it can do
- 153:24
- by writing code to do that data analysis
- 153:28
- Again
- 153:28
- you don't need to write any code yourself
- 153:30
- Just tell AI what you want
- 153:32
- and or try to write code to do it for you
- 153:35
- and this can lead to helpful insights in
- 153:37
- your work and personal life
- 153:40
- In the next video
- 153:41
- let's take a look at using AI
- 153:43
- for data analysis
- 153:45
- If you have data from your personal health records
- 153:47
- like maybe if you use one of the
- 153:49
- apps that track your heart rate or running time
- 153:52
- or if you have data of sales records from your company
- 153:56
- or really most other types of data
- 153:58
- like what you might saw in a spreadsheet table
- 154:01
- AI
- 154:02
- can be pretty good at writing code to
- 154:04
- analyze this data for you
- 154:06
- and to try to extract useful insights
- 154:09
- Let's look at some examples that hope will
- 154:11
- inspire you to try out this capability
- 154:14
- If you have running tracker data
- 154:17
- that you can download into a spreadsheet
- 154:19
- you might try uploading that spreadsheet to AI
- 154:22
- model and asking it
- 154:24
- how are my pace and distance progressing?
- 154:28
- And an AI system might spend some time
- 154:30
- to write some code to analyze it
- 154:32
- and maybe generate a plot for you
- 154:35
- and also potentially provide some insights
- 154:40
- Or if you run a small business
- 154:42
- and you have a spreadsheet of sales data
- 154:44
- you can upload that data and ask the AI model
- 154:48
- what if you tell me about this month's
- 154:49
- sales?
- 154:51
- The AI could analyze for a little while
- 154:54
- and it might actually write some code to
- 154:57
- do things like compute the monthly revenue
- 155:00
- or to create a graph
- 155:01
- and
- 155:02
- try to show you whatever insights it finds
- 155:05
- I find that AI data analysis
- 155:08
- often isn't as sophisticated as a really
- 155:11
- really good human data scientist
- 155:13
- but for pulling out basic insights and doing so efficiently
- 155:17
- an AI model can be pretty good
- 155:20
- How do AI models write and execute code?
- 155:25
- It turns out that the way it works under the hood
- 155:29
- the ability to write and execute or to
- 155:31
- run code
- 155:33
- is like any other tool
- 155:34
- that the AI model might use
- 155:37
- You've seen how AI model might have tools
- 155:39
- to
- 155:40
- carry out web search or to read and
- 155:42
- write files and so on
- 155:44
- And some AI models
- 155:46
- also have a tool
- 155:48
- to run
- 155:48
- a computer program or to run code
- 155:52
- And this capability
- 155:53
- will often be used when
- 155:55
- data is present
- 155:57
- or when
- 155:57
- some sort of calculation
- 155:59
- or some sort of
- 156:00
- plotting of a graph is needed
- 156:03
- In that case
- 156:04
- the AI could
- 156:05
- generate a bunch of code
- 156:07
- and then use this tool
- 156:09
- to run the code in order
- 156:11
- to generate the result for the user
- 156:14
- We saw previously how a reasoning model
- 156:16
- might input a user prompt
- 156:19
- reason for a while using the context
- 156:21
- and occasionally decide whether or not it needs
- 156:23
- to use a web search tool to get
- 156:25
- more context
- 156:27
- In this case
- 156:28
- it has the option of not just using
- 156:30
- a web search tool
- 156:31
- but also a run code or code execution tool
- 156:35
- in which the AI can
- 156:37
- write a basic computer program to analyze data
- 156:40
- compute averages, plot a graph
- 156:42
- or whatever is needed
- 156:43
- in order to get to the final answer
- 156:47
- Let's look at a concrete example
- 156:49
- If you're running a bubble tea shop
- 156:52
- you may want to look at your sales
- 156:53
- trends over time to answer questions like
- 156:56
- did your new drinks sell well compared to
- 156:58
- your existing drinks?
- 156:59
- If you have sales data handy
- 157:01
- AI can help you with this
- 157:03
- You can attach your sales data as a
- 157:05
- file and write a prompt like this one
- 157:07
- Which drinks had the biggest changes in sales?
- 157:10
- Graph it
- 157:11
- The AI will then go through an agentic
- 157:13
- process to analyze the data and graph it
- 157:16
- inspect the data, then calculate the monthly changes in sales
- 157:20
- and do things like analyze intelligently your data
- 157:24
- to help you get useful insights
- 157:26
- For example, it may say things like
- 157:29
- I'm noticing some clear patterns
- 157:31
- Most drinks are flat
- 157:32
- but four stand out
- 157:33
- Now draft those
- 157:35
- So not just drafting all the drinks
- 157:37
- but
- 157:37
- identifying and potentially focusing on the most interesting
- 157:40
- ones
- 157:41
- Then it might generate a graph like this
- 157:43
- one
- 157:44
- Time is on the horizontal axis
- 157:47
- and number of sales on the vertical axis
- 157:50
- with each drink graphed in its own color
- 157:53
- So the strawberry matcha took off in spring
- 157:56
- mango green tea and strawberry lemonade in summer
- 157:59
- and your new coconut milk tea did well
- 158:01
- in fall
- 158:03
- It also added these colorful highlights to make
- 158:05
- these trends stand out
- 158:07
- So in just a few minutes
- 158:09
- you can get a highly useful graph like
- 158:11
- this
- 158:12
- From this graph
- 158:13
- you might conclude that your spring promotion for
- 158:15
- strawberry matcha was strong
- 158:17
- and maybe you want to try that again
- 158:19
- next year
- 158:20
- You can keep going and iterate on this
- 158:22
- graph and ask for different insights or changes to the graph
- 158:27
- or
- 158:27
- give more context about your business
- 158:30
- to the AI
- 158:31
- to give it a better shot at finding
- 158:33
- more useful insights
- 158:35
- But
- 158:35
- let me show you a more long-thinking example
- 158:38
- You could take your sales analysis further
- 158:41
- and ask for a more comprehensive graphic
- 158:44
- For example
- 158:45
- you might want to create a
- 158:46
- year-end review
- 158:48
- graphic for your business
- 158:49
- to present to your team
- 158:51
- This involves analyzing all your data
- 158:53
- in
- 158:53
- a few different ways to find out most
- 158:55
- interesting to share
- 158:57
- So you might write a prompt like
- 158:59
- create a one-slide year-end review graphic for our bubble tea shop
- 159:03
- analyze the data carefully for insights
- 159:05
- and again, attach your data file
- 159:08
- Using the words carefully in your prompt
- 159:11
- may trigger the AI to think for
- 159:13
- several minutes to complete this analysis
- 159:16
- and it might go through an agentic thinking
- 159:18
- process like you saw on the previous slide
- 159:21
- and it will likely write and run codes
- 159:23
- to calculate things like revenue and items sold
- 159:27
- And in the end
- 159:28
- you might get a graphic like this with
- 159:31
- a lot of interesting insights
- 159:32
- like brown sugar and classic being the most ordered drinks
- 159:36
- and most customers choosing a large drink
- 159:39
- This graphic also has a creative bubble tea
- 159:41
- color scheme
- 159:43
- You probably want to double-check these figures to
- 159:45
- make sure they match your expectations
- 159:48
- since the AI sometimes can hallucinate
- 159:51
- but you can get a good analysis in
- 159:52
- a relatively short period of time just by prompting
- 159:56
- that's fairly likely to be
- 159:58
- accurate since AI calculated these numbers
- 160:01
- by writing and running code
- 160:03
- I encourage you to try this type of
- 160:05
- analysis yourself
- 160:06
- If you have data like sales data
- 160:08
- or personal data you're interested in getting insights
- 160:11
- from
- 160:13
- If you're using an AI model that can run code
- 160:17
- when we choose to do so
- 160:19
- we're seeing that for
- 160:20
- some questions it can use its pre-trained knowledge
- 160:25
- If you're asking for a question that can
- 160:27
- be answered using common knowledge on the internet
- 160:30
- what it already knows
- 160:31
- may well be good enough
- 160:34
- If you're asking a specific question that is real-time
- 160:37
- then web search-enabled AI
- 160:40
- will help you get a better answer
- 160:42
- You may want to
- 160:43
- ask it to use a deep researcher
- 160:45
- If you have a more complex question
- 160:48
- that may require multiple related searches
- 160:51
- such as come up with a
- 160:52
- complete plan for your Halloween house
- 160:56
- And for queries that require calculation or graphing
- 161:00
- Those are the total questions where
- 161:03
- it's most likely to
- 161:04
- write and run some code in order to
- 161:07
- carry out that task precisely for you
- 161:11
- One of the most powerful things that
- 161:13
- AI can do now is
- 161:14
- write and run code
- 161:16
- in order
- 161:16
- to carry out the task for you
- 161:19
- If you have some data that you want
- 161:20
- to get insights from
- 161:22
- I encourage you to try exploring it with
- 161:24
- an AI model
- 161:26
- AI isn't always reliable
- 161:28
- and it's better at simple analysis than really complex ones
- 161:32
- so consider double-checking these conclusions
- 161:35
- But it can be much faster
- 161:37
- than having to do the analysis yourself in
- 161:40
- Excel or Google Sheets
- 161:41
- and it has on many occasions
- 161:43
- helped me discover useful insights in my data
- 161:47
- To put your skills into practice
- 161:49
- we have one last practice lab
- 161:51
- and an optional final project for you
- 161:54
- Let's go into the next video
- 161:56
- to see what these are
- 161:58
- I'm excited for you to test out building
- 162:00
- games and applications using AI
- 162:03
- This should give you a taste for what's
- 162:04
- possible to build with AI
- 162:08
- When you get into the lab
- 162:08
- you can read through these instructions shown here
- 162:11
- but I'm just gonna dismiss this for now
- 162:14
- And I'm just going to
- 162:16
- click the prompt for a fireworks show
- 162:19
- I encourage you to
- 162:20
- read through this prompt carefully
- 162:23
- so that you
- 162:24
- understand what it takes to build an app
- 162:26
- like this
- 162:27
- And I'm just gonna hit run
- 162:31
- And this short prompt will build this application
- 162:35
- I did leave a lot of things unspecified
- 162:37
- but it looks like my AI has made reasonable decisions
- 162:40
- and if I click my mouse
- 162:43
- it launches pretty fireworks
- 162:46
- If I don't want to launch them by myself
- 162:48
- the auto show runs my automatic firework show
- 162:52
- and then here's the grand finale
- 162:57
- So, pretty cool
- 162:59
- One neat feature
- 163:00
- is that you can actually share the apps you built
- 163:02
- So, I'm gonna click share
- 163:04
- just copy the link to my clipboard
- 163:06
- and I can open up a new tab with this URL
- 163:12
- and this actually launches the
- 163:13
- Fireworks app in my web browser
- 163:17
- So, if you share this URL with a friend
- 163:20
- they'll be able to run the Fireworks app
- 163:22
- that you just created
- 163:25
- One of the things I most enjoy about
- 163:26
- building simple applications is being able to share it with friends
- 163:30
- so I encourage you to
- 163:31
- maybe take what you built and
- 163:33
- consider sharing that
- 163:34
- with friends and see what they think
- 163:40
- Feel free to take the Fireworks show prompt
- 163:42
- modify it
- 163:43
- and see if you can get a
- 163:44
- different result that
- 163:45
- could be even more to your liking
- 163:48
- Second example
- 163:49
- let me click on the color palette prompt
- 163:52
- If you are designing a website
- 163:54
- something you may have to do is choose
- 163:55
- a color palette for the website
- 163:57
- and so with a prompt like this
- 164:01
- you can build a color palette picker
- 164:06
- She has a base color in RGB values
- 164:11
- So, this is actually roughly the color of my shirt
- 164:14
- and you can choose complementary
- 164:17
- analogous, and so on color palettes that
- 164:20
- you know, they actually go pretty well together
- 164:24
- Hope you have fun playing with this
- 164:27
- So
- 164:27
- I hope you try out all four of these built-in prompts
- 164:31
- and additionally, try using your own prompt
- 164:35
- For example
- 164:36
- here I'm going to ask you to
- 164:37
- build a flashcard app to help me practice
- 164:39
- basic French vocabulary
- 164:42
- This process may take a few seconds
- 164:44
- and when it's done
- 164:46
- this is what you might get
- 164:48
- Where, what is, yes, oui
- 164:51
- Doctor Shabi, and so on
- 164:54
- Looks like I got that right
- 164:57
- It looks like it's giving me
- 164:58
- all the answers here on the right
- 165:00
- but if I don't want that
- 165:01
- I can go back and modify the prompt
- 165:02
- to redesign my flashcard app
- 165:05
- Please have fun with this
- 165:06
- and
- 165:07
- I find it inspiring that with just a one English prompt
- 165:10
- you can build a webpage to build
- 165:12
- applications like these
- 165:15
- The final project is to build a simple
- 165:17
- app
- 165:18
- from research on a topic that interests you
- 165:21
- We're going to go through three steps
- 165:22
- First, brainstorm a research question
- 165:25
- then run research
- 165:27
- and then to build an app
- 165:30
- I encourage you to read through these instructions
- 165:32
- but I'm just going to dismiss them
- 165:34
- for this demo
- 165:36
- The first step is to brainstorm a
- 165:38
- research question on
- 165:40
- either a topic you want to explore or
- 165:42
- a decision you want to make
- 165:43
- like
- 165:44
- researching health supplements or buying a car
- 165:46
- or
- 165:47
- things to do with your career
- 165:48
- or some fun things with astronomy
- 165:50
- Let me pick the careers one
- 165:53
- Given this prompt or some other one that you may choose
- 165:57
- I would encourage you to then
- 165:59
- give the AI more context
- 166:02
- So I'm just explaining different careers
- 166:03
- know about me, my situation
- 166:06
- Let's say I'm a at university studying deep learning
- 166:09
- I want to work in office
- 166:11
- And I encourage you to write more than
- 166:13
- I am here in this demo walkthrough
- 166:17
- But let me send it to the AI
- 166:20
- And here it will
- 166:22
- help me to brainstorm a few specific research questions
- 166:24
- such as
- 166:25
- what careers let me
- 166:27
- spend a lot of time collaborating
- 166:29
- and what does it look like day to
- 166:30
- day
- 166:31
- So hopefully you pick something that's
- 166:33
- relevant to your
- 166:35
- life and this will help you brainstorm
- 166:37
- a range of different questions that
- 166:39
- hopefully will be interesting to you
- 166:42
- And
- 166:42
- we also have a AI mentor to give
- 166:44
- you some feedback on how
- 166:46
- the brainstorming process is going
- 166:49
- So in this brainstorming workflow
- 166:52
- I would read through these questions and give
- 166:54
- feedback. So let's say
- 166:56
- most interesting question one
- 166:58
- but I like the specificity of question three
- 167:00
- also want to make sure I have time
- 167:01
- to go to my job
- 167:03
- So this type of feedback
- 167:04
- the
- 167:05
- AI
- 167:05
- now has additional context about what make an
- 167:08
- interesting research question for you
- 167:10
- And so it refines it to
- 167:13
- slightly better options
- 167:16
- Let's take one more turn
- 167:17
- Based on these three questions
- 167:20
- my feedback is
- 167:22
- and
- 167:22
- I also want to make sure to consider
- 167:24
- nonprofit work
- 167:28
- And so by going a few rounds
- 167:30
- with looking at the brainstorm research questions and giving feedback
- 167:34
- we've now refined it to a handful of research questions
- 167:40
- So, if you're using an OM
- 167:42
- you can actually go for multiple rounds
- 167:44
- ask the refined questions
- 167:45
- combine them
- 167:46
- but
- 167:47
- let me just end this
- 167:48
- part of the exercise for now and say
- 167:50
- I'd like this question the best
- 167:53
- and I'm just going to
- 167:54
- pick this question to go on to set two
- 167:56
- now that I've formulated a research question
- 167:59
- And what we just did was go through
- 168:01
- a brainstorming exercise
- 168:03
- that's an example of how you may
- 168:05
- iterate with your AI
- 168:07
- to brainstorm any of other possible topics as well
- 168:11
- not just identify a research question to work
- 168:13
- on
- 168:15
- Now that I've picked my research question
- 168:17
- let me give it a little bit more
- 168:19
- context
- 168:19
- So I also want to know about salaries and typical compensation
- 168:24
- and I wanted to use these sources
- 168:26
- forums for personal experience and so on
- 168:28
- and I encourage you to add more
- 168:30
- personal context and add more sources
- 168:32
- that I'm doing in this quick Run through
- 168:35
- but let's just
- 168:36
- have it carry out this research
- 168:41
- And let's have it go ahead and do
- 168:42
- so
- 168:44
- The AI mentor gives some
- 168:46
- feedback on my prompts
- 168:49
- And after running for a while
- 168:51
- here my web search enabled AI
- 168:55
- gives a pretty decent set of results with
- 168:58
- lots of citations
- 169:01
- For the third and final step
- 169:02
- after you've gotten your research report
- 169:05
- let's go build an app
- 169:08
- We can build a quiz
- 169:10
- to test my knowledge based on the reports
- 169:12
- or there's a mini game
- 169:15
- or build an infographic
- 169:18
- All three of these are actually pretty fun
- 169:19
- but I'm gonna choose the quiz option
- 169:23
- So
- 169:24
- here I'm gonna build a five-question multiple-choice quiz
- 169:28
- And the research report that we had
- 169:30
- created from step two is uploaded here as
- 169:33
- an attachment to give it more context
- 169:37
- And to keep the results more predictable in our website
- 169:41
- this prompt is grayed out
- 169:43
- so it's not editable
- 169:45
- But you can always copy the report
- 169:48
- into a third-party AI system
- 169:51
- such as ChatGPT or Gemini or Cloud
- 169:53
- to experiment more
- 169:55
- And I encourage you to play with all
- 169:56
- of these. All of these are actually pretty fun
- 169:58
- but let me just generate my quiz app
- 170:02
- So that'll again take a little bit of
- 170:03
- time
- 170:04
- And so here's the app
- 170:06
- Let me pick that
- 170:08
- Yep, looks like I got that right
- 170:10
- And so on. And same as before
- 170:12
- you can
- 170:13
- copy this link
- 170:15
- using the share button
- 170:16
- and share this app with your friends and
- 170:18
- see if they like it
- 170:20
- So I hope you enjoy going through the iterative brainstorming work
- 170:24
- though, which is a very useful skill
- 170:26
- to have when using AI
- 170:29
- and then
- 170:30
- provide enough context for it to do research for you
- 170:34
- and then based on or inspired by that research
- 170:37
- build one of the
- 170:39
- fun apps that we just showed you and
- 170:40
- maybe even share it with some friends
- 170:43
- I hope you enjoy exploring the lab
- 170:45
- and the optional final project
- 170:48
- and I hope you'll share your final project
- 170:50
- with others as well
- 170:52
- And again, if you're interested in learning more
- 170:54
- I encourage you
- 170:55
- to take the Build with Andrew course
- 170:58
- Once you're done
- 170:59
- I'll see you in one final video
- 171:13
- Congratulations on making it to the end of
- 171:15
- this course
- 171:16
- With all the techniques you've learned
- 171:18
- you're now ready
- 171:19
- to be an AI power user
- 171:21
- I hope you find a lot of places
- 171:23
- in your personal life and work Ways skills will benefit you
- 171:27
- like using AI to help brainstorm
- 171:30
- or use deep researcher where you need thoroughly research reports
- 171:34
- or use AI to help you with writing
- 171:37
- and even generate multimodal outputs and code
- 171:41
- And even as you're doing all this
- 171:44
- AI models will keep on getting better
- 171:48
- So
- 171:48
- please keep trying new models
- 171:51
- and give AI hard tasks
- 171:53
- and provide high quality context
- 171:55
- Does it help you
- 171:56
- to keep honing your intuitions
- 171:58
- about what AI can and cannot do?
- 172:01
- I'm confident
- 172:02
- you get a lot out of this incredibly
- 172:04
- powerful technology
- 172:17
- Thank you for sticking with me to this point
- 172:19
- and I hope you use these powers to help yourself
- 172:22
- your friends, your community
- 172:25
- And
- 172:25
- go make the world a better place
- 172:27
- for yourself and others
- 172:31
- Hi
- 172:32
- I'm Andrew Ng
- 172:33
- I want to show you how to build fun
- 172:35
- useful software using AI
- 172:37
- regardless of your background or training
- 172:40
- If
- 172:40
- you're intrigued about using AI or vibe coding
- 172:43
- to build cool software
- 172:45
- but aren't sure where to start
- 172:47
- this course is for you
- 172:49
- Let me show you
- 172:50
- a funny
- 172:51
- birthday card app
- 172:52
- that you learned to build in minutes using ChatGPT
- 172:55
- Gemini, or any other similar AI system
- 172:58
- even if
- 172:59
- you've never written a line of code before
- 173:02
- This birthday card app is a computer program
- 173:04
- called a web application
- 173:06
- which means that it runs
- 173:07
- right in your web browser
- 173:09
- The app takes whatever you type in
- 173:10
- and creates a custom card for you on
- 173:12
- the spot. In the next lesson
- 173:14
- you build your own version by describing what
- 173:16
- you want the AI to do
- 173:17
- and the AI
- 173:18
- will write all the code for you
- 173:20
- To use the app
- 173:20
- you enter a name
- 173:21
- an age, a hobby, penguin fashion design, an adjective, suspiciously fun, and a plural noun
- 173:28
- lucky socks
- 173:29
- Then hit generate card
- 173:31
- and it says, Karen, against all odds
- 173:33
- you've turned 27
- 173:35
- Your devotion to penguin fashion design is the stuff of legends
- 173:38
- and your suspiciously fun nature
- 173:39
- iconic
- 173:40
- Seriously speaking, most people like you
- 173:43
- I think that's pretty funny
- 173:44
- Or if I don't feel like filling in
- 173:45
- all of these words myself
- 173:46
- I can hit I'm feeling lucky
- 173:48
- and have this
- 173:49
- fill in automatically and then generate a different
- 173:51
- card
- 173:52
- And there's another birthday card
- 173:54
- And you get a different funny
- 173:56
- story or a different funny birthday card this
- 173:58
- time
- 173:58
- Notice the app stores all the cards we've
- 174:00
- generated. If I click copy on any of them
- 174:03
- the birthday message goes to my clipboard so
- 174:05
- I can paste the message into an email
- 174:07
- or text and send it as a birthday
- 174:09
- card
- 174:09
- You might be excited to jump right in
- 174:11
- and start building out your own idea
- 174:13
- And if so, that's great
- 174:14
- And before doing that
- 174:16
- spending a little time building this birthday card
- 174:18
- app yourself will help a lot
- 174:20
- Doing so will give you a better feel
- 174:22
- for how to talk to the AI so
- 174:23
- you actually get the results you want
- 174:25
- And this practice process will build your intuition
- 174:28
- for how to shape what the AI
- 174:31
- actually creates and
- 174:32
- you get a better sense of how even
- 174:34
- small tweaks can lead to very different outcomes
- 174:37
- Once these
- 174:38
- set of concepts become a bit more familiar to you
- 174:41
- building your own ideas
- 174:42
- Faster
- 174:43
- and easier because you have
- 174:45
- a better sense of how to steer the
- 174:47
- process
- 174:49
- So an app like this might seem complicated to build
- 174:51
- but in the next video
- 174:53
- I'll show you how you can build something
- 174:54
- like this
- 174:55
- in
- 174:55
- minutes. Let's go on to the next video
- 174:58
- The easiest way to create software in the
- 175:00
- AI era is no longer to type out code yourself
- 175:04
- Instead
- 175:05
- you should tell AI to do it
- 175:06
- for you
- 175:08
- Telling AI what to do is called prompting
- 175:12
- When giving precise instructions
- 175:14
- AI can do
- 175:15
- a lot of things for you
- 175:17
- What I want to do
- 175:17
- in this video is share with you some
- 175:19
- best practices for prompting AI to create software for you
- 175:24
- First, we'll walk through an example together
- 175:26
- Then, you try yourself using any AI chat system
- 175:29
- like ChatGPT, Gemini, Cloud
- 175:31
- or the one built right into this website
- 175:34
- To use any of these AI systems
- 175:36
- you give it a prompt or set of instructions like this
- 175:40
- to tell it to
- 175:41
- create a webpage to help me write birthday cards
- 175:43
- and when I give a person's name
- 175:45
- age, and hobby
- 175:45
- it should give me back a funny message
- 175:47
- And if you do this
- 175:48
- it might actually generate an app that looks like this
- 175:51
- which is a good start
- 175:52
- You can enter the name
- 175:54
- age, and hobby, and it generates
- 175:56
- you know, an okay message like this
- 175:58
- But if you aren't satisfied with this
- 176:00
- you might then continue the conversation with AI and say
- 176:03
- make it prettier by adding a festive title and colors
- 176:06
- and this would give you a second version
- 176:08
- Of the app
- 176:09
- that now looks a bit better
- 176:11
- And if you're so unsatisfied
- 176:12
- you might say, display the card on the right side
- 176:14
- we can look inside the birthday card
- 176:16
- and get back a third version
- 176:17
- And if you have ideas to make it even better
- 176:21
- you could give it more instructions like
- 176:23
- add a fun title at the top
- 176:24
- and so on
- 176:25
- This is how
- 176:26
- I actually work in practice when using AI
- 176:28
- to write code for me
- 176:30
- I'll often start off with a basic set
- 176:31
- of instructions and see what I get
- 176:33
- and then repeatedly tell the AI
- 176:35
- how I want to improve it
- 176:37
- It turns out that when you're building software applications
- 176:40
- there's a number of basic building blocks you
- 176:42
- might end up including in the prompt
- 176:44
- One thing I often include is the goal
- 176:47
- So here the goal is to create a
- 176:48
- webpage to help me write birthday cards
- 176:50
- And another thing you might include is to
- 176:51
- specify what is the input
- 176:53
- That is
- 176:54
- what does the user have to tell
- 176:56
- the software?
- 176:57
- This is
- 176:58
- us telling the software what we want it to output
- 177:02
- as well as the layout
- 177:03
- So what's on the left
- 177:04
- what's on the right
- 177:05
- how do you arrange the different parts
- 177:07
- of this app?
- 177:08
- And lastly
- 177:09
- any special instructions for additional features
- 177:12
- you wanted to include. There are many ways
- 177:14
- to write good prompts
- 177:15
- but as you start your journey of telling
- 177:17
- AI to build software for you
- 177:19
- I encourage you to consider
- 177:20
- these five building blocks as
- 177:23
- Common pieces
- 177:25
- that you might choose to include in the
- 177:26
- prompt. That is the goal
- 177:28
- namely what you want to create
- 177:30
- what input the user will provide
- 177:32
- what is the layout of the software app
- 177:34
- what special features you want it to have
- 177:36
- as well as
- 177:37
- what you want the software to output
- 177:40
- In the previous slide
- 177:40
- we went through a back and forth process
- 177:42
- of me incrementally in four steps
- 177:45
- adding
- 177:46
- to the instructions I gave to AI to
- 177:48
- tell it what I wanted to do
- 177:49
- But
- 177:50
- if you already know in advance
- 177:52
- more or less what you wanted to build
- 177:54
- you can also specify all of the building
- 177:56
- blocks in a single prompt
- 177:57
- For example
- 177:59
- if I already knew the specification I want for the software
- 178:02
- I could write a single much longer prompt
- 178:04
- like this to create a webpage
- 178:06
- tell me write birthday cards
- 178:08
- what do you have a person's name
- 178:10
- Asian hobby, give me a friendly message
- 178:12
- use festive color and colors
- 178:13
- and so on. And in this example
- 178:15
- I've taken all five building blocks and written
- 178:17
- them all into a single much longer prompt
- 178:20
- So instead of building things step by step
- 178:23
- Like you saw in the last video
- 178:25
- you can also
- 178:26
- give it all the instructions in a single longer prompt
- 178:29
- and that might give you a better
- 178:31
- first version of your app
- 178:32
- which you can then further refine if it's
- 178:34
- still not quite what you want
- 178:36
- Now, whether you're writing a single long prompt all at once
- 178:40
- or whether you're giving it one building block incrementally
- 178:43
- one piece at a time
- 178:44
- I will often start by telling the AI
- 178:46
- what is my goal
- 178:47
- and then of the remaining building blocks
- 178:49
- there are multiple ways to put them together
- 178:51
- and you don't have to use all the
- 178:53
- building blocks every time
- 178:54
- and the order is also not very important
- 178:57
- Start with the goal
- 178:57
- and tell the input output layout
- 178:59
- and maybe not list any special features
- 179:01
- Or you might put the building blocks together this way
- 179:04
- and it could work fine
- 179:05
- Or you might list out the building blocks
- 179:07
- in a different order
- 179:08
- and the AI is usually pretty good at
- 179:10
- understanding these different rearrangements of the building blocks
- 179:14
- And if you're feeling like
- 179:15
- boy, this is a lot
- 179:16
- I would say, don't worry about it
- 179:18
- If you just tell the AI whatever's on your mind
- 179:20
- even if it's partial and imperfect
- 179:23
- you can Then work
- 179:24
- and go back and forth with the AI
- 179:25
- a few times to hone
- 179:27
- it towards something that you want together with
- 179:29
- the AI
- 179:30
- One of the skills you hone over time
- 179:32
- is the ability to give AI
- 179:35
- more specific instructions
- 179:37
- Because it turns out that the results you
- 179:38
- get may vary
- 179:40
- even if you give fairly specific prompts
- 179:42
- So here's a long detailed prompt that you
- 179:44
- saw just now
- 179:45
- specifying all of the five building blocks
- 179:47
- And if you
- 179:48
- give these same instructions to the same AI system multiple times
- 179:52
- maybe one time give an app that looks like this
- 179:54
- which is pretty nice
- 179:55
- And maybe a second time it'll build something
- 179:57
- like that
- 179:57
- And the third time it'll build something like
- 179:59
- that. And all of these look pretty good
- 180:00
- because see that there's some variation among them
- 180:03
- In contrast
- 180:04
- if someone were to give a much less specific prompt
- 180:07
- a much less clear prompt
- 180:08
- so it's a very short prompt that just says
- 180:10
- create a webpage, tell me my birthday cards
- 180:12
- Because this is a relatively vague prompt
- 180:14
- the results you get
- 180:15
- if you run this multiple times through an AI system
- 180:17
- maybe once you get this
- 180:19
- a second time you get this
- 180:20
- which looks totally different with different fields
- 180:23
- and A third time you will get this
- 180:25
- which again looks totally different than the first
- 180:27
- two times. The more
- 180:29
- specific and the more precise the prompts you write
- 180:32
- the more predictable will be the results
- 180:34
- But even then
- 180:35
- there'll be a little bit of variability
- 180:36
- So
- 180:37
- if you get results that are a little
- 180:38
- bit different than what I show in this video
- 180:41
- don't worry about it. That's the
- 180:42
- normal cause of how AI systems behave
- 180:44
- But if it surprises you with something that
- 180:46
- you really don't like
- 180:47
- that's okay too. Just give it additional instructions
- 180:49
- to
- 180:50
- steer or to move the AI closer to
- 180:52
- whatever you do want it to do
- 180:54
- The best way to learn this is to
- 180:55
- put your hands on your keyboard and to
- 180:57
- try using AI yourself
- 180:59
- Let me show you what it'll look like
- 181:01
- What I'd like you to do after this
- 181:02
- video
- 181:03
- is go to this section on the website
- 181:05
- and go through this exercise yourself
- 181:07
- So the instructions here that you can read later
- 181:10
- but this is an AI system
- 181:12
- similar to
- 181:13
- ChatGPT and Gemini and Cloud and so on
- 181:16
- And I'm going to
- 181:18
- Select and then copy and paste the first prompt here
- 181:23
- where it tells the AI
- 181:24
- take care of the webpage
- 181:25
- help me write birthday cards
- 181:26
- and so on
- 181:27
- And I'm going to hit this
- 181:29
- to send it to the AI
- 181:30
- So here, it'll think for a little bit
- 181:32
- and then it will
- 181:34
- generate HTML
- 181:35
- page that it can then download and run
- 181:37
- Notice while it's still running that this
- 181:39
- download button here is grayed out
- 181:41
- so I can't actually click on it yet
- 181:43
- but the AI system will
- 181:45
- take just a little bit of time to
- 181:47
- write something called a HTML page
- 181:49
- This is what web pages are made of
- 181:52
- that will be this birthday card generator
- 181:56
- Now that the AI has finished generating all
- 181:58
- of this HTML code
- 182:00
- I can click this download button
- 182:02
- and here I'm running Chrome on a Mac
- 182:04
- I'll show you later what to do
- 182:06
- if you're on a different machine
- 182:07
- and I can
- 182:08
- go to this downloaded menu here and open up file.html
- 182:13
- and this
- 182:14
- has created a little birthday card generator
- 182:16
- So it's Karen
- 182:19
- 27
- 182:20
- and I can create a simple birthday message
- 182:24
- Not bad
- 182:25
- And notice that this is actually a
- 182:28
- piece of code written in HTML
- 182:31
- that is running on your computer
- 182:34
- right now if you were to do this
- 182:36
- yourself
- 182:37
- The code is in this file called file.html
- 182:41
- and it's actually saved to my computer
- 182:43
- right now. And if you do this
- 182:44
- it'll be saved to your computer
- 182:46
- Now, if you want to improve the code
- 182:49
- you can then prompt it
- 182:51
- add a festive title and colors
- 182:53
- and then it updates the code
- 182:55
- Same as before
- 182:56
- have to wait until it's finished writing the code
- 182:58
- but if I can download it
- 183:00
- and now I can download it
- 183:02
- and same as before
- 183:04
- open it
- 183:05
- Oh wow, now it looks much more festive
- 183:07
- And so
- 183:08
- what I'd like you to do is
- 183:09
- try it out for yourself
- 183:11
- You can, you know, add this third prompt
- 183:13
- run it, add this fourth prompt
- 183:15
- or try some other prompt
- 183:16
- if you feel so inclined
- 183:18
- But it's also
- 183:19
- fine to use just these four prompts one
- 183:21
- at a time
- 183:22
- and see what birthday card generator you get
- 183:25
- out of this process
- 183:26
- Even though I'm showing you this process on the website
- 183:29
- these same prompts should give you similar results on OpenAI's ChatGPT
- 183:33
- Google's Gemini, Anthropix Cloud
- 183:35
- or any other popular AI system
- 183:36
- What you're learning isn't tied to any one platform
- 183:39
- and these skills apply to any AI system
- 183:41
- you choose to use. When you click the download button
- 183:44
- your web browser will usually download the File.html
- 183:49
- or whatever is the file the AI generated with the code
- 183:52
- to your downloads folder
- 183:55
- So these videos show you
- 183:57
- how you can
- 183:58
- navigate on either Windows
- 184:00
- or Mac to the downloads folder to find that file
- 184:04
- and then if you double-click on it
- 184:07
- it should open up in a web browser
- 184:09
- and let you see
- 184:10
- what the code you just generated looks like
- 184:12
- when it runs in your web browser
- 184:15
- Please try this out
- 184:16
- After you try out this process and generate a web app
- 184:20
- one mindset I hope you have
- 184:22
- is that getting feedback is often a
- 184:26
- great step in building software applications
- 184:29
- Whenever I write software
- 184:30
- I'll often show it to friends
- 184:32
- show it to family
- 184:33
- or sometimes
- 184:34
- respectfully approach strangers and ask if they're willing
- 184:37
- to look at whatever I'm building and see
- 184:38
- if they can
- 184:39
- let me know what they think
- 184:40
- or
- 184:41
- email it to a colleague or post it
- 184:42
- online for to get feedback
- 184:44
- Because I find that when people look at it
- 184:47
- they often have suggestions
- 184:49
- for how to make it even better
- 184:50
- Or sometimes
- 184:51
- if you get a laugh out of a
- 184:53
- friend by showing them something funny
- 184:55
- I find that really encouraging as well and
- 184:57
- gives me the energy to keep on going
- 184:59
- So
- 185:00
- what I'd like you to do now is
- 185:01
- go on to the next item in this
- 185:03
- course and try it out yourself
- 185:05
- Get AI to generate some code for you
- 185:08
- and download the
- 185:09
- HTML file and see what results you get
- 185:12
- If you feel so moved
- 185:14
- I hope you also
- 185:15
- show it to a friend or show it
- 185:16
- to someone else to get their feedback
- 185:18
- After that
- 185:19
- please come back to the next video
- 185:21
- where we'll
- 185:22
- keep on working on the app
- 185:24
- and we'll look at how you can add
- 185:25
- even more features to the birthday card app
- 185:27
- to make it even more fun and interesting
- 185:40
- Now that you've built a basic birthday card app
- 185:42
- let's see how you can add additional features
- 185:44
- to it
- 185:45
- so it can do
- 185:45
- additional things or be more fun
- 185:48
- If you showed your app to anyone else
- 185:50
- maybe they'll have ideas for things to add
- 185:52
- too
- 185:53
- And
- 185:54
- I also want to show you what to
- 185:55
- do
- 185:56
- if there's something wrong with whatever the AI
- 185:58
- had built for you
- 185:59
- such as
- 186:00
- if it doesn't work the way you were expecting
- 186:03
- As before
- 186:04
- the primary way we will
- 186:06
- tell AI what features it had
- 186:08
- is by
- 186:09
- prompting
- 186:10
- To understand why it's important to be specific
- 186:12
- in the way you prompt
- 186:13
- let me pick an analogy to ordering food
- 186:16
- If you go to a food truck and say
- 186:20
- give me a sandwich
- 186:21
- who knows what type of sandwich they will
- 186:23
- give you if that's all you say
- 186:24
- But if you say something more specific
- 186:26
- like I would like a vegetarian sandwich
- 186:28
- then that narrows down
- 186:29
- the range of what you may get back
- 186:31
- Or if you say
- 186:31
- I'd like a vegetarian sandwich with hummus and
- 186:33
- cheese and multigrain bread
- 186:34
- then that
- 186:35
- specific instruction really
- 186:37
- means what you get back becomes much more
- 186:40
- predictable. Or if you say
- 186:41
- I'd like such and such vegetarian sandwich
- 186:43
- and add a drink and make it to go
- 186:45
- please
- 186:45
- then it becomes much more predictable what
- 186:48
- the food truck will give you back
- 186:51
- At a food truck
- 186:52
- by being specific with what you ask for
- 186:54
- you're more likely to get what you want
- 186:56
- And so too it is with prompting AI
- 186:58
- By writing more specific prompts
- 187:00
- you're more likely to get exactly
- 187:02
- what you want
- 187:05
- this is exactly what I wanted
- 187:08
- As you gain skills in getting AI to write for you
- 187:12
- you get better at
- 187:13
- being more specific with the instructions you give to AI
- 187:16
- and this will give you better results
- 187:18
- I previously built an app using a prompt like this
- 187:21
- either on one go
- 187:23
- or by going back and forth a few
- 187:24
- times
- 187:25
- What I'd like to do now is
- 187:26
- add a few new features and modify some
- 187:29
- existing ones. For example
- 187:30
- instead of having three input fields like we have so far
- 187:34
- name, age, and hobby
- 187:35
- maybe you want to gather five pieces of information
- 187:39
- so
- 187:39
- five input fields that users can input
- 187:42
- so you can create more personalized messages
- 187:44
- Or
- 187:44
- maybe you want a I'm feeling lucky button
- 187:47
- that automatically fills out all the fields for you
- 187:49
- so you don't have to type everything in
- 187:50
- each time you use the app
- 187:52
- Or you can also update the look and feel
- 187:54
- say change the title or
- 187:56
- add a button that copies the message to
- 187:58
- your clipboard so you can easily
- 188:00
- email it to friends or
- 188:01
- even redesign the entire The scheme to better
- 188:04
- match your style. After having built the basic app
- 188:07
- you can then
- 188:08
- make the app yours by deciding what features
- 188:11
- you want to add
- 188:12
- So
- 188:13
- you might continue the chat conversation by telling
- 188:15
- the AI to make it have five clearly labeled inputs
- 188:19
- name, age, hobby, an adjective, and a plural noun
- 188:22
- and it'll update the app
- 188:24
- to look like this
- 188:25
- And notice that I'm writing very
- 188:27
- specific instructions here. I'm not just saying make
- 188:30
- it have five fields
- 188:31
- I'm telling it what are the five fields
- 188:32
- I want it to have
- 188:34
- Or if I add an I've feeling lucky button
- 188:37
- again
- 188:37
- I try to be specific in telling it
- 188:39
- what I want the I'm feeling lucky button
- 188:41
- to do
- 188:42
- I want to automatically fill in the blanks with random words
- 188:44
- and I want the
- 188:45
- predefined word choices to be funny
- 188:47
- And Here
- 188:49
- I'm actually adding two features at the same
- 188:50
- time. I might update the
- 188:52
- titles and subtitles and also add a button
- 188:54
- to copy the cards to clipboard
- 188:56
- So this is actually two of the bullets
- 188:58
- from the previous slide that I'm adding at
- 188:59
- the same time. And I actually like the color blue
- 189:02
- so let's make the color theme blue
- 189:04
- If you feed a child these prompts yourself
- 189:07
- or even better
- 189:08
- pick a different color theme
- 189:10
- or
- 189:10
- tell it to implement any other idea that
- 189:12
- you have
- 189:13
- If you like pink like my daughter does
- 189:15
- or green like my son does
- 189:17
- tell it to use a pink or green color theme instead
- 189:20
- or make some other changes depending on what
- 189:21
- you feel like doing. And if you ever change your mind
- 189:24
- you can do that too
- 189:25
- So after I've made the color theme blue
- 189:27
- if I decide that I don't want it blue after all
- 189:30
- I actually want it to be purple
- 189:31
- you can write a prompt like that
- 189:32
- and the AI will obediently change it to purple
- 189:36
- Now, a lot of the time
- 189:38
- AI is pretty good at giving you
- 189:41
- exactly what you ask it to do
- 189:43
- but sometimes there's a chance
- 189:45
- to generate an HTML page that doesn't work
- 189:48
- In software, we call this a bug
- 189:51
- when there's something that doesn't quite work the
- 189:53
- way it's supposed to
- 189:54
- Here's actually a broken version of an
- 189:57
- app that AI generated for me earlier
- 190:00
- where I've typed in all of these fields
- 190:02
- but if I create generate card
- 190:04
- Nothing happens. So that's kind of weird
- 190:06
- Here I'm clicking my mouse
- 190:07
- but it's not actually generating a card
- 190:09
- It turns out I'm feeling lucky works
- 190:11
- but again
- 190:12
- generate card doesn't create any card for me
- 190:14
- If that happens
- 190:15
- what I'd encourage you to do
- 190:17
- is to clearly tell the AI what happens
- 190:20
- So here, typed
- 190:21
- nothing happens when I click on the generate
- 190:23
- card button. Can you fix it for me?
- 190:25
- And if you do that
- 190:27
- the AI
- 190:28
- is usually pretty good at finding
- 190:31
- at least basic bugs and
- 190:32
- fixing whatever is wrong
- 190:34
- When you do this
- 190:35
- sometimes the AI will
- 190:36
- write some technical explanations of what went wrong
- 190:39
- So here it says
- 190:40
- no JavaScript attached, it was a click event
- 190:42
- Boy, that's a lot of technical terminology
- 190:44
- I would say for now
- 190:45
- just don't worry about what the AI is
- 190:46
- saying here. You don't need to understand these
- 190:48
- technical details. If you're really curious
- 190:50
- you can actually ask the AI
- 190:51
- what do these terms mean?
- 190:53
- But you don't have to do that
- 190:54
- Let the AI do its thinking
- 190:56
- say what it wants about the technology
- 190:58
- and then focus on downloading the new
- 191:01
- HTML file to see if that works
- 191:03
- I hope you will start to the basic
- 191:04
- birthday card app and
- 191:07
- either try building the features
- 191:09
- I suggested earlier in this video or
- 191:11
- some features that
- 191:12
- maybe your friends have suggested
- 191:14
- or your own ideas. And in fact
- 191:16
- if you're not sure what else to do
- 191:18
- you can also ask the AI for ideas
- 191:20
- So if you ask it
- 191:21
- how can I make this birthday card app
- 191:23
- cool? It may
- 191:24
- give a few
- 191:25
- suggestions and then you could pick one or
- 191:27
- more
- 191:27
- and use the AI's idea to make your
- 191:30
- app
- 191:31
- Cool. I actually use AI
- 191:33
- as a brainstorming tool a lot when I'm building software
- 191:37
- and
- 191:37
- you should do the same too if you
- 191:39
- don't already have ideas that you're excited about
- 191:41
- Please go to the next item in this
- 191:43
- website and have fun adding features or
- 191:46
- play the color theme or
- 191:48
- implement whatever you like
- 191:50
- When you come back
- 191:51
- we'll use the skills you've already learned
- 191:53
- to build a second app
- 191:54
- a ping pong game
- 192:06
- Using the skills you now have in prompting
- 192:08
- let's build another app
- 192:10
- this time a ping pong game
- 192:13
- Because of AI
- 192:13
- it's now possible to go from an idea
- 192:16
- to a working app
- 192:17
- sometimes in minutes
- 192:19
- One of the earliest video games in computer
- 192:21
- history was a game called Pong
- 192:23
- which was sort of a two-person ping pong game
- 192:26
- and that had taken the team weeks to build
- 192:29
- But now, thanks to AI
- 192:31
- you can build something like this in minutes
- 192:34
- Let's apply the prompting techniques we learned
- 192:36
- I'm going to start quickly and write a
- 192:38
- moderately specific prompt and say
- 192:40
- build me a table tennis game as a single issue file
- 192:43
- user plays against computer
- 192:44
- moves the paddle, the arrow keys
- 192:46
- And if you do that
- 192:47
- you might get
- 192:48
- a first version of the app that looks
- 192:49
- like this. I go
- 192:50
- okay, that's a good start
- 192:51
- But
- 192:52
- I wanted to add three difficulty levels
- 192:54
- and let the user specify
- 192:57
- the number of points required to win
- 192:58
- and also have it keep score
- 193:00
- And this builds me
- 193:03
- a second version like that
- 193:05
- And this now looks like a fun game
- 193:07
- but I want the graphics to be fancier
- 193:09
- So I'm going to say
- 193:10
- make the playground green
- 193:11
- paddles beige, ball white
- 193:12
- and I'm going to tell it to
- 193:14
- insert this image into the background
- 193:16
- And so I end up with a game
- 193:18
- that looks like this
- 193:18
- And here is the game
- 193:20
- It's actually pretty Pretty fun
- 193:22
- The ball is bouncing back and forth
- 193:24
- I could actually play this for quite a long time
- 193:26
- maybe longer than you'd actually want to watch me playing it
- 193:29
- but I hope you will have fun with
- 193:30
- this
- 193:32
- As a reminder
- 193:33
- being specific in how you write your prompt
- 193:36
- gets you better results. And so
- 193:38
- you can look over the prompts I was using
- 193:40
- and it was fairly specific in a number of details
- 193:43
- like the color, the points, and so on
- 193:46
- If you're not sure what to include in the prompt
- 193:48
- think about the building blocks
- 193:49
- Do you want to specify the goal
- 193:50
- the
- 193:51
- output, the input, the layout, the features
- 193:53
- as a set of things to consider including
- 193:56
- And lastly, use your chatbot to iterate
- 193:59
- improve
- 193:59
- and troubleshoot. You don't need to get it
- 194:01
- right the first time. You can tell what you already know
- 194:03
- what you already have in mind
- 194:04
- see what you get
- 194:05
- and then
- 194:06
- use that to further refine what you tell
- 194:08
- the AI to do
- 194:10
- I actually enjoy ping pong in real life
- 194:11
- and now you can build
- 194:13
- a game to play it also with your
- 194:14
- computer
- 194:16
- Please go to the next learning item to
- 194:17
- give this a shot yourself
- 194:19
- And in addition
- 194:20
- beyond building a perfect card generator or ping pong game
- 194:24
- if you have an idea for something else
- 194:25
- you want to try to build
- 194:27
- give it a shot. May or may not work
- 194:28
- but it's by
- 194:29
- practicing exploring our ideas that all of us
- 194:32
- get better at building things
- 194:44
- If you made it this far
- 194:46
- congratulations
- 194:47
- You're an AI builder now
- 194:50
- I've met people who've been clearly building software of AI
- 194:53
- maybe even for months
- 194:55
- that somehow still wonder if they're actually a
- 194:58
- builder
- 194:59
- I'm here to tell you
- 195:00
- yes, you are, and I count you as one of us
- 195:04
- a builder
- 195:06
- I hope you do one last activity
- 195:08
- and get a certificate for this course
- 195:11
- The final
- 195:12
- project is to make your own fill-in-the-blanks
- 195:16
- story builder
- 195:17
- I hope you use the prompt building blocks
- 195:19
- to build your own fill-in-the-blanks app
- 195:21
- Your final project must include three to five input fields
- 195:24
- a button that takes the input and then generates the output
- 195:28
- and a place to display the output
- 195:30
- You've already seen the birthday card app
- 195:32
- Here are some other ideas for maybe inspiration
- 195:36
- This is a funny product review generator
- 195:40
- where the blanks are the product name
- 195:43
- some number, a noun, a body part, and so on
- 195:46
- and it generates product reviews
- 195:48
- Notice it says there are totally honest product
- 195:50
- reviews that look like this
- 195:53
- Feel free to pause the video and read
- 195:55
- this if you want
- 195:55
- but I love this
- 195:57
- Or here's another Time off request generator
- 196:01
- where you specify how much time you went off
- 196:03
- a noun, a body part, an object, and then it says
- 196:06
- I need an entire month off
- 196:08
- effective immediately
- 196:09
- because the vintage forever requires my urgent attention
- 196:12
- And so whether a
- 196:13
- birthday card or a product review
- 196:15
- or a time off or some other type of story generator
- 196:19
- please write your own prompts and
- 196:21
- use it to generate an app like this
- 196:23
- On the website, when you go to the next item
- 196:26
- you will see the usual chatbot that you
- 196:27
- can use to write your prompts
- 196:29
- to get an HTML file
- 196:31
- and download it onto your own computer
- 196:33
- Please do that, run it
- 196:35
- see if it works to your satisfaction
- 196:36
- And you could either use our website
- 196:39
- or
- 196:39
- you're also welcome to use an external website
- 196:42
- like ChatGPT or Gemini to generate your
- 196:44
- HTML file. When you're done building this project
- 196:47
- please go to this final page
- 196:49
- It might take a second to load
- 196:51
- and then click upload file
- 196:53
- And here I'm going to navigate to
- 196:56
- the file that I had downloaded
- 196:58
- And
- 196:59
- here's a preview of the app I actually
- 197:00
- generated. When you're ready
- 197:02
- hit submit assignments
- 197:03
- and our AI will look at your
- 197:05
- HTML file
- 197:06
- and see if it
- 197:07
- seems to work correctly and give you feedback
- 197:09
- if any
- 197:11
- Assuming it works
- 197:12
- this will bring you
- 197:14
- to the last
- 197:15
- exercise of this course and will earn you
- 197:17
- a certificate
- 197:18
- Most effective builders keep taking courses and also
- 197:21
- keep building
- 197:22
- I find that if someone only builds
- 197:25
- they often end up unaware of core concepts
- 197:28
- and can end up spending months reinventing the wheel or worse
- 197:31
- doing things in very strange ways
- 197:34
- But conversely, if someone only takes courses
- 197:37
- then they end up with
- 197:38
- theoretical knowledge without knowing how to apply it
- 197:41
- So both building and taking courses are important
- 197:45
- Please keep building whatever you feel like
- 197:48
- and
- 197:48
- I'll also suggest some additional courses for you
- 197:51
- to consider in the next item
- 197:53
- Keep sharing your applications with friends to get the feedback
- 197:56
- or maybe
- 197:57
- just to get a laugh out of them
- 198:00
- Being a builder
- 198:01
- is one of the most fun things
- 198:03
- in the world
- 198:04
- I'm glad you and I have started this journey together
- 198:08
- and I hope we'll keep building
- 198:09
- and learning together
- 198:11
- Thank you also to Deep Learning AI's Summer Rae
- 198:14
- Andreas Gonzalez
- 198:15
- and Tommy Nelson who had contributed to this
- 198:17
- course
Watch these videos on your TV
XTTV brings video content from X to your living room. Download now and start watching on the big screen.