← Back to videos

Kimi创始人杨植麟这39分钟,比很多2万欧的硕士课还实在。 从中国起步,把�

Hux

Hux

@huxlab

November 19, 2026· 39:43

Kimi创始人杨植麟这39分钟,比很多2万欧的硕士课还实在。 从中国起步,把月之暗面做到估值超500亿美元,他几乎没卖课、没包装成功学,就把真正在用的判断全掏出来了。 几件最有冲击力的事: 1. 他死磕生产力场景,主动砍掉陪伴/娱乐。因为提升智能能直接带来留存,而娱乐的天花板已经被移动互联网打满了。组织天然想做更多,他硬抗这个惯性,只做一件事做到极致。 2. 长上下文是他早期非共识的押注,现在被验证了。真正有价值的不是“能读多长”,而是在超长窗口里做长推理,从几分钟的任务,慢慢扩展到需要两周甚至一个月的复杂目标。人类根本标不了这种数据,AI已经在某些维度超过人了。 3. 先把模型做到足够有用、足够真实(少幻觉),信任和“长期伙伴”关系才会自然长出来。里程碑很具体:哪天AI干的工作超过你自己的一半。 4. 长上下文成本会再降一个数量级以上(缓存、架构、选择性记忆),目标是今年让两百万字真正人人用得起。 最后他反复强调: 技术的第一性原理是人和组织。 招顶级技术人才、建能让他们在scaling law框架里持续创新的机制,比任何预测都重要。 AI发展太快,快速调整比押对远方更关键。 这更像在爬一座山,每一步都能感觉到模型和流程在变好,而不是在海上漂着看不到进展。 在这个还要用10-20年视角看的马拉松里,工具和模型会越来越便宜,真正拉开差距的是谁先把“聚焦 + 长推理 + 组织能力”这套工程师长期主义跑通。 https://t.co/Vo3s4aIhf5

Transcript

I feel that the industry as a whole
0:02
is
0:02
certainly a marathon from our point of view
0:04
I think technology develops relatively quickly
0:07
but looking at the whole picture
0:09
the spread of this technology
0:11
including the product and its commercialization
0:14
I believe we must look at this from
0:15
a ten to twenty year perspective
0:17
So I think in the last year
0:19
we
0:22
we basically focused on just one primary objective
0:24
which was to thoroughly explore and validate this
0:27
early stage product market fit
0:28
ensuring that we truly understood the needs of
0:30
our target audience before moving forward
0:33
And maybe in terms of on the technical side
0:35
we have, iterated on some progress of the models
0:39
enabling the model to better serve the needs of the users
0:41
I believe
0:44
There might be some very
0:45
very early stage PMF
0:46
but I think at the same time
0:47
there are still many challenges
0:50
I think the most critical thing is that within this
0:56
For example, the scaling law
0:57
Regarding the scaling law
0:58
I think we are still in the very early stages today
1:00
so maybe
1:01
we still need to see how we can
1:03
scale to the next generation model and the
1:05
one after that
1:06
And through this
1:07
we can form a stronger PMF in the
1:08
product
1:10
Then it can truly have a higher penetration
1:12
in everyone's work and life
1:13
I think there are still many challenges
1:15
I think it might be a longer-term thing
1:17
Well, I hear you being a bit modest
1:19
so let me narrow down the topic a
1:20
bit. So you see
1:22
you actually entered the market relatively late
1:25
but our impression from the outside is that
1:27
you've accelerated very quickly
1:29
and now you have great momentum
1:31
I think you were being very objective just now
1:33
and I believe you might have higher expectations
1:35
for the development of this matter itself
1:37
But from our objective viewpoint
1:39
you are still developing quite well
1:41
Could you summarize for us what you think
1:42
you've done that was relatively successful
1:45
to achieve today's objective results?
1:46
Which choice do you think was right and
1:48
led to this outcome?
1:52
that's a great question
1:53
I think, strictly speaking
1:55
I certainly can't say it's been very successful
1:57
or anything like that
1:58
I think there might be some relatively small starts
2:00
and then
2:02
I think the things we continue to focus on here
2:05
and at the same time
2:06
the things we are currently optimizing
2:10
one might be the hope to truly solve
2:11
this problem from the ground up
2:13
focusing on first principles
2:15
Because I think every industry has a basic assumption
2:18
just like the basic assumption of
2:21
the personal computer industry is Moore's law
2:23
right?
2:24
Then I think
2:25
the basic assumption of the
2:26
AI industry is still the scaling law
2:29
So, if we look at this issue from a ten-year perspective
2:32
I think it's more
2:33
about how we can
2:35
Continuously optimize in terms of technology and model effectiveness
2:39
and form a stronger PMF in the process
2:41
So many of our actions
2:43
will likely start from this point of view
2:45
and then we'll plan out
2:46
exactly what we want to build
2:48
So I think this is very important
2:49
meaning we want our biggest investment
2:52
and probably everyone's time and energy
2:54
to be focused on how to iterate and create better models
2:58
thereby unlocking more scenarios
3:00
This is likely what we care about most
3:02
And of course
3:03
this in itself probably also means focusing
3:05
It's likely that in terms of technology and product
3:08
we all want to be more focused
3:10
for example
3:10
we might now be more
3:12
focused on productivity scenarios
3:14
And for scenarios outside of productivity
3:16
we won't be doing much for the time being
3:19
because if you try to do everything
3:20
it's hard to do anything well
3:22
So
3:23
There are trade-offs in this
3:24
right? You have to make some choices
3:26
Yes, exactly
3:29
Because we're still a startup
3:30
and I think a startup needs to have
3:32
a very clear focus
3:34
For instance, we might aim for extreme optimization
3:37
specifically for productivity. Often
3:39
the product might just look like a simple
3:41
text box appearing unchanged
3:43
but behind the scenes
3:45
many experiences have actually been greatly improved
3:48
Of course
3:48
there's still plenty of room for growth
3:51
So I think in this process of making trade-offs
3:53
it often means
3:55
that you need to cut certain things out
3:57
You can't do everything
3:58
because an organization's natural inertia
4:00
is to want to do more and more
4:02
right? So we have to fight against that gravity
4:05
meaning you actually want to do fewer things
4:07
but do them to the extreme
4:10
This applies to, you know, technology too
4:12
because the AI space is so vast
4:14
you might consider, since intelligence itself is very heterogeneous
4:18
a very diverse thing
4:20
the intelligence of an accountant
4:22
a painter, or a mathematician is completely different
4:26
So we'll look at our core user base
4:27
to see what kind of
4:29
intelligence is the priority now
4:30
and what the corresponding core capabilities should be
4:34
Yeah, and then focus more intently on that
4:37
You mentioned that achieving today's results is largely
4:39
due to your focus
4:41
specifically on the productivity dimension
4:43
There were trade-offs where you chose not to
4:45
pursue certain directions
4:47
I'm interested in specifics. For example
4:49
the current trend of AI companionship and chatting
4:52
I'm sure you've looked at it
4:53
and I bet your team discussed it
4:55
but ultimately decided against it
4:56
What was the logic behind that?
4:58
What was the reason for giving it up?
5:01
Right
5:03
We did indeed discuss this issue
5:05
and I think there are a few main
5:06
points
5:07
One is considering what we ultimately want to achieve
5:10
because what we want to build is general
5:12
intelligence
5:13
So between casual chat scenarios and actual productivity scenarios
5:17
I think there's a high probability
5:19
combined into a single product
5:21
So
5:22
now it's more about having different options for
5:23
your path
5:25
So regarding the path
5:26
I think we should focus on productivity first
5:29
because, it improves intelligence much faster
5:32
For instance
5:33
if we were to build a product like Character today
5:35
then most of your effort wouldn't be spent on optimizing intelligence
5:39
because optimizing intelligence might not help you improve
5:41
product retention at all. But if you focus on productivity
5:43
then
5:44
optimizing intelligence will lead to a significant boost
5:47
in retention
5:48
So I think it allows for a much
5:50
tighter integration between your company's mission and your product roadmap
5:54
which is a very important reason
5:57
And of course
5:57
there are likely other reasons as well
5:59
For example
5:59
we've observed that different development paths in the US market
6:02
since the US market is generally one to
6:04
two years ahead of China's
6:06
we can look at how different companies have
6:07
evolved
6:09
In this context
6:10
companies that chose productivity and excelled at it
6:13
are doing much better today
6:14
whether in terms of business volume
6:15
financing, or attracting talent
6:17
so I think this is also a very
6:18
important reference for our decision
6:21
And another key point is that
6:23
in entertainment scenarios, the baseline today is already extremely high
6:27
right?
6:28
Because the mobile internet's growth over the last
6:30
decade created many great entertainment experiences
6:33
However
6:34
I believe there's still
6:35
massive untapped potential in the
6:38
fundamental experience or value of the productivity dimension
6:41
Even the best productivity products today
6:44
haven't really deeply integrated into actual workflows yet
6:47
and I think this is the huge new
6:48
value that AI brings to the table
6:50
So that's likely a major reason why I
6:52
made this specific choice
6:53
Yeah
6:54
I see
6:55
so if we take a step back
6:57
how exactly should we define Kimi then?
6:59
Even though we all use it every single day
7:01
it really seems like we haven't actually made a conscious
7:04
deliberate effort to sit down and define it
7:05
properly
7:06
But I think that's interesting how you would
7:08
define it
7:09
So
7:09
what exactly is Kimi and what problem is
7:11
it solving?
7:12
Is it an AI workstation in the long
7:14
run or something else entirely?
7:16
Have you had these kinds of discussions or definitions internally?
7:20
Well
7:20
I think we should look at this from
7:21
both a short-term and long-term perspective
7:23
In the short term
7:25
its current definition is that we want it
7:27
to
7:27
provide significant value in productivity scenarios by offering
7:30
more and more intelligence
7:32
Of course
7:33
today the main focus is likely on
7:35
things like better information retrieval
7:37
analysis, and even content creation
7:40
Where we hope it can provide even greater
7:42
value in these tasks
7:43
So that's the short-term view
7:45
as for the long-term ideal scenario
7:47
we've discussed it
7:48
and I think there are two different viewpoints
7:50
now
7:50
One is that this is essentially about the
7:52
ultimate form of an AGI product
7:55
or what its ultimate definition should be
7:57
There are essentially two distinct possibilities to consider
8:00
One is that it's an alter ego
8:01
which is a version of you existing out
8:03
there in the world
8:04
And the second possibility is something else entirely
8:06
another me
8:07
meaning it could have all of your inputs
8:10
And its thoughts are basically the same as yours
8:11
so it's like you've replicated another version of yourself
8:14
and this other you can do
8:15
many things in the digital world
8:17
and even in the physical world
8:19
So I think, well
8:20
another definition is that it could be your
8:22
partner, a long-term, even nearly lifelong partner
8:26
And this partner can also help you do many things
8:28
but it might be different
8:29
in that it could offer you new perspectives
8:31
and it's not a complete replica of you
8:33
but rather it's a bit like
8:35
But it understands me
8:36
right?
8:37
Not a replica, but it understands me
8:38
It understands you very
8:39
very well
8:40
Yeah. I, I
8:41
I think
8:41
the second one is more likely now
8:44
So this might be what we want to build
8:45
Well, I think it will have a few key characteristics
8:48
First, I think it first has to be useful
8:50
meaning you can do
8:52
more and more things. I think what it
8:53
can do today isn't quite enough yet
8:56
And so that's why our main focus right
8:57
now is still on
8:59
or further improving the model's capabilities
9:02
because only by improving the model's capabilities can
9:04
you let it do more things
9:06
For example, if you compare it to a human today
9:09
it's still missing many dimensions compared to a
9:11
person
9:12
It might lack memory
9:13
and have no way to do very long-term
9:15
planning
9:16
In my imagination, if this product is eventually done well
9:19
it shouldn't just be about you completing these
9:23
ten or twenty-second tasks
9:25
Instead
9:25
it should be that you could even set
9:28
an OKR for it
9:29
like what its OKR for this quarter should
9:31
be
9:32
And then it could actually just go and
9:33
do that thing
9:34
So I think the first key feature is being useful
9:36
gradually being able to do more complex
9:39
tasks. And then the second key feature
9:41
I think is building this very long-term trust
9:43
and connection between AI and humans
9:45
But the prerequisite for this is still that first step
9:48
which is that it must be useful enough
9:50
For instance, it can't hallucinate every day or give you many
9:54
wrong conclusions
9:56
or make you feel like it's just providing
9:58
a lot of incorrect information
10:01
I think only by
10:02
solving for truthfulness and being able to provide
10:04
that in increasingly complex
10:07
and long-context tasks with truthfulness and accuracy
10:11
do I think it can build that great
10:12
sense of trust
10:14
So we'll probably consider the future of this product
10:17
from these perspectives
10:20
So actually from this angle
10:22
we see that today
10:23
is likely just an early form of the
10:25
product
10:26
This also explains why we're seeing various new
10:28
features in the product
10:29
Essentially
10:30
you're working hard to earn a more useful
10:31
place by the user's side
10:34
So this role will gradually expand from a few capable
10:37
tasks to more and more things
10:39
That might be the path
10:41
Yes
10:41
I think there will be a milestone like
10:43
this
10:43
If one day you find that AI is
10:45
doing more in your work than you are
10:48
meaning AI's share exceeds fifty-five percent
10:51
I think that will be a very important milestone
10:54
And then, of course, the next milestone might be that you
10:58
for example
10:59
the number of robots might exceed the number of humans
11:01
but that's in the physical world
11:03
However
11:03
I think the digital world might see this
11:05
first. It could be realized much sooner there
11:09
So I'm curious about something
11:10
Kimmy
11:10
looking at it from the perspective of Moonshot
11:12
AI's Kimmy
11:13
You have a very successful product today that everyone is using
11:16
and in the future
11:17
you might have new features to solve more
11:20
specific user problems
11:21
Would you prefer to solve everything within the Kimmy platform
11:24
or is there a chance for other more
11:26
specialized apps to emerge?
11:27
Have there been any principled discussions on this?
11:30
Right. Our current thinking is that we will
11:32
definitely focus on building just one app
11:35
because I believe that future
11:37
Just Kimmy?
11:38
Yes
11:39
No second one
11:41
Right. We'll definitely focus on this
11:42
because
11:43
I think a very important point is that future intelligent products
11:46
should satisfy universal needs
11:48
just like I mentioned earlier
11:50
Actually, your entertainment needs, including productivity needs, Even your
11:56
Actually
11:56
these two things might not have very clear
11:58
boundaries
11:59
I think they should be within the same
12:00
product
12:01
It's just that
12:01
there will be differences in your choice of path
12:04
right? Because I think this is also what's most interesting about
12:07
general intelligence
12:08
It's not just for one thing
12:10
but you really can't do everything right from
12:12
the start in terms of path selection
12:14
At first
12:15
there will likely be focused scenarios and a core user group
12:18
but it can gradually generalize from there
12:21
So we still hope to focus on one
12:23
product and make it the absolute best it
12:24
can be
12:27
It sounds like, to reach Kimmy, what you just said
12:30
that ultimate goal of becoming a partner who
12:32
understands the user
12:34
The starting point is to first help the
12:35
user do one or two things well
12:37
then gradually do more and more
12:39
and that partnership and trust will be built
12:42
So this is a forward-looking worldview and roadmap
12:45
Yes, exactly
12:46
And yes
12:46
I think a key indicator is that right
12:48
now you might just have a single query
12:50
asking it one question at a time
12:52
or having a very specific task
12:55
And for example, if a person, if a professional, were to do this task
12:59
they could probably finish it in a minute or two
13:00
but now with your new approach
13:02
you are just making the whole process significantly
13:04
faster and more efficient
13:06
But I think gradually
13:07
you'll move from these short duration tasks to
13:10
these are complex things that
13:11
might take a typical person
13:13
two weeks or even a full month to complete
13:15
and honestly
13:15
even humans might not do them particularly well
13:18
That is because AI today
13:20
already possesses some forms of superintelligence
13:22
Now, to be clear, it is not all around superintelligence
13:24
but for instance, if you consider long context
13:27
and I mean really long context
13:28
it is quite remarkable
13:30
It is actually a form of superintelligence because
13:32
humans can't read through
13:34
say
13:35
a multi-million word article in one go and then immediately find
13:39
the answer to a question because many answers
13:41
can't be found just by using control F
13:43
That's a pitfall we encountered early on
13:45
Because we thought humans could do this
13:46
we hired a bunch of people to label
13:48
that long context data
13:49
you label the data
13:50
you learn it
13:51
right? But we found that humans simply can't label it
13:54
or their efficiency is extremely low
13:56
right? So I think today
13:58
some capabilities can indeed outperform humans
14:01
but we probably need to gradually
14:04
expand the scope of these tasks because today
14:05
I think it's still within a very limited range
14:08
yeah
14:09
You mentioned long context
14:11
which was really one of the key things
14:13
that made Kimi so incredibly impressive right from
14:15
the very start
14:16
Even back then, it was actually quite a bit of a
14:18
a non-consensus view
14:19
when you first proposed that this specific feature
14:22
was going to be so important for the
14:23
future of AI
14:24
But it seems like the industry is reaching
14:26
more of a consensus lately
14:27
isn't it?
14:28
Regarding long context, your understanding is
14:30
since we see many models increasing their supported context length
14:33
so there are two points here
14:35
Is this becoming a consensus?
14:36
And second
14:37
is long context the most important path to
14:39
solving that ultimate problem you mentioned?
14:41
Right, that's a great question
14:43
I think more people are working on this today
14:45
but as for a complete consensus
14:47
I'm not entirely sure
14:48
because there are different viewpoints
14:50
Some might think that long context isn't actually that important
14:54
right?
14:54
And you should actually
14:56
well, there's a lot to explore in short context scenarios
14:59
so you might not be in a rush
15:00
to do long context
15:01
So I tend to think a little differently about this
15:03
but honestly, I think this is completely normal
15:06
Because every single product or company is going
15:08
to have its own unique judgment
15:09
its own specific choices
15:11
and its own way of handling differentiation
15:13
So
15:14
I really do think this is just a
15:15
very normal process
15:17
For us
15:17
we realized quite early on that this shift
15:19
might be a very critical thing for our
15:21
workflow
15:22
there are a few key aspects to consider here
15:25
first
15:25
if we really want to move from simple
15:27
tasks that take just one or two minutes to complete
15:29
up to tasks that take thirty or fifty minutes
15:32
or even
15:32
projects that span two weeks or perhaps a
15:34
full month to finish
15:36
we need a new approach
15:37
It is a fundamental change in how we
15:39
think about productivity and time management
15:41
and then it inevitably requires a very long
15:43
context to function effectively
15:45
Only by successfully completing increasingly complex tasks within
15:48
this expansive long context can you truly push
15:50
AI capabilities to the next significant step
15:53
So I personally think this is incredibly important
15:55
for the future of the field
15:57
but it is not just about the long context itself
15:59
it is about how we utilize that capacity
16:02
Is it a sufficient condition?
16:03
It might be a necessary one
16:05
It's necessary
16:06
Not enough
16:06
That is simply not sufficient
16:08
right?
16:09
you might have a very long context window
16:10
but within that specific window
16:12
you also absolutely need to have strong reasoning
16:14
capabilities
16:16
So I really think that
16:17
long context itself is just one dimension of the problem
16:21
but it is more like a cycle
16:23
You first need a longer window
16:24
then within that expanded window
16:26
you can develop much stronger reasoning
16:28
and with that stronger reasoning
16:30
you can then effectively handle even longer windows
16:32
than before
16:33
So I think it's like a spiral iterative
16:35
process
16:36
A more accurate term might be long reasoning
16:39
being able to reason well within a very
16:41
long window
16:42
I think that's what will ultimately generate great value
16:45
yeah
16:47
The essence of long context should be long
16:48
reasoning
16:49
Right
16:50
This actually aligns with solving more complex problems
16:53
for users
16:54
The exchange ratio between the gain in intelligence
16:56
and the service value exchange for user instructions must improve
17:00
right?
17:01
essentially, it seems you have a comparison parameter
17:03
and only when it reaches that value
17:05
do you hit the critical point
17:07
Yes
17:08
I think the point you're making makes a lot of sense
17:09
yeah
17:10
So
17:11
it's ultimately a calculation of the conversion ratio
17:13
between the user's data and instruction input
17:16
and the service you finally deliver
17:17
right? The higher it is
17:18
the more value it likely brings
17:20
Right, that's very interesting. For instance
17:22
even with, long text technology itself
17:25
its implementation is a process of gradually increasing
17:28
that ratio
17:29
In the early stages
17:30
you might use long context for reading tasks
17:33
Those are definitely the first to be implemented
17:35
because they involve turning a lot of information into a little
17:39
which is relatively easier
17:41
For example
17:41
if I read ten articles and ask for a summary
17:44
that's certainly
17:46
less difficult than me
17:47
giving a simple instruction today that requires a month of work
17:50
and after that month
17:51
you actually satisfy the user's needs
17:54
I think the latter is definitely harder
17:56
So this is indeed
17:59
as you said, a matter of ratio
18:01
the quotient of your input and output might
18:03
be the fundamental thing
18:05
That's the value of the gain that intelligence
18:07
can bring
18:08
Let me ask about another recent technical shift
18:10
in the industry
18:11
For example, we've seen Sora and GPT-4.0
18:14
What's your take on Sora?
18:16
Will this kind of video generation capability
18:18
be something Moonshot AI's Kimi focuses on in the future?
18:22
Yes
18:22
that is definitely an important point to consider
18:24
because
18:25
I really think that when we are talking about general intelligence
18:27
the system must be multimodal to truly understand
18:30
the world
18:30
It's hard to imagine a single mode general intelligence
18:33
so I think eventually
18:36
different modalities will surely be integrated into a
18:38
unified model
18:40
And of course
18:40
I think we can clearly see now that
18:42
technological development might follow two very different dimensions
18:46
The first dimension is still the continuous
18:48
steady rise of artificial intelligence
18:49
which continues to evolve and reshape our world
18:51
in ways we are only just beginning to
18:53
fully understand
18:54
For instance, with Sora or GPT-4.0
18:57
the intelligence gain itself might not be that huge
19:00
They've improved, but not drastically
19:02
If you gave them an IQ test or more complex tasks
19:05
they might still fail
19:07
So I think we must keep investing
19:08
in this direction
19:10
It's the most important one in my view
19:12
but there's another dimension
19:13
which is continuously expanding the modalities
19:17
For example, you might currently have
19:19
video or voice modalities integrated into your system
19:22
and in the near future
19:23
you could even expand those capabilities to include complex sensory data
19:27
precise motion data
19:28
and a variety of other new robotic modalities
19:31
But the real value there is really more
19:33
about being able to effectively handle
19:35
a much wider range of diverse and challenging
19:37
scenarios than you could before
19:39
And then there are richer ways to interact
19:41
which can really take this product from where
19:43
it is right now
19:44
and honestly
19:44
I think it could play a huge pivotal
19:46
role in helping it finally cross the chasm
19:49
This would allow more and more people to
19:50
actually use it in their daily lives
19:52
and by doing so
19:53
it would be making this technology truly accessible to everyone
19:56
everywhere, regardless of their technical background
19:58
I personally believe that they might actually be
20:00
two completely different dimensions
20:02
but I am confident that they will eventually
20:03
unify in the end
20:07
I totally understand
20:08
For instance, with 4.0
20:10
its multimodal capabilities are something that every model company
20:14
every foundation model company
20:15
will likely need to evolve towards
20:18
But regarding something like Sora and video generation
20:22
if we follow your previous theory
20:23
which focuses more on the growth of intelligence
20:26
is the ability to generate video
20:28
on that same path of intelligence growth?
20:30
Or is it more about the service delivered
20:32
to the user?
20:33
Why exactly do you think it's so important?
20:36
Right. Well, actually, this question
20:38
I think it was discussed back in the
20:39
era of pure language models
20:41
I remember between 2019 and 2020
20:44
a very important topic was whether language models
20:47
should focus on understanding or generation
20:51
Right
20:53
So at first, there were models like
20:54
for example, BERT, and then later came things like GPT
20:58
where GPT's generative ability was better
21:00
But the advantage of the BERT model was
21:02
that given the same computing power
21:04
BERT was always more efficient
21:06
meaning the improvement to benchmarks per unit of
21:08
computing power was greater
21:10
So for a while
21:11
everyone focused on BERT
21:12
saying as long as it can understand
21:14
that's enough
21:15
And perhaps most of the industry's value lies
21:16
in that understanding
21:18
But I think
21:19
a very important point might have been overlooked here
21:21
which is that if you want to achieve excellent understanding
21:24
you actually
21:27
Well
21:27
you need to have excellent generation
21:29
to achieve excellent understanding
21:31
So these two issues might ultimately just be
21:32
one and the same
21:33
I think it's the same for video
21:35
today when we want to do great video generation
21:38
it's partly because video generation itself has high value
21:42
For example, there are many creators
21:43
so the user value itself is high
21:45
it's very valuable to users
21:47
But I think there's something even more important
21:49
than that. More importantly
21:50
it's that through generation itself
21:52
if you can optimize the objective function for generation exceptionally well
21:55
it will ultimately inevitably lead to better understanding
21:59
These two goals, well
22:00
I think text has already taught us a
22:02
big lesson over the past few years
22:04
There was a lot of debate at first
22:05
but now there's basically a consensus that
22:08
understanding and generation are inseparable
22:10
It's very hard
22:11
to train an understanding model in isolation
22:13
in the end, they're likely just one model
22:17
Understood
22:18
Let me ask a bit more about long
22:19
context windows or what we might call long
22:21
chain reasoning. Will the cost of this
22:25
Drop rapidly in the future
22:26
because we all know that if you actually
22:28
throw two million words in for a run
22:30
the cost is still quite high
22:32
If I don't have a very reasonable cause
22:34
right, I'd feel bad about it
22:36
I can't just keep freeloading off you
22:37
and you wouldn't let me anyway
22:39
So you'll definitely consider this
22:41
and it also relates to how
22:42
in what scenarios can I use such a
22:44
long text to really
22:46
solve problems with high value
22:48
These two issues are bonded together
22:49
What's your take?
22:51
Right. I think the cost
22:53
A continuous decline is definitely an inevitable trend
22:55
and because we have some new technologies recently
22:58
that is to say
22:59
we can significantly reduce this cost
23:01
To what extent? Give us a sense of
23:03
it
23:04
I honestly think that there could be at
23:05
least an order of magnitude drop
23:07
or
23:07
from what we are seeing in terms of
23:08
current costs
23:09
Much? Maybe more than one order of magnitude
23:11
more than one, yes
23:12
So on one hand
23:14
we are likely to see some truly extreme
23:15
engineering optimizations taking place
23:17
For example
23:18
we might soon have access to advanced technologies like context caching
23:22
which will undoubtedly lead to massive
23:24
game-changing cost reductions for everyone involved in the process
23:27
And secondly
23:28
we've also done a lot of optimization on the architecture
23:30
right?
23:31
So
23:31
I firmly believe that if we are talking
23:33
about a massive two million word context window like that
23:36
and we are looking at making it truly
23:38
affordable for most people to use in their daily work
23:40
I think that
23:41
should be a very important goal
23:42
that will in all likelihood be achieved in
23:44
the near future
23:46
It is just a matter of time
23:49
this year
23:50
This year?
23:51
So this year
23:51
two million words
23:52
is something that more people can actually access
23:55
Yes, I think that's our goal
23:57
And also in the coming long period of time
24:00
the cost will continue to drop
24:01
and
24:02
the cost will drop even faster than it
24:04
does for short context
24:05
For instance
24:06
there's still a lot that hasn't been tapped
24:08
into yet
24:09
For example, if you
24:10
if you consider a person processing a very
24:12
long memory or a very long
24:14
long-term task, they don't actually need to remember everything
24:17
right? Maybe it's a dynamic process of calculation
24:21
You can choose what to keep and what to discard
24:23
and within that
24:25
there's actually huge room for optimization because its
24:27
efficiency today is far higher than a human's
24:30
so there's a lot of room to optimize
24:32
So the overall speed of cost reduction
24:35
will actually be faster than the cost reduction
24:37
for short contexts
24:40
If we take this and look at it from this perspective
24:43
to imagine or think about what this kind
24:44
of variable might bring in terms of application scenarios
24:47
how would people use this capability?
24:50
For example
24:50
the previous case of throwing in a book
24:52
and having it summarize it for me
24:54
that's very straightforward
24:55
So based on your reasoning
24:57
what other scenarios might become available?
25:00
Right
25:00
this is somewhat related to what we were
25:02
just discussing
25:03
That is
25:03
it's a process of adjusting the output ratio
25:07
You might start with reading
25:08
and reading might
25:09
It's one of the most pressing needs right now
25:11
and then it will gradually become
25:12
Basically
25:14
the model itself will have the ability
25:16
within this very long window to perform reasoning and planning
25:19
and then execute multi-step tasks
25:22
For example, if you wanted to
25:23
say, research a certain topic today
25:26
right? Or
25:27
you could even give the AI a very clear objective
25:30
and then it could go and execute that
25:32
kind of multi-step planning
25:33
and then call upon different tools
25:35
right? And even have some internal reasoning and analysis
25:38
I think it will gradually evolve in that direction
25:41
Of course, it's also very important for multimodality
25:44
Like, if you wanted to generate a very continuous video today
25:47
a longer one, behind the scenes
25:48
you'd probably need very good long-context technology
25:52
So now I understand why you said that
25:54
long context might actually be closer to long reasoning
25:57
essentially
25:57
It's not just about how much text I
25:59
give it in the traditional sense
26:01
but rather how much reasoning capacity it has
26:03
and how much information it can
26:05
right?
26:06
Used for further reasoning and creation
26:08
that's what's going to become more important
26:10
right? Because if you only have
26:11
you might have the length
26:12
but the brain isn't powerful enough to process
26:14
it on its own
26:15
that data isn't very valuable
26:17
So
26:18
both elements really have to be done well
26:20
together
26:21
That's quite inspiring
26:22
It will shift from being a product feature
26:25
to a backend product capability
26:27
an underlying capability
26:28
which will then lead to even more powerful features
26:31
right? That might be the path forward
26:34
Yes
26:35
And there's actually
26:36
a lot of exploration involved with the users because
26:38
for instance, many long context scenarios today
26:42
we didn't even foresee on day one that
26:44
there might be scenarios like this
26:46
for example
26:49
Even just the act of reading itself
26:51
we might not have thought before that it
26:52
could be used to
26:54
say, quickly get started in a new field
26:57
right, or things like that
26:58
or in different industries it might
27:00
have these different applications
27:02
For example, some users might use this to perform certain analyses
27:06
but if you don't provide the context
27:08
the analysis might not be as good
27:10
But once you provide it with these three articles
27:12
it might start using things like the McKinsey method
27:15
it'll analyze things better and be more structured
27:18
right? So I think
27:19
it's a process of co-creation with users where
27:21
we'll keep discovering new use cases
27:23
We just feel that the scaling law
27:25
including the transformer, might lead to the future
27:27
right?
27:28
I see we're over time
27:29
but let me ask everyone
27:30
would you like to chat for another
27:32
ten to fifteen minutes? Are you hungry?
27:33
You want to?
27:34
Okay
27:34
let's continue. Let's give Jillian a round of
27:36
applause for more of his time
27:42
everyone wants to keep going
27:43
Okay, then let me ask another question
27:46
Actually, recently, in the academic world
27:49
we often see some discussions
27:50
We just felt that the scaling law and
27:52
the transformer might be the way forward
27:54
Then we see some experts we trust
27:57
including academic scientists
27:58
saying they don't actually have that much confidence
28:00
in it
28:01
They believe that new changes need to occur
28:03
This actually leaves us lay people feeling a bit confused
28:06
because Zilin
28:08
you were once a young scholar
28:10
and are now an entrepreneur
28:12
I'm sure the worldview of a scholar differs
28:13
from that of an entrepreneur
28:15
but how do we reconcile these two?
28:17
How do you view their opinions and judgments?
28:19
That is right
28:20
I think it's really like this
28:22
the academic world is primarily focused on the
28:24
rigorous process of identifying and establishing a correct first principle
28:29
which serves as the foundational bedrock for all
28:32
subsequent research and discovery
28:34
That's its mission
28:35
right?
28:36
And I personally think that the entire industry
28:37
is really all about achieving the absolute best execution possible
28:41
specifically by focusing on a first principle
28:44
By execution
28:45
I don't mean it's just following orders
28:47
It also requires a lot of innovation
28:48
but at different levels
28:52
Your first principle
28:53
might involve innovating at the very bottom layer
28:57
So academia discusses whether the scaling law is
28:59
correct or if next token prediction is right
29:02
I think these questions are very meaningful
29:05
and they should be discussed and challenged
29:06
You know, new viewpoints should emerge
29:08
and everyone has different ideas
29:10
In this process, for example, neural networks
29:13
maybe thirty or forty years ago
29:14
or even twenty or thirty years ago
29:17
didn't get much attention
29:19
People didn't think it was a good technical
29:20
path
29:22
So I think that's
29:23
the greatest value of academia
29:25
The value of industry
29:26
or what it does is to
29:27
solve the most important problems within a technical
29:30
path or based on a first principle
29:33
Take the scaling law
29:34
for instance. Although it's a first principle
29:36
there are still many unsolved problems
29:38
like how to generate data effectively
29:40
how to build good multimodal models
29:42
or how to create a solid data flywheel
29:44
All these issues need to be addressed
29:46
but it's never about inventing a new first
29:48
principle. It's about working and innovating at different
29:50
levels
29:51
That's my understanding
29:53
So academia naturally requires more debate
29:57
we need people to propose many new challenging ideas
30:00
but in industry, I think it's about
30:02
how to move faster
30:04
or better solve what are also major technical
30:06
challenges at the second level
30:07
even when first principles are clear
30:09
right? It's just that your technical path might be fixed
30:11
but
30:12
I don't think there's that much conflict overall now
30:14
for instance
30:15
Yann LeCun always talks about world models
30:17
a world model is essentially
30:19
or rather
30:20
current large language models are actually just a
30:22
special case of a world model
30:24
Right, so I don't think they conflict
30:26
Actually
30:26
there really isn't that much of a conflict
30:28
So for us
30:29
I think it's about working within the framework
30:31
of scaling laws to explore the limits of intelligence
30:35
and then, but I think in the development of human technology
30:38
new technical paths will always be proposed
30:41
However, I feel this is more of a maybe
30:43
academic or research, a pure research mission
30:46
so I think they're at different levels
30:50
right
30:50
Each has its own mission and its own efforts to make
30:53
right?
30:53
there's really not that much conflict
30:55
Let me ask a more specific question related
30:57
to your work
30:58
Actually, last year, when we were chatting
31:00
one thing you mentioned that really stuck with me was that
31:03
a startup in the era of large models
31:05
isn't just about innovating in technology and products
31:09
It might also need organizational innovation
31:11
Because building products today compared to before
31:13
involves many more system variables
31:15
You have models, you have data, you have users, and so on
31:19
what results have you seen this year in
31:21
terms of organizational innovation?
31:23
Any results, like grounded results?
31:26
Right, I think this is still an ongoing process because
31:30
an organization itself needs room
31:32
time to grow, right? I think actually
31:35
many times we see some American companies running
31:37
faster than us. For example
31:39
their overall AI capabilities are still better
31:42
I think that's actually thanks to the fact
31:43
that they probably spent a lot of time
31:45
early on building the organization
31:47
It's not just about hiring the best people
31:49
but forming a mechanism where
31:51
all these people can innovate within this paradigm
31:54
And, because I think Chinese companies in many cases are still
31:58
well, they did start a bit later
31:59
So there are two different types of
32:01
companies here
32:03
One type is those that might have been
32:04
doing other businesses before
32:06
and then maybe now they're pivoting to this
32:08
new business
32:09
the organizational style needed for that new business
32:12
might be increasingly different
32:13
Then there's the other type
32:15
starting from zero to one
32:17
From zero to one
32:17
there might be less organizational baggage
32:20
but you still need to explore a good
32:22
way of doing things
32:23
So overall, I think this
32:25
takes some time
32:26
right? But I think we've made some progress now
32:28
though there's definitely still a lot of room
32:30
for improvement. Yeah
32:31
So it's definitely an important thing
32:33
but it just takes more time to figure out
32:35
Right, right. Because I feel that when we consider technology
32:38
it essentially comes down to
32:40
how that technology is produced
32:42
right? It's produced by people
32:43
and then people are paired with the materials of production
32:47
So
32:47
so if your people can be organized effectively
32:49
So people are the first principle of technology
32:52
or rather human organization is the The first
32:54
principle behind that technology
32:56
Yeah, I think that's actually the case
32:57
So we also pay a lot of attention to
32:59
you know, how you can recruit the best talent
33:02
especially technical talent. So these are all
33:05
the foundation for being able to build great technology
33:09
right? So
33:10
is recruiting where you spend most of your time now
33:12
does it take up a large proportion?
33:14
yeah, the overall proportion is quite high
33:16
It's just quite, it remains quite high
33:18
primarily because this is still the essential fuel
33:20
that powers our ongoing core development efforts
33:23
When you are hiring
33:24
what kind of people do you spend the
33:25
most time recruiting?
33:27
right now the main focus is probably on the technical
33:30
technical talent
33:32
Technical talent. Right, right
33:33
because I think this is still related to
33:35
your company's priority because Well
33:37
I think for us
33:38
the most important thing is to get the technology right
33:40
because only when the technology is good can
33:42
you unlock
33:43
more product scenarios and have better
33:46
retention and better commercialization. I think the foundation
33:49
for everything is having better technology
33:50
So today, in this regard
33:52
although we already have some good people
33:54
we definitely need to continue to
33:56
strengthen and continue to
33:59
have more and better people join us
34:01
Yeah
34:03
In the past six months or so
34:04
or since you started the business over a year ago
34:06
is there anything from this past year that
34:08
you
34:09
feel you predicted correctly and things you predicted
34:11
incorrectly?
34:13
You could mention one of each that stands out
34:15
like one where you thought
34:16
I got that right
34:17
and another where you really guessed wrong
34:20
Is there anything like that you can share with us?
34:23
honestly
34:23
I think that overall making accurate predictions is
34:25
a very difficult thing to do
34:26
So
34:27
I really think a much more important point
34:28
is actually focusing on rapid adjustment
34:30
because, you know
34:32
AI develops so incredibly fast these days
34:34
that a lot of the time
34:35
it is just really hard to predict what
34:37
is coming next
34:38
That is why being able to pivot quickly
34:40
is essential
34:42
For example
34:42
what will models be like next year?
34:44
I think that's
34:45
a very hard question to answer
34:46
You might have some ideas or judgments
34:48
but the most important thing is to adapt to new variables
34:52
which could come from the market or from
34:54
your new iterations after many experiments
34:57
or maybe from user feedback
34:59
and I think reacting very quickly is probably
35:02
probably the most important thing
35:04
And, if I had to talk about predictions
35:06
I think there are a few things
35:08
where the trend is similar to what we
35:10
thought at the start
35:11
like the context length keeps increasing
35:14
and including video generation capabilities
35:16
reaching minute level generation
35:18
and I think the trend is about what we expected
35:20
but Some timing judgments might not be as accurate
35:24
like Sora's release might have been earlier than we imagined
35:27
though maybe not really earlier since reaching actual
35:29
product market fit might still take a
35:32
Anyway, we haven't used it yet
35:33
right? Because we haven't reached that point yet
35:36
It might be due to some cherry-picking factors
35:38
or maybe you're trying to reach
35:40
PMF might still take some time
35:41
and that includes
35:43
the fact that most current intelligence gains come from
35:46
say, doing better post-training on models like GPT-4
35:49
but for something like GPT-5
35:51
it looks like the timeline will be later
35:52
than originally predicted
35:54
It'll be later
35:55
I think making accurate predictions about these
35:58
specific points in time is very difficult
35:59
yeah
36:00
So, overall
36:01
what I'm hearing is that you don't care
36:03
much about those long-term predictions
36:05
You think that, in other words
36:07
those predictions should instead become
36:09
a more continuous and effective form of reasoning
36:12
right? Fast and effective reasoning
36:14
Once a variable appears
36:15
you immediately take the next step
36:17
rather than looking too far ahead
36:19
It's more of a dynamic and rapid response
36:21
I think This is
36:22
well, this might be something that we
36:24
as a small startup
36:25
can do better at
36:26
yeah
36:27
Instead of placing a big bet on something far off
36:29
it's
36:29
You definitely need to make some judgments on
36:31
the direction
36:32
I think basically it's very important to have long-term certainty
36:35
long-term certainty, and moving steadily toward that goal
36:39
And in the short term
36:41
there should also be strong certainty
36:42
because it's tied to how well you can
36:44
execute
36:46
But when you are looking at plans on
36:47
a scale of just a few months
36:48
I really think you can afford to allow
36:50
for some flexible adjustments along the way
36:53
That way
36:53
the overall probability of making a significant mistake
36:56
is actually much smaller
36:57
yeah
36:58
I'll ask one last question that's a bit
37:00
out there. I want you to answer with your intuition first
37:03
and then we'll use
37:04
system one and then system two
37:06
okay?
37:07
For this question
37:08
I want you to answer with your intuition
37:09
first
37:10
Don't start reasoning it out yet
37:12
Do you feel your work today
37:14
your startup life
37:15
is more like climbing a mountain
37:17
or more like sailing the sea?
37:21
Oh, that's a very good
37:22
quick, quick
37:25
maybe more like climbing a mountain
37:27
Okay, good, you said it
37:28
Now let's reason it out
37:29
Why is it more like climbing a mountain?
37:32
Well, my first instinct was climbing
37:34
because we've always felt that
37:38
we're climbing stairs, not looking at the scenery
37:40
That's something we always say internally
37:42
So, that was my first reaction
37:45
But I think now if you use system
37:48
system two to think about it
37:50
well, he is starting to reason
37:52
Your model is actually starting to reason now
37:56
I think it's just
37:59
because sailing in my imagination
38:01
since I've never actually sailed
38:03
in my imagination, you're on the ocean
38:05
and even if you've traveled
38:06
say, hundreds of kilometers
38:07
what you see is basically the same
38:10
Your goal is still very clear
38:11
but well, what you see
38:13
Progress isn't clear
38:14
Right?
38:14
or in a situation where you don't have
38:16
very good positioning technology
38:18
Reference points are unclear. Exactly
38:19
You might
38:20
right? It's not that you can clearly see different things
38:22
but rather you see that nothing around you has changed
38:25
so your sense of loneliness will be stronger
38:27
than
38:28
I'd even say it's just like that
38:29
but with mountain climbing
38:32
you can actually feel yourself improving with every step
38:36
for example
38:37
you can feel that the model's capability seems
38:39
a bit better than a few months ago
38:40
right? And your workflow seems a bit more
38:42
efficient than a few months ago
38:44
and then the view you see is different
38:47
And when you measure the distance between yourself and the goal
38:49
you actually have a much better feeling
38:51
So I think it's something with clearer progress
38:54
so I think that's one aspect
38:56
On the other hand
38:57
I think the development of AI is also a gradual process
39:00
like taking it one step at a time
39:02
right? You might go from ten to the
39:03
twenty-fourth power to twenty-fifth
39:05
twenty-sixth, twenty-seventh, and even if you're at ten to the twenty-fifth
39:09
your training efficiency can keep improving
39:12
making every bit of computing power produce more
39:14
intelligence
39:15
that's a bit like taking a few more
39:17
steps when you're climbing a mountain
39:18
right? So
39:21
I think overall it's more like
39:22
it's more like mountain climbing
39:23
right? Yeah
39:24
Right, why I ask this question isn't really important
39:26
your answer is more important
39:28
I'm very grateful to Gillian for taking the
39:30
time to share with us in our community today
39:32
and we all look forward
39:33
to you enjoying the climb more and more
39:36
and seeing even richer scenery
39:37
All right
39:38
let's give a warm round of applause for
39:39
Gillian. Thank you
39:40
Okay, thank you, everyone

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.