[00:00]
Hmm, why does this deep sea quirk exist?
[00:02]
I mean, it adds vision capabilities to
[00:05]
the deep sea AI system, but that's not
[00:07]
new. A lot of other AI systems have
[00:10]
vision capabilities. You just drop an
[00:12]
image here and it works. Even video and
[00:15]
even for open models. So, why do we need
[00:18]
this paper? Well, they did something
[00:21]
incredible here and it is an absolute
[00:24]
game changer. Why? You see, if you ask a
[00:27]
previous technique to count the number
[00:29]
of people in this photo, it will think
[00:32]
something like this. Okay, there are
[00:34]
people on the upper left and a bunch of
[00:36]
stripy guys in two rows. That is kind of
[00:40]
three rows. Some of them are standing,
[00:42]
some of them are sitting.
[00:44]
Ah, it's just so confusing to just count
[00:46]
them up using only words. Two problems
[00:49]
with this one. One, this is prone to
[00:51]
error. Two, you have to think a lot.
[00:55]
Just describing stuff. Why? What would
[00:58]
we, humans, do? Of course, we would use
[01:01]
our finger and would point at the image.
[01:04]
One, two, three, and so on.
[01:07]
Done. Don't describe images like a poet.
[01:10]
Point like a human. Now, that is exactly
[01:13]
what this new technique does. It allows
[01:16]
an AI system to point at things while
[01:19]
thinking and it is absolutely brilliant.
[01:22]
This makes it more accurate and it also
[01:25]
makes it faster. In a world where
[01:27]
hardware and tokens cost a fortune, it
[01:30]
is fantastic to have something that
[01:33]
gives us results faster and cheaper.
[01:35]
But, it turns out thinking with visual
[01:38]
primitives has even more advantages. It
[01:41]
can also do topological reasoning. For
[01:43]
instance, if you give it a maze with a
[01:46]
start and end point, you not only get a
[01:49]
correct answer to your questions, but
[01:51]
you can also trace back the whole
[01:54]
thought process visually.
[01:56]
I love that. Also, here you can ask
[01:59]
where the crown connects and look.
[02:02]
To the octopus. Yeah, it answers
[02:05]
correctly, but you can also see how it
[02:08]
came to that conclusion. Now, make no
[02:11]
mistake. These are simple examples. I'll
[02:14]
show you in a moment if it is as good as
[02:16]
these billion-dollar frontier models.
[02:19]
Also, if something goes wrong, this will
[02:21]
make it easier to find mistakes and fix
[02:24]
them to create an even better model.
[02:26]
This puts us one step closer to AI
[02:29]
systems we can actually understand that
[02:32]
do not just give us a soup of numbers.
[02:34]
So good. So, how good is it? Well, hold
[02:38]
on to your papers, fellow scholars, and
[02:41]
I dropped my papers here. Look, it needs
[02:43]
about 90% fewer visual tokens than most
[02:47]
frontier models. Now, wait, wait, wait.
[02:50]
It doesn't matter how little you think
[02:52]
if you just say three as an answer
[02:55]
without thinking. Thinking time doesn't
[02:57]
matter if it is incorrect. So, how
[03:00]
accurate is it?
[03:02]
Are you kidding me? This free system
[03:04]
matches or beats almost everything. And
[03:08]
once again, we are talking about this,
[03:10]
which is free, going up against
[03:12]
billion-dollar systems here. Wow. Now,
[03:16]
we are fellow scholars here, so at this
[03:18]
point we ask, are these results real?
[03:21]
You know, benchmarks are being gamed
[03:23]
left and right. Now, here is what many
[03:26]
people missed. Average over seven
[03:29]
benchmarks, but in-house benchmarks
[03:31]
excluded.
[03:33]
That is the key. They did not rig their
[03:35]
own benchmarks. You know why? Well,
[03:38]
everyone loves it because it's one of
[03:40]
the oldest tricks in the book. If you
[03:42]
are not performing well, just create a
[03:45]
new benchmark that fits you. Let's make
[03:47]
a YUNUS benchmark. You will always be
[03:50]
world first in being you. And this is
[03:53]
not the case here. Amazing. This is free
[03:56]
and open research. So, this technique
[03:58]
can potentially be added to many
[04:00]
existing models, including free ones.
[04:03]
This paper does not have a model
[04:05]
attached that I know of. It describes
[04:07]
the concept of how to do it in detail.
[04:10]
It's a blueprint, if you will. More
[04:12]
intelligence for all of us for free.
[04:16]
The world needs more papers like this.
[04:18]
Love it. But, this all sounds like
[04:21]
magic. How did they do this? Well, look,
[04:23]
this is their own policy distillation
[04:26]
objective. We need exactly this. You
[04:29]
see, normally, we have a bunch of expert
[04:32]
AI models. Now, at the risk of
[04:34]
simplifying things, imagine that one of
[04:37]
these guys is great at boxes. Nobody
[04:40]
does boxes better than this guy. The
[04:42]
other one is great at tracing mazes with
[04:45]
points. But, that's not what we want.
[04:48]
What we want is one AI that can do all
[04:51]
of these things. And that is where this
[04:53]
comes into play. We train a student
[04:56]
model that learns from all of these
[04:58]
teachers. It says what it would try to
[05:01]
do, then the teachers say, "Okay, here's
[05:04]
what I would have done." Do this enough
[05:06]
and the student will be pretty good at
[05:09]
all of these different kinds of visual
[05:11]
thinking. This is why they used the name
[05:13]
distilling the knowledge of a bunch of
[05:15]
expert teachers into a student. So,
[05:19]
where does this put us? Okay, so here's
[05:21]
what I think. Dear fellow scholars, this
[05:23]
is Two Minute Papers with Dr. Károly
[05:25]
Zsolnai-Fehér. You know, we always
[05:27]
thought that we would make AI systems
[05:29]
smarter by giving it higher resolution
[05:32]
images to train on. More pixels, more
[05:34]
smarts. It turns out not true.
[05:37]
Sometimes, that's not what we need at
[05:39]
all. Deep Seek just cut down those
[05:42]
visual tokens by 90% and still beat
[05:45]
frontier models. Less is more. Now, is
[05:48]
this perfect? All problems solved? No.
[05:52]
Limitations. One, the AI does not
[05:55]
automatically do this kind of pointy
[05:57]
thinking. It needs a word as a cue for
[06:00]
this kind of thinking. Two, bounding
[06:03]
boxes are nice for people, but if you
[06:05]
are counting blades of grass or strands
[06:08]
of hair, now, in this case, not having
[06:11]
those in very high resolution is a
[06:13]
problem.
[06:15]
>> [laughter]
[06:15]
>> Yep, once again, the two-minute papers
[06:18]
special, thin structures. Every time,
[06:22]
man. It's so painful. And three, this
[06:25]
kind of topological reasoning does not
[06:27]
generalize as well as we'd like. It
[06:30]
might not be as robust when you show it
[06:32]
something completely new. So, careful
[06:35]
with the misleading media headlines,
[06:37]
careful with the hype everywhere. There
[06:39]
is still plenty to improve here. But, I
[06:42]
feel that this might be a breakthrough.
[06:44]
And that makes it
[06:46]
maybe the third one this month in AI
[06:48]
research. What a time to be alive. Also,
[06:52]
with large AI companies going to IPO,
[06:55]
they are about to become ventures that
[06:57]
look to maximize their profits. More
[06:59]
money needed every quarter. So, it's
[07:02]
going to become more and more crucial to
[07:05]
own your own AI systems with free open
[07:07]
weights models. And this one makes them
[07:10]
better.
[07:11]
Love it. Here you see me running the
[07:13]
full DeepSeek AI model through Lambda
[07:16]
GPU Cloud. 671
[07:20]
billion parameters running super fast
[07:22]
and super reliably. This is insane. I
[07:26]
love it and I use it on a regular basis.
[07:29]
Lambda provides you with powerful Nvidia
[07:32]
GPUs to run your own chatbots and
[07:35]
experiments. Seriously, try it out now
[07:37]
at lambda.ai/papers
[07:40]
or click the link in the description.