00:00Hmm, why does this DeepSeq work exist?
00:02I mean, it adds vision capabilities to the DeepSeq AI system, but that's not new.
00:08A lot of other AI systems have vision capabilities, you just drop an image here, and it works.
00:14Even video, and even for open models.
00:17So why do we need this paper?
00:19Well, they did something incredible here, and it is an absolute game changer.
00:25Why?
00:25You see, if you ask a previous technique to count the number of people in this photo,
00:31it will think something like this.
00:33Okay, there are people on the upper left, and a bunch of stripy guys in two rows, that
00:39is kinda three rows.
00:40Some of them are standing, some of them are sitting, it's just so confusing to just count
00:46them up using only words.
00:48Two problems with this one.
00:50One, this is prone to error.
00:52Two, you have to think a lot.
00:54What?
00:54Just describing stuff, why?
00:58What would we humans do?
01:00Of course, we would use our finger, and would point at the image.
01:04One, two, three, and so on.
01:06Done.
01:07Don't describe images like a poet, point like a human.
01:11Now that is exactly what this new technique does.
01:14It allows an AI system to point at things while thinking, and it is absolutely brilliant.
01:22This makes it more accurate, and it also makes it faster.
01:26In a world where hardware and tokens cost a fortune, it is fantastic to have something
01:32that gives us results faster and cheaper.
01:35But, it turns out, thinking with visual primitives has even more advantages.
01:40It can also do topological reasoning, for instance, if you give it a maze.
01:45With a start and end point, you not only get a correct answer to your questions, but you
01:51can also trace back the whole thought process visually.
01:56I love that!
01:58Also, here, you can ask where the crown connects, and… look!
02:02To the octopus!
02:04Yeah!
02:05It answers correctly, but you can also see how it came to that conclusion.
02:10Now, make no mistake, these are simple examples.
02:14I'll show you in a moment if it is as good as these billion dollar frontier models.
02:19Also, if something goes wrong, this will make it easier to find mistakes and fix them to create
02:25an even better model.
02:26This puts us one step closer to AI systems we can actually understand that do not just
02:32give us a soup of numbers.
02:34So good!
02:35So, how good is it?
02:37Well, hold on to your papers, fellow scholars, and… I'll drop my papers here.
02:42Look, it needs about 90% fewer visual tokens than most frontier models.
02:48Now, wait, wait, wait.
02:50It doesn't matter how little you think if you just say three as an answer without thinking.
02:56Thinking time doesn't matter if it is incorrect.
02:59So, how accurate is it?
03:01Are you kidding me?
03:03This free system matches or beats almost everything.
03:07And once again, we are talking about this, which is free, going up against billion dollar
03:13systems here.
03:14Wow!
03:15Now, we are fellow scholars here, so at this point, we ask, are these results real?
03:21You know, benchmarks are being gamed left and right.
03:24Now, here is what many people missed.
03:27Average over seven benchmarks.
03:29But, in-house benchmarks excluded.
03:33That is the key.
03:34They did not rig their own benchmarks.
03:36You know why?
03:37Well, everyone loves it, because it's one of the oldest tricks in the book.
03:42If you are not performing well, just create a new benchmark that fits you.
03:46Let's make a you-ness benchmark.
03:49You will always be world first in being you.
03:52And, this is not the case here.
03:54Amazing.
03:55This is free and open research, so this technique can potentially be added to many existing models,
04:02including free ones.
04:03This paper does not have a model attached that I know of.
04:06It describes the concept of how to do it in detail.
04:10It's a blueprint, if you will.
04:12More intelligence for all of us, for free.
04:15The world needs more papers like this.
04:18Love it.
04:19But, this all sounds like magic.
04:21How did they do this?
04:22Well, look.
04:23This is their own policy, distillation objective.
04:27We need exactly this.
04:29You see, normally, we have a bunch of expert AI models.
04:33Now, at the risk of simplifying things, imagine that one of these guys is great at boxes.
04:40Nobody does boxes better than this guy.
04:42The other one is great at tracing mazes with points.
04:46But, that's not what we want.
04:48What we want is one AI that can do all of these things.
04:52And that is where this comes into play.
04:55We train a student model that learns from all of these teachers.
04:59It says what it would try to do.
05:01Then, the teachers say, okay, here's what I would have done.
05:05Do this enough, and the student will be pretty good at all of these different kinds of visual thinking.
05:11This is why they use the name distilling the knowledge of a bunch of expert teachers
05:16into a student.
05:18So, where does this put us?
05:20Okay, so here's what I think.
05:22Dear Fellow Scholars, this is Two Minute Papers with Dr. Karol Zsolnai-Fehér.
05:26You know, we always thought that we would make AI systems smarter by giving it higher resolution
05:32images to train on.
05:33More pixels, more smarts.
05:35It turns out not true.
05:37Sometimes, that's not what we need at all.
05:40DeepSeek just cut down those visual tokens by 90% and still beat frontier models.
05:46Less is more.
05:48Now, is this perfect?
05:49All problems solved.
05:51No.
05:52Limitations.
05:53One, the AI does not automatically do this kind of pointy thinking.
05:58It needs a word as a cue for this kind of thinking.
06:01Two, bounding boxes are nice for people, but if you are counting blades of grass or strands
06:08of hair, now, in this case, not having those in very high resolution is a problem.
06:14Ha ha ha.
06:16Yep, once again, the Two Minute Papers special.
06:19Thin structures.
06:21Every time, man.
06:22It's so painful.
06:24And three, this kind of topological reasoning does not generalize as well as we'd like.
06:29It might not be as robust when you show it something completely new.
06:34So, careful with the misleading media headlines, careful with the hype everywhere, there is still
06:40plenty to improve here.
06:41But, I feel that this might be a breakthrough.
06:44And that makes it, maybe the third one this month in AI research.
06:49What a time to be alive!
06:51Also, with large AI companies going to IPO, they are about to become ventures that look
06:57to maximize their profits.
06:59More money needed every quarter.
07:01So, it's going to become more and more crucial to own your own AI systems with free, open weights
07:08models.
07:09And this one makes them better.
07:11Love it!
07:11Here you see me running the full DeepSeq AI model through Lambda GPU Cloud.
07:19671 billion parameters!
07:21Running super fast and super reliably.
07:24This is insane!
07:25I love it!
07:27And I use it on a regular basis.
07:29Lambda provides you with powerful NVIDIA GPUs to run your own chatbots and experiments.
07:35Seriously!
07:36Try it out now at lambda.ai.papers or click the link in the description.
Comments