00:04unbelievable today is the day after gpt4
00:09it's great to have you here I'm
00:12delighted to have you I've known you a
00:14long time the journey and just my mental
00:17hit my my mental memory of your of the
00:21times that I've known you and the
00:23seminal work that you have done
00:25starting in Universal University of
00:29the co-invention of alexnet with Alex
00:35that led to the Big Bang of modern
00:39artificial intelligence your career that
00:42took you out here to the Bay Area the
00:47gpt123 and then of course chat GPT the
00:52AI heard around the world
00:55this is this is the incredible resume of
00:58a young computer scientist
01:00um you know an entire community and
01:02Industry at all with your achievements I
01:05I guess my I just want to go back to the
01:07beginning and ask you deep learning
01:11what was your intuition around deep
01:13learning why did you know that it was
01:15going to work did you have any intuition
01:18that are going to lead to this kind of
01:20okay well first of all thank you so much
01:24for the quote for all the kind words
01:27a lot has changed thanks to the
01:30incredible power of deep learning
01:33like I think this my personal starting
01:36point I was interested in artificial
01:38intelligence for a whole variety of
01:44an intuitive understanding of
01:46appreciation of its impact
01:48and also I had a lot of curiosity about
01:51what is consciousness what is The Human
01:56progress in artificial intelligence will
02:01well back then I was starting in 2002
02:05and it seemed like learning is the thing
02:08that humans can do that people can do
02:10that computers can't do at all in 2003
02:16computers could not learn anything
02:20and it wasn't even clear that it was
02:24and so I thought that making progress
02:28in learning in artificial learning in
02:32that would lead to the greatest progress
02:36and then I started to look around for
02:39and nothing seemed too promising
02:42but to my great luck
02:44Jeff Hinton was a professor at my
02:47and I was able to find him
02:50and he was working in neural networks
02:51and it immediately made sense
02:53because neural networks had the property
02:56that we are learning we are
02:59automatically programming parallel
03:02back then the parallel computers were
03:05but the promise was if you could somehow
03:06figure out how learning in neural
03:10then you can program small parallel
03:12computers from data and it was also
03:14similar enough to the brain and the
03:16brain works so it's like you had these
03:18several factors going for it now it
03:20wasn't clear how to get it to work
03:24but of all the things that existed that
03:27seemed like it had by far the greatest
03:30long-term promise even though in the
03:32time that you first started at the time
03:34that you first started working with deep
03:35learning and and neural networks what
03:38was what was the scale of the network
03:39what was the scale of computing at that
03:41moment in time what was it like an
03:43interesting thing to note is that the
03:45importance of scale wasn't realized back
03:48so people would just training on neural
03:50networks with like 50 neurons 100
03:52neurons several hundred neurons that
03:54would be like a big neural network
03:57a million parameters would be considered
03:59very large we would run our models on
04:02unoptimized CPU code because we were a
04:05bunch of researchers we didn't know
04:08we used Matlab the Matlab was optimized
04:12and we'd just experiment like what is
04:15the what is even the right question to
04:16ask you know so you try to want to
04:19to just find interesting phenomena
04:22interesting observation
04:24you can do this small thing you can do
04:26that small thing you know Jeff Hinton
04:28was really excited about
04:30training neural Nets on small little
04:33digits both for classification and also
04:36he was very interested in generating
04:38them so the beginnings of generating
04:40models were right there
04:42but the question is like okay so you've
04:43got all this cool stuff floating around
04:46what really gets traction
04:49it wasn't so it wasn't obvious that this
04:51was the right question
04:53back then but in hindsight that turned
04:56out to be the right question now now the
05:00Alex net was 2012. now
05:04you and Alex were working on Alex net
05:06for some time before then
05:08and and uh at what point at what point
05:12was it was it clear to you that you
05:14wanted to build a computer vision
05:17oriented neural network that imagenet
05:20was the right set of data to go for and
05:22to somehow go for the computer computer
05:29so I can talk about the context there
05:35two years before that it became clear to
05:40that supervised learning is what's going
05:43to get us the traction and I can explain
05:45precisely why it wasn't just an
05:50I would argue an irrefutable argument
05:52which went like this
05:54if your neural network is deep and large
06:00then it could be configured to solve a
06:04so that's the key word deep and large
06:08people weren't looking at large neural
06:10networks people were you know maybe
06:11studying a little bit of depth in neural
06:13networks but most of the machine
06:15learning field wasn't even looking at
06:16neural networks at all they were looking
06:18at all kinds of Bayesian models and
06:21which are theoretically elegant methods
06:24which have the property that they
06:26actually can't represent a good solution
06:28no matter how you configure them whereas
06:31the large and deep neural network can
06:33represent a good solution to the problem
06:36to find a good solution
06:39you need a big data set which requires
06:41it and a lot of compute to actually do
06:45the work we've also made advanced work
06:47so we've worked on optimization for for
06:49a little bit it was clear that
06:52optimization is a bottleneck
06:54and there was a breakthrough by another
06:56grad student in Jeff hinton's lab called
06:59James Martens and he came up with an
07:01optimization method which is different
07:02from the one we are doing now using now
07:04some second order method but the point
07:07about it is that it's proved that we can
07:10train those neurologics because before
07:11we didn't even know we could train them
07:12so if you can train them
07:15you make it big you find the data and
07:17you will succeed so then the next
07:19question is well what data and an
07:22imagenet data set back then it seemed
07:24like this unbelievably difficult data
07:27but it was clear that if we were to
07:30train a large convolutional neural
07:31network on this data set it must succeed
07:33if you just can have the compute and
07:36write that right at a time gpus you and
07:38I you and I are history and our paths
07:42intersected and somehow
07:45you had the the the the observations at
07:49a GPU and at that time we had this is
07:51our couple of generations into a Cuda
07:53GPU and I think it was GTX 580 uh
07:57generation you had the you're at the
07:59inside that the GPU could actually be
08:01useful for training your neural network
08:03models what was that how did that day
08:06start tell me you know you and I you
08:07never told me that moment you know how
08:09did that day start yeah so
08:13you know the the GP the gpus
08:16appeared in our in our lab in our
08:19thanks to Jeff and he said we should be
08:21good we should try this gpus and we
08:23started trying and experimenting with
08:26it was a lot of fun but but it was
08:31what to use them for exactly where are
08:33you going to get the real traction
08:36but then with the existence of the
08:42and then it was also very clear that the
08:44convolutional neural network is such a
08:46great fit for the GPU so it should be
08:48possible to make it go unbelievably fast
08:51and therefore train something which
08:54would be completely unprecedented in
08:58and that's how it happened and you know
09:00very fortunately Alex krajevski he
09:03really loved programming the GPS
09:06and he was able to do it he was able to
09:10to to program really fast
09:13convolutional kernels
09:19and then trained the neural network data
09:22set and that led to the result but it
09:24was like it shocked the world this
09:26shocked the world it it broke the record
09:29of a computer vision by such a wide
09:31margin that that it was a clear
09:34discontinuity yeah yeah and I wouldn't I
09:36would say it's not just like there is
09:37another bit of context there
09:40it's not so much like when we say break
09:42the record there is an important it's
09:44like I think there's a different way to
09:46phrase it it's that that data set was so
09:50obviously hard and so obviously outside
09:53of reach of anything
09:55people are making progress with some
09:57classical techniques and they were
09:58actually doing something but this thing
10:01was so much better on a data set which
10:03was so obviously hard it was it's not
10:05just that it's just some competition it
10:08was a competition which back in the day
10:10an average Benchmark it was so obviously
10:14difficult so obviously Out Of Reach and
10:19feed the property that if you did a good
10:21job that would be amazing
10:22Big Bang of AI fast forward to now uh
10:26you came out to the valley you started
10:28open AI with some friends
10:30um you're the chief scientist
10:34what was the first initial idea about
10:36what to work on at open AI because you
10:38guys worked on several things some of
10:40the trails of of inventions and and work
10:43uh you could you could see led up to
10:50um but what were the initial inspiration
10:53what were you how would you approach
10:56intelligence from that moment and led to
11:02obviously when we started it wasn't
11:06100 clear how to proceed
11:09and the field was also very different
11:12compared to the way it is right now
11:15so right now you already used we already
11:19you have these amazing artifacts these
11:22amazing neural Nets who are doing
11:23incredible things and everyone is so
11:26excited but back in 2015 2016 early 2016
11:31when you were starting out
11:33the whole these things seemed pretty
11:36there were so many fewer researchers
11:40100 maybe they were between a hundred
11:42and a thousand times fewer people in the
11:44field compared to now yeah like back
11:50most of them are working in Google slash
11:53deepmind and that was that and then
11:55there were people picking up the skills
11:57but it was very very scarce very rare
12:07at the start of open AI
12:09that state that had a lot of staying
12:11power and they stayed with us to this
12:13and I'll describe them right now
12:16the first big idea that we had
12:21I was especially excited about very
12:23early on is the idea
12:29unsupervised learning through
12:36today we take it for granted that
12:38unsupervised learning is this easy thing
12:39you just pre-train on everything and it
12:41all does exactly as you'd expect
12:46unsupervised learning was an unsolved
12:49problem in machine learning that no one
12:53had any insight exactly any clue as to
12:56what to do that's right iyanla Khan
12:58would go around and give a talk give
13:00talk saying that you have this Grand
13:02Challenge and supervised learning
13:05and I really believed
13:07that really good compression of the data
13:10will lead to unsupervised learning
13:15compression is not language that's
13:17commonly used to describe what is really
13:19being done until recently when suddenly
13:22it became apparent to many people that
13:25those gpts actually compress the
13:28you may recall that Ted Chiang Neo Times
13:31article which also alluded to this
13:33but there is a real mathematical sense
13:35in which training these autoregressive
13:38generative models compress the data and
13:42intuitively you can see why that should
13:44if you compress the data really well you
13:46must extract all the hidden secrets
13:50therefore that is the key
13:53so that was the first idea that we're
13:54really excited about and that led to
13:58quite a few Works in openai
14:01to the sentiment neuron
14:06very briefly it is not this work might
14:10outside of the machine learning field
14:13but it was very influential especially
14:20like the result there
14:23was that when you train a neural network
14:26back then it was not a Transformer it
14:28was before the Transformer right small
14:29recurrent neurological lstm sequence
14:33work you've done I mean there's some of
14:34your some of the words that you've done
14:37so the same lstm with a few twists
14:40trying to predict the next token in
14:43Amazon reviews next character
14:46and we discovered that if you predict
14:48the next character well enough
14:50it will be a neuron inside that lstm
14:53that corresponds to its sentiment
14:56so that was really cool because it
14:59showed some traction for unsupervised
15:02learning and it validated the idea that
15:09character prediction next something
15:11prediction compression yeah
15:13has the property that it discovers the
15:16secrets in the data that's what we see
15:18with these GPT models right you train
15:19and people say just statistical
15:21correlation I mean at this point it
15:23should be so clear to anyone that
15:26you know for me intuitively open up the
15:29whole world of where do I get the data
15:31for unsupervised learning
15:34because I do have a whole lot of data if
15:36I could just make you predict the next
15:38character and I know what the ground
15:39truth is I know what the answer is I
15:42could be I could train a neural network
15:43model with that so that that observation
15:46and masking and other other technology
15:48other approaches you know open open my
15:51mind about where would the world get all
15:53the data that's unsupervised for
15:55unsupervised learning well I think I
15:57think so I would I would phrase it a
16:01I would say that within supervised
16:03learning the hard part has been less
16:07where you get the data from though that
16:10part is there as well especially now but
16:14why should you do it in the first place
16:17why should you bother
16:20the hard part was to realize
16:25these neural Nets to predict the next
16:28is a worthwhile goal at all they would
16:31learn a representation
16:32that it would it would be able to
16:34understand that's right but it will be
16:36use grammar and yeah
16:38but to actually to actually
16:40it just wasn't obvious right so people
16:43weren't doing it but the sentiment
16:45neuron work and you know I want to call
16:47out alakratford is a person who really
16:51was responsible for many of the advances
16:54the sentiment this this was this was
16:57before GPT one it was the precursor to
16:59gpt-1 and it influenced our thinking a
17:01lot then the Transformer came out
17:04and we immediately went oh my God this
17:09let me try and GPT one
17:14you've always believed that scaling will
17:18improve the performance of these models
17:20yes larger larger networks deeper
17:23networks more training data would scale
17:26um there was a very important paper that
17:29open AI wrote about the scaling laws and
17:32the relationship between loss and the
17:36size of the model and the amount of data
17:37set the size of the data set when
17:40Transformers came out it gave us the
17:42opportunity to train very very large
17:43models in very reasonable amount of time
17:46but what did the intuition about about
17:51the scaling laws or the size of of
17:55and your journey of gpt123 which came
18:00first did you see the evidence of GPT
18:02one through three first or was there the
18:05intuition about the scaling law first
18:07the intuition so I would say that the
18:10way the way I'd phrase it is that I had
18:13a very strong belief that bigger is
18:18and that one of the goals that we had at
18:21open AI is to figure out how to use the
18:26there was a lot of belief about an open
18:28AI about scale from the very beginning
18:33what to use it for precisely
18:35because I'll mention right now we're
18:37talking about the gpts but there is
18:38another very important line of work
18:40which I haven't mentioned the second big
18:41idea but I think now is a good time to
18:43make a detour and that's reinforcement
18:48that clearly seems important as well
18:49what do you do with it
18:52so the first really big project that was
18:59our effort at solving a real-time
19:05and for context a real-time strategy
19:07game is like it's a competitive sport
19:09yeah right we need to be smart you need
19:12to have faster you need to have a quick
19:14reaction time you there's steam work
19:16and you're competing against another
19:18team and it's pretty it's pretty it's
19:22and there is a whole competitive league
19:25for that game the game is called DotA 2.
19:27and so we train the reinforcement
19:29learning agent to play against itself
19:36with the goal of reaching a level so
19:39that it could compete against the best
19:41players in the world and I was a major
19:44undertaking as well it was a very
19:46different line it was reinforcement
19:48learning yeah I remember the data that
19:50you guys announced that work and this is
19:52this by the way when I was asking
19:54earlier about about there's a there's a
19:56large body of work that have come out of
19:57open AI some of it seem like detours
20:01um but but in fact as you're explaining
20:03now they might might have been detours
20:05it seemingly detours but they they
20:07really led up to some of the important
20:08work that we're now talking about chat
20:09GPT yeah I mean there has been real
20:12convergence where the gpts
20:16produce the foundation and in the
20:19reinforcement learning from DOTA morphed
20:22into reinforcement learning from Human
20:24feedback that's right and that
20:25combination gave us chat GPT you know
20:28there's a there's a there's a
20:29misunderstanding that that uh chat GPT
20:33is uh in itself just one giant large
20:36language model there's a system around
20:38it that's fairly complicated is it could
20:41could you could you explain
20:43um briefly for the audience the the uh
20:46the fine-tuning of it the reinforcement
20:48learning of it the the
20:50um uh you know the various surrounding
20:53systems that allows you to keep it on
20:55Rails and and let it let it uh give it
20:59knowledge and you know so on so forth
21:04so the way to think about it
21:06is that when we train
21:08a large neural network
21:10to accurately predict the next word
21:14in lots of different texts from the
21:16internet what we are doing is that we
21:20are learning a world model
21:22it looks like we are learning this it
21:24may it may look on the surface but we
21:27are just learning statistical
21:28correlations in text
21:31but it turns out that to just learn the
21:35statistical correlations in text
21:37to compress them really well
21:39what the neural network learns is some
21:42representation of the process that
21:46this text is actually a projection of
21:50there is a world out there
21:52and it has a projection on this text
21:55and so what the neural network is
21:56learning is more and more aspects of the
22:00world of people of the human conditions
22:03their their their hopes dreams and
22:06motivations their interactions and the
22:08situations that we are in and the neural
22:12network learns a compressed abstract
22:15usable representation of that this is
22:19what's being learned from accurately
22:21predicting the next word
22:23and furthermore the more accurate you
22:26are is predicting the next word
22:28the higher Fidelity the more resolution
22:32you get in this process
22:33so that's what the pre-training stage
22:37but what this does not do
22:40is specify the desired behavior that we
22:43wish our neural network to exhibit
22:47you see a language model what it really
22:51is to answer the following question
22:54if I had some random piece of text on
22:57the internet which starts with some
23:02what will it complete to if you just
23:05randomly ended up on some text from the
23:09but this is different from well I want
23:11to have an assistant which will be
23:13truthful that will be helpful that will
23:16follow certain rules and not violate
23:18them that requires additional training
23:21this is where the fine tuning and the
23:24reinforcement learning from Human
23:26and other forms of AI assistance it's
23:29not just reinforcement learning from
23:30Human teachers it's also reinforcement
23:32learning from human and AI collaboration
23:34our teachers are working together with
23:36an AI to teach Rai to behave
23:39but here we are not teaching it new
23:40knowledge this is not what's happening
23:45we are communicating with it we are
23:48communicating to it what it is that we
23:51want it to be and this process the
23:54second stage is also extremely important
23:57the better we do the second stage the
24:00more useful the more reliable this
24:02neural network will be so the second
24:04stage is extremely important too in
24:06addition to the first stage of the learn
24:10learn everything learn as much as you
24:13can about the world from the projection
24:15of the world which is text now you could
24:18tell you could you could fine tune it
24:20you could instruct it to perform certain
24:24things can you instruct it to not
24:26perform certain things so that you could
24:27give it guard rails about avoid these
24:31um you know give it some kind of a
24:32bounding box so that so that it doesn't
24:34it doesn't wander out of that bounding
24:37box and and perform things that that are
24:39you know unsafe or otherwise yeah so
24:44this second stage of training is indeed
24:47where we communicate to the neural
24:49network anything we want which includes
24:52the bowling box and the better we do
24:56this training the higher the Fidelity
24:58with which we communicate is about inbox
25:01and so with constant research and
25:03Innovation on improving this fidelity
25:07we are able to improve this fidelity
25:10and so it becomes more and more reliable
25:14and precise in the way in which it
25:19intended intended instructions chat gbt
25:22came out just a few months ago
25:24um fastest growing application in the
25:30lots of lots of uh uh interpretations
25:36but some of the some of the things that
25:38that is clear it is it is the easiest
25:42application that anyone has ever created
25:44for anyone to use it performs tasks it
25:49performs things it does things that are
25:52Beyond people's expectation
25:55anyone can use it there are no
25:57instruction sets there are no wrong ways
25:59to use it you you just use it and if
26:03it's if your instructions are our
26:05prompts are ambiguous the conversation
26:08refines the ambiguity until your intents
26:11are understood by by the by the
26:14application by the AI
26:16the the impact of course uh clearly or
26:21remarkable now yesterday this is the day
26:25after gpt4 just a few months later the
26:30the performance of gpt4 in many areas
26:33astounding SAT scores GRE scores bar
26:38exams the number of the number of
26:42tests that it's able to perform at very
26:46capable levels very capable human levels
26:50what were the what were the major
26:52differences between chat GPT and gpt4
26:56that led to its improvements in these in
27:07pretty substantial Improvement on top of
27:10chat GPT across very many dimensions
27:14we trained gpt4 I would say
27:19between more than six months ago maybe
27:22eight months ago I don't remember
27:26GPT is the first build big difference
27:29between Shad GPT and gpt4
27:32and that's perhaps is the more
27:35the most important difference
27:37is that the base on top of gpt4 is built
27:40predicts the next word
27:43with greater accuracy this is really
27:47because the better a neural network can
27:50predict the next word in text the more
27:52it understands it this claim is now
27:55perhaps accepted by many at this point
27:58but it might still not be intuitive or
28:01is not completely intuitive as to why
28:03so I'd like to take a small detour and
28:05to give an analogy that will hopefully
28:07clarify why more accurate prediction of
28:11the next word leads to more
28:13understanding real understanding
28:16let's consider an example
28:18say you read a detective novel
28:22complicated plot a storyline different
28:23characters lots of events Mysteries like
28:27Clues it's unclear then let's say that
28:31at the last page of the book the
28:33detective has got all the clues gathered
28:35all the people and saying okay I'm going
28:37to reveal the identity of whoever
28:40committed the crime and that person's
28:42name is predict that word predict that
28:45word exactly my goodness right yeah
28:48there are many different words but by
28:50predicting those words better and better
28:52and better the understanding
28:54of the text keeps on increasing gpt4
28:58predicts the next word better I tell you
29:00people say that the Deep learning won't
29:04lead to reasoning that deep learning
29:06won't lead to reasoning but in order to
29:09predict that next word figure out from
29:11all of the agents that were there and
29:16you know strengths or weaknesses or
29:19their intentions and the context and to
29:23be able to predict that word who who was
29:26the murderer that requires some amount
29:29of reasoning a fair amount of reasoning
29:31and so so how did how did the how is it
29:36that it's able to pre to learn reasoning
29:39and and if if it learn reasoning
29:42um you know one of the one of the things
29:44that I was going to ask you is of all
29:45the tests that were that were taken
29:48um between Chad GPT and gbd4 there were
29:52that gpt3 or Chan GPT was already very
29:56there were some tests that gbd3 or chai
29:59GB was not as good at
30:01um that gbt4 was much better at and
30:04there were some tests that neither are
30:07I would love for it you know and some of
30:09it has to do with reasoning it seems
30:11that you know maybe in calculus that
30:13that it wasn't able to break maybe the
30:17um into into its reasonable steps and
30:19solve it is is it but yet in some areas
30:23it seems to demonstrate reasoning skills
30:26and so is that an area that that um uh
30:30that in predicting the next word you're
30:32you're learning reasoning and um uh what
30:36are the limitations uh now of gpt4 that
30:39would enhance his ability to reason even
30:45reasoning isn't this super well-defined
30:48concept but we can try to Define it
30:53which is when you maybe maybe when you
30:57where you're able to somehow think about
31:00it a little bit and get a better answer
31:03because of your reasoning
31:05and I'd say I'd say that there are
31:10you know maybe there is some kind of
31:12limitation which could be addressed
31:15by for example asking the neural network
31:17to think out loud this has proven to be
31:19extremely effective for reasoning but I
31:21think it also remains to be seen just
31:23how far the basic neural network will go
31:25I think we have yet to uh
31:28tap fully tap out its potential
31:35but yeah I mean there is definitely some
31:37sense where reasoning is still not
31:43as some of the other capabilities of the
31:46neural network though we would like the
31:48reasoning capabilities of the neural
31:50network to be high higher
31:53I think that it's fairly likely that
31:56business as usual will keep will improve
31:58the reasoning capabilities of the neural
32:02I wouldn't necessarily confidently rule
32:05out this possibility
32:06yeah because one of the things that that
32:09is really cool is you ask you as a
32:12treasury video question but before it
32:14answers the question tell me first first
32:16of what you know and then to answer the
32:19um you know usually when somebody
32:20answers a question if you give me the
32:22the foundational knowledge that you have
32:24or the foundational assumptions that
32:26you're making before you answer the
32:27question now that really improves the my
32:30believability of of the answer you're
32:33also demonstrating some level of Reason
32:35well you're demonstrating reasoning and
32:37so it seems to me that chat GPD has this
32:39inherent capability embedded in it
32:42yeah to some degree yeah this cup the
32:45the the the the the way the one way to
32:47think about what's happening now is that
32:50these neural networks have a lot of
32:52these capabilities they're just not
32:54quite very reliable in fact you could
32:57say that reliability is currently the
33:00single biggest obstacle for these neural
33:02networks being useful truly useful if
33:06sometimes it is still the case
33:09that these neural networks
33:12hallucinate a little bit
33:13or maybe make some mistakes which are
33:15unexpected which you wouldn't expect a
33:19it is this kind of unreliability that
33:22makes them substantially less useful
33:24but I think that perhaps with a little
33:27bit more research with the current ideas
33:28that you have and perhaps a few more of
33:32ambitious research plans will be able to
33:35achieve higher reliability as well and
33:37that will be truly useful that will
33:39allow us to have very accurate guard
33:43rails which are very precise that's
33:45right and it will make it ask for
33:47clarification where it's unsure or maybe
33:52say that it doesn't know something when
33:54it does anything he doesn't know and do
33:57so extremely reliably so I'd say that
34:01some of the bottlenecks really so it's
34:03not about whether it exhibits some
34:06particular capability but more how
34:08reliable exactly yeah you know one is
34:11speaking of speaking of factualness and
34:14faithfulness uh hallucination I I I saw
34:19in in uh one of the videos uh a
34:22demonstration that that
34:24um uh links to a Wikipedia page uh to it
34:28does retrieval capability uh has ever
34:31been included in the gpt4 is it able to
34:34retrieve information from a factful
34:36place that that could augment its
34:39response to you so the current gpt4
34:45does not have a built-in retrieval
34:47capability it is just a
34:50really really good next word predictor
34:54which can also consume images by the way
34:56we haven't spoken about yeah it is
34:59really good at images which is also then
35:01fine-tuned with data and
35:05various reinforcement learning variants
35:08to behave in a particular way
35:10it is perhaps I'm sure someone will it
35:16wouldn't surprise me if some of the
35:17people who have access could perhaps
35:18request gpt4 to maybe make some queries
35:23and then populate the results inside
35:25inside the context because also the
35:27context duration of gpd4 is quite a bit
35:29longer now yeah that's right so
35:35although lgbt4 does not support built-in
35:39retrieval it is completely correct
35:43that it will get better with retrieval
35:45yeah yeah multi-modality gpt4 has the
35:49ability to learn from text and images
35:52and respond to input from text and
35:59the foundation of multi-modality
36:04of course Transformers has made it
36:07possible for us to learn from
36:08multimodality tokenized text and images
36:15yeah but at the foundational level help
36:18us understand how multimodality enhances
36:21the understanding of the world
36:25um Beyond text by itself
36:27and uh and my understanding is that that
36:31that when you when you um I do
36:35multi-modality learning that even when
36:39it is just a text prompt the text prompt
36:41the text understanding could actually be
36:45um tell us about multi-modality at the
36:48foundation why it's so important and and
36:50um what was what's the major
36:51breakthrough in the the and the
36:53characteristic differences as a result
36:56so there are two Dimensions to
36:59two reasons why it is interesting
37:06a little bit humble the first reason is
37:09that multi-modality is useful
37:12it is useful for a neural network to see
37:15Vision in particular
37:17because the world is very visual human
37:20beings are very visual animals
37:23I believe that a third of the visual
37:26core of the human cortex is dedicated to
37:33by not having vision
37:36the usefulness of our neural networks
37:38though still considerable
37:40is not as big as it could be so it is a
37:44very simple usefulness argument it is
37:51and gpt4 can see quite well
37:54the there is a second reason to division
37:58which is that we learn more about the
38:02by learning from images in addition to
38:07that is also a powerful argument though
38:11it is not as clear-cut as it may seem
38:14I'll give you an example
38:16or rather before giving an example I'll
38:18make the general comment
38:21for a human being us human beings we get
38:24to hear about 1 billion words
38:29only only one billion words that's
38:31amazing yeah that's not a lot yeah
38:35so we need to company we need does that
38:37include my own words in my own head
38:40make it 2 billion but you see what I
38:43mean yeah you know you can see that
38:47um a billion seconds is 30 years so you
38:51can kind of see like we don't get to see
38:52more than a few words a second then if
38:54you're asleep half the time so like a
38:56couple billion words is the total we get
38:58in our entire life so it becomes really
39:00important for us to get as many sources
39:02of information as we can and we
39:04absolutely learn a lot more from vision
39:07the same argument holds true for our
39:10neural networks as well
39:13except for the fact that the neural
39:15network can learn from so many words
39:19things which are hard to learn about the
39:22world from text in a few billion words
39:28from trillions of words and I'll give
39:37one needs to see to understand colors
39:40and yet the text only neural networks
39:44who've never seen a single Photon in
39:49if you ask them which colors are more
39:51similar to each other it will know that
39:53red is more similar to Orange than to
39:55Blue it will know that blue is more
39:57similar to purple than to Yellow
40:01how does that happen
40:03and one answer is that information about
40:06the world even the visual information
40:08slowly leaksane through text
40:12slowly not as quickly
40:13but then you have a lot of text you can
40:16of course once you also add vision and
40:21learning about the world from Vision you
40:22will learn additional things which are
40:24not captured in text but it is you know
40:27I would not say that it is a binary
40:29there are things which are impossible to
40:32learn from the from text only I think
40:34this is more of an exchange rate and in
40:37particular as you want to learn if we
40:39are if you if you are if you are like a
40:42human being and you want to learn from a
40:45or a hundred million words then of
40:47course the other sources of information
40:48become far more important
40:52yeah and so so the the you learn from
40:57is there is there a sensibility that
40:59that would suggest that if we wanted to
41:03um also the construction of the world as
41:05in you know the arm is connected to my
41:07shoulder that my elbow is connected that
41:10somehow these things move the the the
41:12the animation of the world the physics
41:15of the world if I wanted to learn that
41:17as well can I just watch videos and
41:21and and if I wanted to augment all of
41:24that would sound like for example if
41:27um the meaning of of great
41:29uh great could be great
41:32or great could be great you know so one
41:36is sarcastic one is enthusiastic uh
41:39there are many many words like that you
41:41know uh that's sick or you know I'm sick
41:46or I'm sick depending on how people say
41:49would audio also make a contribution to
41:52the learning of the the model and and
41:54could we put that to good use soon yes
41:57yeah I think I think it's definitely the
42:00case that well you know what can we say
42:02about audio it's useful it's an
42:05additional source of information
42:06probably not as much as
42:11there is another there is a case to be
42:13made for the usefulness of audio as well
42:15both on the recognition side and on the
42:19when you when you um uh
42:22on the on the context of the scores that
42:25um the thing that was really interesting
42:27was was uh the the data that you guys
42:30published which which one of the tests
42:32were were um uh performed well by GPT
42:35three and which one of the tests
42:37performs substantially better with gbt4
42:40um how did multi-modality contribute to
42:44those tests you think oh I mean
42:47in a pretty straightforward
42:48straightforward way anytime there was a
42:51test where a problem
42:53would were to understand the problem you
42:55need to look at a diagram like for
42:57example in some math competitions like
43:01cont math competition for high school
43:03students called AMC 12 right
43:08presumably many of the problems have a
43:12so GPT 3.5 does quite badly on that on
43:17that text on that on the test gpt4 with
43:22I think I don't remember but it's like
43:24maybe from two percent to 20 accuracy of
43:28success rate but then when you add
43:30Vision it jumps to 40 success rate so
43:32the vision is really doing a lot of work
43:34the vision is extremely good and
43:37I think being able to reason visually as
43:40well and communicate visually will also
43:42be very powerful and very nice things
43:45which go beyond just learning about the
43:48world you have several things you got to
43:49learn you can learn about the world
43:51you can then reason about the world
43:53visually and you can communicate
43:55visually where now in the future perhaps
43:58in some future version if you ask your
43:59neural net hey like explain this to me
44:01rather than just producing four
44:03paragraphs it will produce hey like fear
44:05is like a little diagram which clearly
44:07conveys to you exactly what you need to
44:09know and so that's incredible you know
44:12one of the things that you said earlier
44:13about about an AI generating generating
44:16uh tests to train another AI
44:19um you know there's there was a paper
44:21that was written about and I I don't I
44:24don't completely know whether whether
44:25it's factual or not but but um that
44:27there's there's a total amount of
44:29somewhere between 4 trillion to
44:31something like 20 trillion useful you
44:34know tokens in in language tokens that
44:39that the world will be able to train on
44:41you know over some period of time and
44:43that would have run out of tokens to
44:44train and and um I I well first of all I
44:48wonder if that's you feel the same way
44:51and then this secondary secondarily
44:53whether whether the AI generating its
44:59data could be used to train the AI
45:03itself which you could argue is a little
45:06circular but we train our brain with
45:12data all the time by self-reflection
45:16working through a problem in our brain
45:19uh you know and and uh or you know some
45:23I guess I guess neuroscientists suggest
45:25sleeping we do a lot of fair amount of
45:28you know developing our neurons
45:30um how do you see this this area of
45:32synthetic data generation is that going
45:34to be an important part of the future of
45:36training Ai and and the AI teaching
45:41like I wouldn't underestimate the data
45:44that exists out there I think this
45:48I think is probably more more data Than
45:52and as to your second question certainly
45:55a possibility remains to be seen yeah
45:59yeah it's it really does seem that that
46:02um one of these days our AIS are are
46:05you know when we're not using it maybe
46:08generating either adversarial content
46:10for itself to learn from or imagine
46:12solving problems that that it can go off
46:15and and then and then improve itself
46:17tell us whatever you can about about uh
46:23where we are now and and what do you
46:25think will be and and not not too
46:27distant future but you know pick pick
46:30your your horizon a year or two what do
46:33you think this whole language Model area
46:35would be in some of the areas that
46:36you're most excited about you know
46:38predictions are hard
46:41speed it's a bit although it's a little
46:45say things which are too specific
46:47I think it's safe to assume
46:52progress will continue
46:55and that we will keep on seeing systems
46:58in there in the things that they can do
47:02and the current Frontiers are will be
47:04centered around reliability around the
47:08system can be trusted really get into a
47:11point where you can trust what it
47:14really get into a point where if it
47:16doesn't understand something it asks for
47:20says that it doesn't know something says
47:23that it needs more information I think
47:25those are perhaps the biggest the areas
47:28where Improvement will lead to the
47:30biggest impact on the usefulness of
47:33those systems because right now that's
47:35really what stands in the way you have
47:37an area of asking neural network you ask
47:38a neural net to maybe summarize some
47:40long document and you get a summary
47:43like are you sure that some important
47:45detail wasn't omitted it's still a
47:47but it's a different story when you know
47:49with all the important points have been
47:54at some point like and in particular
47:56it's okay like if some if there is
47:58ambiguity it's fine but
48:00if a point is clearly important such
48:02that anyone else who saw that point
48:04would say this is really important
48:06when the neural network will also
48:07recognize that reliably that's when you
48:10same for the guardrail say same for its
48:12ability to clearly follow the intent of
48:15the user of its operator
48:18so I think we'll see a lot of that in
48:20the next two years yeah that's terrific
48:21because the progress in those two areas
48:24this technology trusted by people to use
48:28and be able to apply for so many things
48:30I was thinking that was going to be the
48:31last question but I did have another one
48:33sorry about that so Chad uh chat gbt to
48:36gpt4 gpt4 when when it first when you
48:40first started using it uh what are some
48:44it demonstrated that surprised even you
48:51there were lots of really cool things
48:52that it demonstrated which
48:57which is which were quite cool and
49:02it was quite good so I'll mention two
49:05excess so let's see I'm just I'm just
49:08about the best way to go about it
49:12is that the level of its reliability was
49:16where the previous neural networks if
49:19you ask them a question sometimes they
49:22misunderstand something in a kind of a
49:26silly way whereas the gpt4 that stopped
49:29its ability to solve math problems
49:31became far greater it's like you could
49:34really like say sometimes you know
49:35really do the derivation and like long
49:38complicated derivation you can convert
49:40the units and so on and that was really
49:41cool you know like many people it works
49:44who are proof it works through a proof
49:45it's pretty amazing not all proofs yeah
49:47naturally but but quite a few or another
49:50example would be like many people
49:52noticed that it has the ability to
49:56produce poems with you know every word
49:59starting with the same letters or every
50:01word starting with some it follows
50:04instructions really really clearly not
50:06perfectly still but much better before
50:08yeah really good and on the vision side
50:10I really love how it can explain jokes
50:13it can explain memes you show it a meme
50:16and ask it why it's funny and it will
50:18tell you and it will be correct the the
50:21vision part I think is very
50:24was also very it's like really actually
50:26seeing it when you can ask questions
50:28follow-up questions about some
50:30complicated image with a complicated
50:32diagram and get an explanation that's
50:35but yeah overall I will say to take a
50:39you know I've been I've been in this
50:41business for quite some time actually
50:43like almost exactly 20 years
50:49the thing which most of which I find
50:51most surprising is that it actually
50:55yeah like it it's turned out to be the
50:58same little thing all along which is no
51:00longer little and it's a lot more
51:02serious and much more intense but it's
51:05the same neural network just larger
51:07trained on maybe larger data sets in
51:10different ways with the same fundamental
51:12training algorithm yeah
51:17I would say this is what I find the most
51:19surprising yeah whenever I take a step
51:21back I go how is it possible those ideas
51:23those conceptual ideas about well
51:25the brain has neurons so maybe
51:28artificial neurons are just as good and
51:30so maybe we just need to train them
51:31somehow with some learning algorithm
51:33that those arguments turned out to be so
51:39the biggest surprise I'd say in the in
51:42the 10 years that we've known each other
51:44uh your your uh the near the models that
51:48you've trained and the amount of data
51:50you've trained from uh the what you did
51:54on alexnet to now is about a million
51:57times and and uh uh no no one in the
52:02world of computer science would have
52:04would have believed that the amount of
52:06computation that was done in that 10
52:08years time would be a million times
52:10larger in that that you dedicated your
52:13career to go go do that
52:16um you've done two uh many more uh your
52:20body of work is incredible but two
52:22seminal works and the invention the
52:24co-invention with Alex Ned and that that
52:26early work and and now with uh GPT at
52:29open AI uh it is it is truly remarkable
52:33what you've accomplished and it's great
52:35to catch up with you again Ilya my good
52:37friend and and um uh it is uh it is
52:41quite an amazing moment and it's a
52:43today's today's talk the way you you uh
52:46break down the problem and describe it
52:48uh this is one of the one of the uh the
52:51best PhD Beyond PhD descriptions of the
52:55state of the art of large language
52:56models I really appreciate that it's
52:58great to see you congratulations thank
52:59you so much yeah thank you so I had so