00:00there is a lot of Focus right now on
00:01building more models but you know
00:04building good products on top of these
00:06models is incredibly difficult
00:36I would love it if you're comfortable
00:38it's giving kind of the longer form of
00:40your background what brought you to open
00:41AI you know just kind of bring us up to
00:43speed and then we'll go from there yeah
00:47it was born in Albania just after the
00:51fall of Communism very interesting times
00:54in this you know very isolated countries
00:57sort of similar to North Korea today and
01:01I bring that up because it was I think
01:04very Central to sort of my
01:07education and focus in math and sciences
01:11because there was a lot of focus on
01:17in in Coast communist Albania and you
01:20know the humanities like history and
01:23sociology and these type of topics they
01:27were a bit questionable like the source
01:29of information and truthfulness was uh
01:32was hard it was ambiguous so anyway I I
01:36got very interested in math and sciences
01:39and that's what I pursued relentlessly
01:42until you know still still working
01:44fundamentally in mathematics
01:47and over time my interests grew more
01:51from this theoretical space into
01:54actually building things and figuring
01:58apply that knowledge to build stuff and
02:01I studied mechanical engineering
02:04and went on to work in Aerospace as an
02:07engineer and then joined Tesla shortly
02:11after where I spent a few years
02:13initially I joined to work on Model S
02:16dual motor and then I went on to model X
02:21from the early days of the initial
02:24design and eventually led the whole
02:29and this is when I got very interested
02:32in applications of AI and
02:35specifically with autopilot and so
02:39I started thinking more and more about
02:40different applications of AI okay what
02:43happens when you when you you're using
02:46Ai and computer visually in a different
02:48domain instead of autopilot
02:51and after Tesla I went on to work on
02:56augmented reality and virtual reality
02:59because I just wanted to get experience
03:02with different domains and I thought at
03:08the time that it was the right time to
03:10actually work on special computing
03:12obviously in retrospect too early
03:16back then but anyways I learned a lot
03:18about the limitations of pushing this
03:22technology to the practicality of using
03:25it every day and at this point I started
03:28thinking more about what happens if you
03:31just focus on the generality like forget
03:33the competence in different domains and
03:36just focus on generality and there were
03:38two places at the time there were laser
03:42focused on this issue and open Ai and
03:47uh and I was very drawn to open AI
03:50because of its Mission and I felt like
03:53there's not going to be a more important
03:55technology that we all built than than
04:00back then I certainly did not have the
04:02same conviction about it as I do now
04:06but I thought that fundamentally if
04:10you're building intelligence it's such a
04:14it is such a core unit then the
04:17university it affects everything and so
04:21you know what what else is there is
04:23there to do more inspiring than than
04:26Elevate and increase Collective
04:29collective intelligence of humanity
04:32whenever I meet somebody that's a like a
04:35um uh influencer and has done major
04:38contributions to the space they almost
04:40invariably have a physics background or
04:41a math background which is actually very
04:43different than it was 15 years ago like
04:4415 years ago I was like you know the
04:46kind of you know it was like engineers
04:47and you know they came from lexical
04:49engineering mechanical engineering but
04:52um you know there's something and I
04:54don't know if it's like some like quirk
04:56in the network or like it's it's more
04:58fundamental like systemic and I mean do
05:01you think that this is kind of the time
05:02for the physicist to step up and kind of
05:04contribute to computer science and
05:05there's something about that or do you
05:07think it's just more of a coincidence
05:09so I think maybe one thing to draw on
05:13from the theoretical space of math
05:17but also the kind of the natural
05:19problems with math is that you know you
05:22kind of need to sit with a problem for a
05:24really long time and you have to think
05:26about it sometimes to sleep and you wake
05:29up and you have a new idea and over the
05:31course of maybe a few days sometimes or
05:34weeks you get to the final solution and
05:37so it's not like a quick reward
05:39and sometimes it's not this iterative
05:42thing and and I think it's almost like a
05:44different way of thinking where you're
05:47building an intuition but also a sort of
05:50discipline to sit with the problem and
05:53have faith that you're going to solve it
05:54and over time you build an intuition on
05:58what problem is the right problem to
06:00actually work on so do you think it's
06:02now more of a systems problem more like
06:04kind of more of an engineering problem
06:05or do you think that we still have a lot
06:06of like kind of pretty real kind of
06:10um both I think the systems and the
06:13engineering problem is massive
06:16um is as we're deploying these
06:17Technologies out there
06:19um and we're trying to scale them we're
06:23trying to make them more efficient we're
06:25trying to make them easily accessible so
06:28you don't need to have you know to know
06:30the intricacies of ml in order to use
06:33them and actually you can see sort of
06:36the contrast between making these models
06:39available through an API and making the
06:42technology available through child GPT
06:45it's fundamentally the same technology
06:48maybe with with a small difference with
06:50reinforcement learning with human
06:52feedback for chat GPT but it's
06:54fundamentally the same technology and
06:56the reaction and the ability to to grab
07:00people's imagination and to get them to
07:04just use the technology every day is
07:06totally different I also think the API
07:09for example PPT is such an interesting
07:10thing so it's my program against these
07:12models myself for fun right and it
07:14always feels like whenever I'm using one
07:16of these models in a program I'm like
07:18I'm wrapping a super computer with an
07:19abacus it's like the code itself just
07:22seems so kind of flimsy compared to the
07:25model that it's wrapping sometimes I'm
07:27like listen I'm just going to give the
07:29like a keyboard and a mouse and like and
07:32let it do the programming and then
07:33actually the API is going to be English
07:37and I'll just tell it what to do and
07:38it'll do all the programming and I'm
07:40just kind of curious as you designed
07:42stuff like chat GPT do you view that
07:44over time the actual interface is going
07:45to be like the like natural languages or
07:47do you think that there's still a big
07:48role for programs the programming is
07:51becoming less abstract where we can
07:53actually talk to computers in high
07:56bandwidth in natural language but
07:58another Vector is one where we're using
08:01the technology and the technology is
08:04helping us understand how to actually
08:06collaborate how to collaborate with it
08:09versus program it and I think there is
08:13definitely the layer of programming
08:15becoming easier more accessible because
08:17you can program things in natural
08:18language but then there is also this
08:20other side which we've seen with child
08:22GPT that you can actually collaborate
08:26with the model as if it was a companion
08:28a partner Co worker you know that's the
08:31interesting thing like it'll be very
08:32interesting to see what happens over
08:33time like you've made decision to have
08:35an API and whatever but like you don't
08:36like have an API to a co-worker right
08:38like you talk to a co-worker so it could
08:40be the case that like over time these
08:42things evolve into like you just speak
08:43natural languages or do you think it
08:44will always be a component of a finite
08:46State machine a traditional computer
08:47that's it yeah I think this is right now
08:50an inflection point where we're sort of
08:52you know redefining how we interact with
08:55with digital information and it's it's
08:59through you know the form of this AI
09:02systems that we collaborate with and uh
09:06maybe we have several of them and maybe
09:07they all have different competences and
09:11maybe we have a general one that kind of
09:13follows us around everywhere knows
09:15everything about uh you know my context
09:18what I've been up to today what my goals
09:23um sort of in life at work and kind of
09:26guides me through and coaches me and so
09:28on and you know you can imagine that
09:31being super super powerful so I think it
09:34is we are right now at this inflection
09:36point of redefining what this looks like
09:43we don't know exactly what the future
09:47we are trying to make these tools
09:50available and the technology available
09:52to a lot of other people so they can
09:55experiment and we can see what happens
09:57it's a strategy that we we've been using
10:00from the beginning and also with child
10:02GPT where you know the week before we
10:04were worried that it wasn't good enough
10:07and we also what happened you know we
10:11put it out there and then people told us
10:13it is good enough to discover new use
10:15cases and you see all this emergent use
10:18cases that I know you've written about
10:21um and that's what happens when you make
10:23this stuff accessible and easy to use
10:26and put it in the hands of everyone so
10:28this leads to my my next question which
10:32um so you invent cold fusion and then
10:34part of it you're like okay listen I'll
10:36just give people like electrical outlets
10:37and they'll use the energy but like I
10:39think when it comes to AI people don't
10:41really know how to think about it yet
10:42and so like there has to be some
10:43guidance like you have to make yeah some
10:45choices and so you know you're an
10:47opening you know you have to decide what
10:50to work on next well if you could like
10:51walk through that decision process like
10:54how do you decide like what to work on
10:56or what to focus on or what to release
10:57or how to position it if you consider
10:59how Chad GPT was born it was not born as
11:03a product that we wanted to put out
11:05there in fact the real roots of it go
11:08back to like more than five years ago
11:11when we were thinking about how do you
11:15how do you make this safe AI systems
11:19um and you know you don't necessarily
11:21want humans to actually write the the
11:26goal functions because you don't want to
11:28use proxies for complex call functions
11:30or you don't want to get it wrong it
11:33could be very dangerous this is where
11:35reinforcement learning with human
11:37feedback was was developed where you
11:41but what we were trying to really
11:43achieve was to align the the AI system
11:46to human values and get it to receive
11:50human feedback and based on that human
11:52feedback it would be more likely to do
11:54the right thing less likely to do the
11:57thing that you don't want it to do
11:59and uh you know after we developed gpt3
12:02and we put it out there in the API this
12:05was the first time that we actually had
12:07Safety Research become practical into
12:12the real world and this happened through
12:14instruction following models so we use
12:21prompts from customers using the API and
12:26then we had contractors
12:29generate feedback for the model to learn
12:32from and we fine-tuned the model
12:35on on this data and build the
12:37instruction following models there were
12:40much more likely to to follow the intent
12:44of the user and to do the thing that you
12:47actually wanted to do and so this was
12:49very powerful because AI safety was not
12:52just this theoretical concept that you
12:55sit around and you talk about but it it
12:57actually became you know was sort of
13:00going into AI Safety Systems now like
13:03how do you integrate this into into the
13:08and obviously with large language models
13:12we see great representation of Concepts
13:17ideas of the real world but on the
13:20output front there are a lot of issues
13:24um and one of the biggest ones is
13:26obviously hallucinations
13:31we we had been studying the issue of
13:34hallucinations truthfulness
13:37um how do you get these models to
13:39express uncertainty the precursor to
13:42child GPT was actually another
13:44project that we called Web GPT
13:47and it used retrieval to be able to get
13:51information and cite sources and so this
13:54project then eventually turned into
13:56child gbt because we thought the
13:59dialogue was really special because it
14:01allows you to sort of you know ask
14:03questions to correct the other person to
14:05express uncertainty there's just
14:07something down the error because you're
14:08interacting exactly there is this
14:11interaction and you can get to a deeper
14:15um and and so anyway we we started going
14:18down this path and at the time we were
14:20doing this with gpt3 and ngpts 3.5
14:24um and and we were very excited about
14:26this from a safety perspective
14:29um but you know one thing that
14:32people forget is that actually at this
14:35time we had already trained gbd4 and so
14:39internally at open AI we were very
14:41excited about gbt4 and sort of put
14:45chagibit in the rear view mirror
14:48and uh then you know we kind of realized
14:52okay we're gonna take six months
14:55to focus on alignment and safety of gbt4
14:59and we started thinking about things
15:01that we could do and uh one of one of
15:05the main things was actually to put
15:07charge ubt in the hands of researchers
15:10out there that could give us feedback
15:12since we had this dialogue modality and
15:15so this was the original intent to
15:19get feedback from researchers and use it
15:22to make contributive form or aligned and
15:24safer and more robust more reliable I
15:27mean just for clarity when you say align
15:29and safety do you actually
15:31do you include in that like correct and
15:33does what it wants or do you mean
15:35safety like actual like protecting from
15:38some sort of harm by alignment I
15:41generally mean that it aligns with the
15:43user's intents so it does exactly the
15:46thing that you wanted to do
15:48but safety includes other things as well
15:51like misuse where the user is you know
15:55intentionally trying to use the model to
15:58generate to create harmful outputs
16:03so yeah you can we were trying in this
16:06in this case with charge GPT we were
16:09um trying to make the model more likely
16:12to do the thing that you wanted to do to
16:14make it more lines and uh we also wanted
16:18to figure out the issue of
16:21hallucinations which is obviously an
16:23extremely hard problem but I do think
16:25that with this method of
16:27to reinforcement learning with human
16:28feedback maybe that is all we need if we
16:31push this hard dinner so there's no
16:33grand plan it was literally like what do
16:34we need to do to like get to AGI and
16:36it's just one step after right yes and
16:38it's you know all the little decisions
16:40that you make along the way but
16:43maybe what made it more likely to happen
16:46is the fact that we did make a strategic
16:49decision a couple of years ago to pursue
16:52products yeah and we did this because we
16:56thought it was actually crucial to
16:58figure out how to deploy these models in
17:00the real world and it would not be
17:02possible to just you know sit in a lab
17:04and develop this thing in a vacuum
17:07without feedback from users from The
17:10Real World so there was there was a
17:12hypothesis and and I think that that
17:14helped us along the way make some of
17:16these decisions build the underlying
17:18infrastructure so that we could actually
17:20eventually deploy things like
17:23I would love if you would Riff on
17:25scaling laws I think this is the big
17:27question that everybody has like I mean
17:28like the pace of progress has been
17:30phenomenal and you would love to think
17:32that the the graph always does this but
17:34like the history of AI seems to be that
17:36like you hit diminishing returns at some
17:38point and it's not parametric it kind of
17:40like tapers off and so
17:42from your standpoint was probably like
17:44the most informed vantage point in the
17:46entire industry do you think the scaling
17:47laws are going to hold and we're going
17:48to continue to see advancements or do
17:49you think we're hitting diminishing
17:54any evidence that we will not get much
17:59better much more capable models as we
18:02continue to scale them across the access
18:04of data and compute whether that takes
18:07you all the way to AGI or not that's a
18:10different question there are probably
18:11some other breakthroughs and
18:13advancements needed along the way but I
18:16think there's still a long way to go
18:19in in the scaling laws and to really
18:22gather a lot of benefits from from these
18:25larger models how do you define AGI
18:29um in our chart Opening Our Charter we
18:31we Define it as a computer system
18:34basically that is able to perform
18:37autonomously the majority of
18:39intellectual work okay I am was that I
18:44was at a lunch and Robert nishihara from
18:46any scale was there and um and he asked
18:49what I called it Robert nishihara
18:50question which I thought was actually
18:51very good characterization he said okay
18:53so like you've got to continue in
18:54between like say a computer and Einstein
18:56so you can go from a computer to a cat
18:58you know from a cat to an average human
19:00and you go from an average human to
19:01Einstein then they ask a question of
19:03okay so where are we on the continue
19:05what problem have we solved and the
19:06consensus was we know how to go from a
19:08cat to an average human like we don't
19:10know how to go from like a computer to a
19:12cat because like that's you know that's
19:14the general perception problem or very
19:15close but we're not quite there yet and
19:17then we don't really know how to do the
19:19Einstein which is kind of set to set
19:21reasoning with fine tuning you can get a
19:25um but in general I think we're sort of
19:27the most tasks kind of like in turn
19:30level I would say that's what I I
19:32generally say the issue is reliability
19:34right of course you know you can't fully
19:37rely on the system to do the thing that
19:40you wanted to do all the time and you
19:42know how do you increase that
19:44reliability over time and then how do
19:46you obviously expand to the the
19:51um the new the emergent capabilities the
19:53new things that these models can do I
19:55think though that it's important to pay
19:57attention to these emerging capabilities
20:00even if if they're highly unreliable and
20:04especially for people that you know are
20:06building companies today you really want
20:08to think about okay what what's somewhat
20:14um what do you see glimpses of today
20:17because you know very quickly this could
20:22these models could become reliable so
20:24I'd love I've been asking just a second
20:26to prognosticate on what that looks like
20:28but before very selfishly I've got uh
20:31on on how you think the economics of
20:34this are going to pencil out which is
20:36I'll tell you what it reminds me of it
20:37reminds me very much of the Silicon
20:39so I remember in the 90s when you buy a
20:41computer there are all these weird
20:42co-processors there's like here's like
20:44string matching here's a floating Point
20:46here's crypto and like all of them got
20:48consumed into basically the the CPU it
20:51just turns out generality was very
20:52powerful and that created a certain type
20:54of economy one where like you had you
20:56know Intel and AMD and like you know it
21:00and of course because a lot of money to
21:02build these chips and so like you can
21:04imagine two Futures there's one future
21:06where like you know generality is so
21:08powerful that over time the large models
21:11basically consume all functionality
21:13and then there's another future where
21:15there's going to be a whole bunch of
21:16models and like things fragment and you
21:19know different points of the design
21:20space do you have a sense of like
21:23is it open Ai and nobody or is it
21:26everybody it kind of depends what you're
21:28trying to do so obviously the trajectory
21:30is one where these AI systems will be
21:33doing will be doing more and more of the
21:35work that we're doing and they'll be
21:37able to operate autonomously but we will
21:40need to provide Direction and guidance
21:42and oversee but I don't want to do a lot
21:44of the repetitive work that I have to do
21:46every day I want to focus on other
21:47things and maybe we don't have to work
21:5210 12 hours a day and maybe we can work
21:55less and Achieve even higher
21:57outputs and so that's sort of what I'm
22:01hoping for but in terms of like how this
22:03how this works out with with the
22:07you can see even today you know we make
22:10a lot of models available through our
22:12API and from the various from the very
22:15small models to to our Frontier models
22:17and people don't always need to use the
22:21most powerful the most capable model
22:23sometimes they just made the model that
22:26actually fits for their specific use
22:28case and it's far more economical so I
22:31think there's going to be a range
22:34um but yeah in terms of how we're
22:36Imagining the platform play
22:38um we definitely want people to build on
22:43top of our models and we want to give
22:45them tools so to make that easy and give
22:48them more and more access and control so
22:50you know you can bring your data you can
22:52customize these models and you can
22:55really focus on the layer beyond the
22:58model and defining the products which is
23:01actually really really hard there is a
23:03lot of Focus right now on building more
23:04models but you know building good
23:07products on top of these models is
23:09incredibly difficult okay we only have a
23:12couple more minutes sadly I would love
23:14for you to prognosticate a little bit
23:15unlike where you think this is all going
23:17like yeah like three years or five years
23:20I think that um the the foundation
23:23models today obviously have this great
23:27representation of the world in text and
23:30we're adding other modalities like
23:32images and video and various other
23:34things so these models can get a more
23:38comprehensive sense of the world around
23:41us similar to how we understand and
23:44observe the world the world is not just
23:46in text it's also in images so I think
23:49that will certainly expand in in that
23:53direction and we'll have these bigger
23:54models that will have all these
23:57um and that's kind of the pre-training
23:59part of the work where we really want to
24:04these pre-trained models that understand
24:06the world like we do and then there is
24:09the output part of the model where we
24:12introduced reinforcement learning with
24:13human feedback and we want the model to
24:16do the actually the thing that we ask it
24:19to do and we want that to be reliable
24:21and there is a ton of work that needs to
24:25happen here and maybe introducing
24:26browsing so you can get fresh
24:29information and you can cite information
24:32and solve hallucinations I don't think
24:35that's impossible I think that's
24:36achievable on the product side I think
24:39we want to put this all together in this
24:43collection of agents that people
24:46collaborate with and you know really
24:48provide a platform where people can
24:50build on top of and you know if you
24:53extrapolate really far out these models
24:56are going to be incredibly incredibly
24:58powerful and with that obviously comes
25:00fear of them being misaligned having
25:02this very powerful model that are
25:05misaligned with our intentions so then a
25:08huge challenge becomes the the challenge
25:11of of super alignment which is a
25:14difficult technical Challenge and we've
25:17we've assembled an entire team at open
25:20to just focus on on this problem so very
25:23very very last question are you a Doomer
25:25an accelerationist or something else
25:28let me say something else all right
25:30perfect thank you so much fantastic