Fraser Kelton (GP at Spark Capital, former Head of Product at OpenAI)

0:00 I remember honestly like coming out of my room with a a printed paper and waving and being like Kristen you you can't believe you can't believe what happened like they they found a way they found a way to turn all of the internet into training data I couldn't shake I couldn't shake from my my my head from my soul from my bones honestly what was going to happen if you played with gpt2 it was horrible right it was it was

0:30 a gimmick it was novelty I scraped a decade of Fred Wilson your partner's blog post and I I fine-tuned a version of gpt2 and he was super dismissive he's like I don't think I'm out of a job anytime soon but if you appreciated what had happened to enable that and what that meant for a year from then two years from then 3 years from then you knew it was going to get really really wild and wonderful

1:19 welcome to the slow hunch podcast I'm Nick Rossman from Union Square Ventures the slow hunch is an idea from one of my favorite books where good ideas come from by Stephen Johnson the idea is that big breakthroughs Innovations in science and technology society and elsewhere are often not the result of a single aha moment but rather the slow accumulation of happenings and observations and intuitions that build over time with

1:49 luck it's always been one of my favorite ideas I think it's encouraging and empowering that if we just keep following our noses lots of small and seemingly Rand steps May ultimately lead to Big breakthroughs today I'm joined by my old friend Fraser Kelton over a decade ago I served on the board of Fraser's last company Coco which was an AI company focused on mental health and trust and safety in online systems which was ultimately acquired by Airbnb

2:19 Fraser then went on to become head of product at open aai where he was part of the launch of chat GPT which you could think of as the iPhone moment for AI a massive break through moment today we'll have a chance to unpack both the story of that breakthrough and Fraser's personal path as part of it Fraser it's so great to have you here I'm so happy to see you and thanks so much for joining us today right I'm

2:43 Fraser Kelton I'm a GP at spark Capital I've been here for a year and before that I've been a founder and operator in startups forever before this immediately before this I was had a product at open AI helped transition it from a research lab to a group that shipped a lot of product wrote the first couple of product road maps and helped build the product development org and before that was fortunate enough to have you invest

3:13 in my startup and before that had been involved in a bunch of young technology companies as it goes I'm Canadian so we're going to have to tolerate that I love your subtle Canadian accent phraser I always will yeah the the nice thing is that Yahoo is not as prominent as they once were which was like the dead giveaway but that's it like I had the joy of meeting you what in 2014 I think right I think that's when we may be met

3:44 I remember here's how I remember hearing about you for the first time you were the Chief Operating Officer of a company called get glue which you can talk about that was a TV a t interactive TV kind of second screen company back in I don't know 2011 12 13 is that right when you were there and that company ended up getting acquired but I think most people probably don't know about

4:15 Coco and that's when I really got to know you and I remember sitting with your co-founder Rob at MIT I think when he was still a student there hearing the vision for Coco that that he had coming out of his grad research and and you're interested in it and maybe you could just talk a little bit about that because I think that's a story that or a product that you know is a little less known in your

4:44 in your story my the other co-founder Kem and I had the Insight that modern therapy cognitive behavior therapy could be delivered at online at scale and without a therapist we set off to to try to understand what it would take to build a product in that area and I I cold emailed this guy who was I thought a tenur profit MIT on a Friday night because I came across one of his papers and it turned out that he had

5:08 just the week before defended his dissertation where he had built a prototype of of like what what Kareem and I were testing in WhatsApp groups and long story short a couple of months later he he joined us as a co-founder and we worked to commercialize what was his dissertation he had ran a randomized control trial showing that it like it it worked so this is a really interesting story right because you and Kareem and then also Rob obviously was coming at it

5:35 from a research perspective but how did it occur to you at that point that such an approach might be possible they had been selling like workbooks on Amazon for cognitive behavior therapy where you could buy them for $20 and if you if you looked at the research those workbooks worked and and then there was a very very small set of researchers one group in Australia one group in Chicago

6:06 who were showing that online interventions actually worked but if you if the the challenge if you read them was that after the actual study people didn't continue to use them and then you would sign up and you would try it and it it was you know no no shade you know if an academic was to build a product it's what it looked like I discovered at that moment that if you actually read a researcher's paper and

6:31 then email them they'll give you all the time in the world right and so I was getting up at I don't know 4 in the morning to talk to this group in Australia and they said oh we haven't published this new paper and it's the most amazing thing we're now understanding that I don't know if this is gonna blow your mind people don't want to read full PDFs online like no you're not kidding huh and they were

6:55 they were rediscovering the basics of HCI right and good product design and you just realize that if you if you took the the basics of of Designing developing crafting a good product experience and applied therapeutically effective interventions at least our hypothesis was that there would be something that would work the key idea behind Coco was not just that CBT

7:26 delivered in an online format could work but that it could could be scaled in a peer-to-peer fashion right so that you know I'm studying cognitive behavior therapy for my own well-being maybe it's starting to work for me you know if there's a way where I could then become your coach you know by coaching you I get better myself right and then as you get better you could then Coach someone else and then you get better and they get better and

7:51 then it can multiply to Infinity that's just an amazing concept for like any kind of potentially consumer viral education you know product or platform and I'm a little bit surprised that that hasn't taken over the world but I think there may be reasons why but in this particular context where you know I think we were back in 2014 sort of just at the beginning of this idea that there could be a mental health crisis bubbling up on the Internet

8:21 and the idea that maybe it would be possible to use the power of the internet to fix some of the the problems on the internet it was such a was just such a compelling idea so we we started iterating on a consumer version of that app and we re we realized pretty quickly that it's not going to work it was hard to build a destination website or or app where people are seeking this type of thing out and I remember a conversation

8:50 with you and Brad and and Kareem and Rob where we're you know we have a well-defined problem in front of us and I don't know who it was that said it but they said Hey listen you guys are having no no problem recruiting individuals who are turning to the large social networks when they're in need of help what would it mean to take your your your product there we did what we called the reverse touring test where we we

9:15 we copy and pasted a script into a messaging experience and had people interact with it to see if we could pull our product experience into like a messenger a again like within a couple of days we realized that that was going to work we could easily Port our product to platforms like Facebook messenger and others and we reshaped it to basically be an embedded experience into

9:47 these social networks and we we set out to do like Partnerships and Business Development with them and this is worth just mentioning because we got a lot of interest and and a group called Kick a messaging app as well as a group called Pinterest said you know let's let's let's do it and up until that point because of the the nature of the product we had had to moderate 100% of the

10:17 content that got created before it went to other users but because the volume was so low Rob and I could do that in slack ourselves and we we start the ramp up with these groups and within like five minutes we're like nope turn it off we can't handle this volume like just stop because the human interaction of moderating the content wasn't going to work and so we set out to figure out what would be a viable solution and this

10:47 was now 2015 2016 and there's just like a a great moment of serendipity where deep learning for natural language processing word to V these other types of Technologies early AI Technologies are just becoming viable and we started to invest into that direction and you remember that I mean I I I remember it like it was yesterday and I remember a feeling back then that we were on the

11:20 Forefront or the frontier of like a new way of using computers to figure stuff out right and I was just looking back through some of the old Coco deex this morning and you know after the after the transition from the consumer app to this idea of like let's embed this you know into social media where there are already millions and millions of people expressing you know their their their situations and and what if you could run

11:50 a classifier across all these social media posts to try and you know distinguish you know this type of in in maybe inappropriate cont content from this type of potentially harassing content and and it it took us took me down this Rabbit Hole of God the the level of nuance in language is infinite and then also the tools that we're starting to have at that moment 20 2016 2017 or so are like pretty good at

12:22 susing some of these things out not like fully good and I remember you had some charts in some of the old decks of you know the Coco class ifier can do you know 98% you know accuracy versus you know the old one which was you know 86% it's like a real lift but still like like hard yeah that that's right and it was a like a grind to get there both in terms of figuring out how to

12:49 get exceptionally good labeled data of enough diversity for that one specific task we we ended having to build a lot of our own you know tooling and services to allow us to iterate and improve these models for our our own purpose and it was really interesting because because we had done human moderation ourselves we had a really deep appreciation for the the challenge of the task and I remember you

13:21 know I think our first prototype was using of all things IBM's API for Watson's NLP remember that yep and it was eye openening because you just you you it felt weird in the sense that this algorithm was catching the Nuance that you and I as a human would would look at and and you know and catch and you're like how how is it even doing that you know double double negations and things like that where it's able to

13:49 figure out the intent in a sentence we just continued to invest in that for our own product purpose and then there was this other board meeting where we realized you know our our Crisis Support tool is scaling it's working you know we help two million people we we published an awful lot of research but we had no idea how to turn it into a a business and another one of those like happy discussions where we had a well-

14:16 defined problem and and it comes up well listen maybe our maybe our our back end our our tooling to create these classifiers is the commercially interesting product and so we commercial ized a a set of apis that allowed large internet companies to send us content that we would train classifiers and then deploy and serve in production for them I also think that there were Shadows of what still exists across

14:48 social media where I'm not I wish that these companies took certain types of concerns more seriously and that that was really the first time that we bumped up against that and it still is very troubling to me is maybe where I just want to leave it even when you have the tools at hand to help assist with really important things like mental health and social media there's not always a perfect alignment in terms of how to get

15:18 that deployed and and I think where we ended up with Coco was sort of trying to sell into the trust and safety organizations and I remember a distinct feeling that it was hard it was hard to make that sale because as as serious an issue as this is it wasn't always the the top driver of you know business no that's right and you know this was previous this was before Zuck gets called in front of Congress for Content

15:49 moderation issues and so it's you know I I remember us talking about it in the sense that this is a group of individuals in trust and safety who care deeply about the problem but they they're not well positioned within the org to affect change and I think our our Mantra became that they could change content not code and it became very hard for them to do anything that required you know engineering and so that was just then a new properly

16:20 defined problem for us to try to resolve we we did you know we did our best to get around that and make it easy for these Champions to to integrate it and then you know and then the ultimately the next challenge was it's just hard it's hard to build a business in that area for that exact reason because a it's a little it was hard to get people to pay and B even if they did pay

16:42 they were limited in what they could do to integrate you know on the back end and I guess maybe to move the story forward where where that ended right was that well you tell me where that ended we commercialize our backend we sell it as an API and we start actually getting some traction Airbnb is a customer Tumblr like we we had a number of people using the

17:12 API this is now when Zuck gets called in front of Congress and it's a a fortunate time to be selling an API content moderation AI solution and Airbnb says hey you're a small team you haven't raised a lot of capital do you want to be acquired and you you remember what that was like that was a very hard set of decisions for us to wrestle with but ultimately we we elected to to go to Airbnb we moved from as I say Brooklyn

17:40 to the Bay Area the importance of the sort of the societal importance of content moderation and sort of social media content was still sort of swelling right it was like yeah like right around or right maybe right before that inflection point and then but I think it wasn't it wasn't big enough to to have a business a big big business around or at least in the format that we were building it in but it also left you

18:08 sort of sitting at Airbnb you know I don't know maybe with some time on your hands so I feel I feel so this this part of the story is so much easier now that Brian chesky has been quite public that his company was a mess and he's trying to clean up his company because we we go there and I am shocked this is also the first time I you know we were eight people at Coco and I think at Airbnb at

18:33 the time was 8,000 right and so I was shocked with how much of a mess it was as you said it was a wonderful time to have the experiences that I had just had over the past five years iterating and building on the on the edge of AI mhm to then have a lot of time on your hands right because think about it is the

19:04 the Transformer paper gets released and and then very early versions of the models that are based on it become available open source what year was the Transformer paper I think it was 2017 and then the the the models land and start to land in like 2018 in an open source way and then gpt2 comes out and is broadly available with open source or open weights in 2019 so

19:34 there's a couple of years there yeah Coco is acquired by Airbnb in 2018 I think right and you're you're there for two years give or take 2018 to 20 20 I tried my best I I didn't make it two years okay I was trying to be generous I tried my best Kareem and Rob the the co-founders made it made it past that Mark but I I I couldn't but kind of an amazing time to be there I guess

19:57 right it's an amazing time to have the experience that we have that we have had right so we we it's hard we we we glossed over it but you know we fast forward through years of us going from using an early version of Watson's NLP API to building our own ml engineering infrastructure and hiring you know a guy named Rohan and others who are just like exceptionally talented who are training these models deploying these models we're like figuring out how to

20:29 improve them iteratively and I think our pitch at the end was like Hey listen you give us 50,000 pieces of labeled data and in a month or two weeks and we're going to train you a great classifier that's how quickly we can get it down right and I'm in the bowels of Airbnb I have nothing but time on my hands and I take Bert and I I go into a a Google collab notebook and in 20 minutes

21:02 I'm I'm training a Model fine-tuning A version of Bert that's outperforming the ml engineering teams that are sitting around me it's so mundane except for the fact that I had just experienced what it took to train these models and I remember I remember it was a twoe stretch man it was a twoe stretch where I'm training these models I'm I'm I'm dunking on Rohan I'm like Rohan come and look what I just did man I'm going to

21:28 put you out of a job like look at look look at look at this confusion Matrix like it just outperformed you and I have no right doing this and I did it in 20 minutes and I I then spent my evenings you know I get my kids to bet I had three young daughters I'd get my kids to bed and I'd spend my evenings trying to understand the research that was behind all of this I was

21:52 reading you know blogs like towards data science and trying to make sense about masked training like a lot of the different techniques that were all bubbling up at that time and it was it was honestly electrifying it was probably one of the more electrifying moments in my life from a career perspective to appreciate the Elegance of what what they had just done and and more importantly what that meant was about to happen I remember I

22:23 remember honestly like coming out of my room with a a printed paper and waving and being like ch you you can't believe you can't believe what happened like they they found a way they found a way to turn all of the internet into training data and she you know she she like patiently sat through as I as I enthusiastically told her what was happening and I I couldn't shake I couldn't shake from my my my head from my soul from my

22:55 bones honestly what was going to happen because if you played with gpt2 it was horrible right it was it was a gimmick it was novelty I don't know if you remember I I scraped a decade of Fred Wilson your partner's blog post and I I fine- tuned to version of gpt2 and he was super dismissive he's like I don't think I'm out of a job anytime soon and if you read the post that it generated it was

23:24 horrible but if you appreciated what had happened to enable that and what that meant for a year from then two years from then three years from then you knew it was going to get really really wild and wonderful gosh it it it reminds me a lot of people who lost their minds when they read the Bitcoin white paper and I you know I don't know if you're you're one of those people you know doesn't matter but but I

23:51 think I hear stories from people about oh my I read the Bitcoin white paper and I just saw the future you know it all unfolded because if this is possible then the next 20 years are possible there's a a Prof called Rich Sutton who who has a post called The Bitter lesson and it basically lays out an argument in 2019 early 2019 that all

24:22 of the traditional approaches to like finicky machine learning or AI techniques don't matter all you need is scale you just you just need to scale up data and compute and it was a beautiful articulation of what this set of of breakthroughs if you will had had just enabled because you could now basically turn well you and I remember right like we talked about how hard it was to get 30,000 pieces of of labeled data that

24:55 would you know how do you scale how could you possibly get 30,000 pieces of labeled data let's figure out how to do this and all of a sudden they had an algorithm and a method to turn the Corpus of the internet's text into training data right so you go from 30,000 examples to everything and and you you just you but you realize you realize that gpt2 and Bert were trained on a small sliver of

25:25 what was now going to become possible and we were at the the first the first step of a of a 20-year run that was going to be like Bountiful in a way that people couldn't appreciate wow it it reminds me a little bit of you know yah Yahoo and Leos to Google you know and the idea of of pay drank and using using the latent information in the internet maybe we come up with one of these

25:47 breakthroughs every 20 years who knows but but that's a real it's a real it's a wild thing and to see that so all of a sudden sort of in you know I you know 20 2019 or so this becomes a building block I'm curious to understand how you move from Airbnb into open AI I'd love to talk about like the there was this massive unlock between having gpts and having people get excited about them which happened you know while you

26:18 were there so I had that two we stretch and I realized that the world's going to change on a trajectory that I just wanted to go be a part of and I think it's worth highlighting that I think the research world was quite open that the capabilities were going to improve dramatically what excited me was what I was doing parallel and in lock step with the capability Improvement was a similar increase in accessibility and I was able to do

26:48 things as I said that Rohan and team required like exceptional skill and expertise to do and I'm like I have no right doing this and the world's going to go in this way and what happens when profound Technologies become dramatically more accessible to to anybody everybody creative people product Builders it's like the history is it's amazing I was going to go do exactly what I did with Coco I was going to go start a company around this

27:12 insight and I I was emailing researchers because I had known if you email researchers with real interest they give you all the time in the world I was going through old emails I cold emailed Ilia from open AI and I sent him the the bitter lesson blog post I'm like I don't know if you've come across this but like here's why I really believe that this is going to happen like how naive in hindsight right like and he goes yeah he goes you know

27:41 I'm very familiar with this and we had like we had an exchange and I ended up meeting him and Greg for breakfast and in late 2019 I got a demo from them of of an early version of what became gpt3 and at that point they had already realized that they wanted to commercialize and they wanted to bring it to Market as an API and there was

28:13 just like a great overlap in in in like Vision right like I I my excitement was what happens when it becomes dramatically more accessible they thought that the API was a great way to make it dramatically more accessible I looked at I looked at gpt3 and like there's no way I'm doing this and I want to go be a part of that parade and so I joined there in 2020 so much to unpack here but but the part I love

28:38 about about how you are and and I just I love hearing it in people's stories is is sort of not being afraid to just go go cold contact somebody that that's doing something interesting and it's not just researchers it's other people too I mean maybe researchers are particularly accessible but the idea that it's possible in the world to to be reading stuff and just reach out to someone and they might just write write you back and

29:03 then that might lead to something I mean it's just so beautiful 100% I think it's pretty unique for Tech and research and probably a number of other fields like I'm not sure that it's entirely common but I think it's I think it is beautiful and then I think it's important for everybody to like pay it forward if you will th% and that's that's such a great reminder and a great point and I think if if you think about things in

29:29 kind of Cycles right people like that are tend to be very open and accessible really early in a technology cycle when things seem early and not there's no Spotlight and everybody's just curiously chewing on ideas that's that's the best time to find access to people because they're all just right there later it gets harder I think the relationship between gpt3 as an API and Chad gpt3 as a product is really interesting we come across this all the time in our

29:59 work which is when is the right time to vert to go vertical and productize something when does something need a real consumer interface versus when can you go horizontal we do a lot of investing in crypto and web 3 this this question comes up all the time you know launch protocol that's horizontal which is like an API versus launch an app on top or do both it does feel like there's a long history of both being the

30:26 right answer you know it's kind of like you know HTTP and HTML and you know the Netscape browser right or the gpt3 and chat GPT or and on and on and on was as you kind of got into the role at open open AI like what was the conversation like around should this be an AI vers an API horizontal versus should this be a product was that controversial was that did that feel like an unlock when you finally got

30:57 there they had already made that decision for the first time before I already got there and I think the story is that John and Greg had hacked together an API on an early checkpoint of gpt3 and they did that you know over the the winter break one year and people were then like hacking and building on top of it and I think the internal discussion was hey here's here's eight really promising product

31:28 ideas which one should we do and then it was well why do we have to pick from any of these eight why don't we just Empower anybody to build whatever they want because the likelihood of the the actual interesting products being within this you know finite set of eight from our small number of Minds is you know you know how it is and so they they were already committed in that direction you know being there when when seeing everything

31:57 unfold right as you were from the research from the products you know all of a sudden it must have felt like a very expansive you know universe and then to see chat gbt actually launch it's got to be one of the most amazing things ever in history but I it's also I'm curious to hear how that all felt especially in the context of you know for the prior 10 years having been kind of grinding areas

32:28 where it felt like you were working on something really important it was pushing uphill versus pushing downhill did it feel like you were pushing downhill finally I think there's an interesting parallel as to why it caught us off guard that's relevant to your interest in having these discussions and so the organization has been very upfront that it was an accident right it was supposed to be a research release we spent two weeks to push it out before

32:54 we got back to you know real work if you try to can you know think about and sit with why it caught us off guard it's because we had been sitting on the edge of iteration after iteration after iteration as this technology was was going through the exponential that we you know everybody there had had believed we were on but if you're sitting on the exponential and you look a little bit forward and a little bit back it doesn't feel that

33:26 profound right and so the first version of gpt3 in the API gets a little bit better we introduce instruction fine tuning it gets a lot more controllable it gets a little bit better it gets a little bit better it gets a little bit better and then all of a sudden what has felt like a grind for us for you know some number of years and and an expected trajectory people get their first taste of what modern AI actually looks like

33:55 and feels like and does and they they were going they were going from no experience no insight to having to appreciate it in that moment like having it snap into focus and we had you know had been brought along month after month with like grinding Improvement after grinding Improvement I mean like a insane moment for the world because to your point like the research Community working on these things you know saw this coming years ago right me decades ago right U because

34:27 a lot of this is really Decades of research you know piling up and sometimes you never know how long it's going to take but for the for the average person you know the expressions of this technology were so hidden for so long right so every tech company had ml in some way under the hood to make things a little bit smarter a little bit better but you could couldn't really taste it and then all of a sudden you

34:53 know I think just generative text was a at the Quality that that that if gpt3 finally got to a quality bar and then the expression and the CH where it it felt like something completely brand new even though it was I don't know I don't know how many decades of research you know a creative research and the the clues the clues were all there we just it's not that we missed them it's just that we we

35:22 didn't appreciate that there could be a well how could you appreciate that there'd be a moment like that right there there's so many elements of luck and Zeitgeist and tapping into the cultural moment and all that other stuff was it the incremental quality of gpt3 or was it something about the the packaging you know of packaging it up into a chat format like do to which of those do you give more weight if if either I think everything I

35:55 I think a very simplistic UI that wasn't complicated that also met a chat model that was a little bit more intuitive for people to interact with that you know you can then trace it back that was built on top of this instruction fine-tuning effort and it's a bunch of intangible things that are that often don't get discussed but I think are are worth mentioning it's you know Nick

36:26 basically I I an internal guy at open AI the a very strong product guy went to the mat to not have a weit list and you know in some degree like put his career on the line and this isn't that this isn't that product if he doesn't do that they had hired shortly after me the VP of comms from Apple like the guy who had ran the communication strategy at the best company that's done it ever

36:56 joins this you know know weird ragtag bunch of Misfits and I think that a lot of it also traces back to the investment that steveen team build o over a handful of years right the storytelling the positioning the brand all of that contributes in that moment to create that type of outcome wow the no weight lless decision is an incredible decision let's just say if you rewound the clock and you you Did you sort of hide the product more behind a wait list

37:26 I think we're probably still where we are today but but maybe not right it's going to come out and people are going to build stuff but but the cultural moment that happened wouldn't have happened otherwise no way no way then you have to ask how much of that cultural moment is is then the acceleration of this entire moment right now right is it became dinner table conversation around the world everywhere how much of that then factored into all Fortune 500's CEOs

37:57 ripping up their road map and saying we need to have an AI strategy right how much of that was academics saying we need to get in front of this because this is going to change education I actually think that that's the biggest thing that's happened here is Society had a A Moment Like An inflection moment where what these researchers knew was going to happen at some point and globally we appreciated

38:28 it all at the same time which is so wild to to even consider that is incredible yeah cuz we like we had a wait list for do 2 which was launched earlier in that year and I'll tell you I think internally we all thought that that was the the real magic certainly gp4 right like the the model that came later but we thought you type in text and get an image was was it and what happened I you know you

38:57 know we saw a huge crush of people sign up to the wait list in the first three days and then it you know it it tapered off I suspect something similar would have happened here had it not been for for that that very bold bold individual well and and what might have happened maybe is that somebody else might have built something that looked like chat GPT on top of the API right if there had been a weight list on chat GPT

39:26 you know maybe maybe but that that's where I then come back to I I think that we're on a timeline now that's a reflection of that cultural moment happening instantly around the world and that is a result of things like Steve's and comm's team investing for years into a brand and positioning and everything else that helps catalyze that that like fire yeah and and and also another way of saying it is I think it takes a pretty sub stantial company to

39:59 stand in the spotlight the way that open AI has in the past couple years yeah cuz it's the it is like Global hyperfocused attention you know I mean the other thing was the amount of inference that was was getting served for the free product was you know nose bleeding Le High and there's there's no like that's a great observation that you had but I don't think it's like it couldn't

40:29 have happened because nobody building on top of the API at API price points would have been able to have tolerated that we had to take gpus from the research org into the inference side to be able to even like approach meeting demand and so there's therefore then a very small number of of Institutions that could have done that but even think about that that decision is if that decision gets made differently if if we decide as

41:00 an organization nah we're going to continue to train the next model let's not bring this over here and let's just add a weit list or what have you this isn't this isn't unfolding the way that it's unfolding man it it it really does feel like there's a little a world of you know before chat gbt and after chat gbt right just in terms of how the world is looking at technology and this particular set of Technology like really

41:25 everything has changed since then and to think that that you know might not have happened I I I think it's been incredibly important for the world to have had to focus on this because now people understand where the future's heading in a different way than than they might have you know now it's like really in people's full attention and it's going to require a fair bit of attention for people to figure out how to live with all this you know cuz it's

41:57 it's going to change so many things mhm Bill Gates has that line that we tend to overestimate in a year and underestimate in a decade I I I think I think we're living through that 's you know multiple orders of magnitude where I think that we are grossly over over our skis in terms of the impact in the short term and I think even the the most aggressive ambitious projections of where things are going to

42:27 go in 10 20 30 years is still mostly under appreciating the change that's going to happen I I just think it takes that long for this dramatic of a change to Ripple through society and it's probably best that we start having those discussions today agree so much of what has been part of your story here has been trying to understand the next card that's about to show up right in the research what are you reading today it

42:56 it is simultaneous L A mundane answer and and one that should be incredibly exciting given this entire conversation there are a very small handful of groups who are applying this same concept into biology they're taking the Transformer architecture they found a way to do you know mased learning where any sequence of proteins in the world becomes training data and they're able to train a model that generates entire L

43:27 new proteins that don't exist in nature for various purposes and tell me if this sounds familiar the models right now are about the size of gpt2 and if you if you look at what's coming out of them it it like it functionally works you can design A protein that doesn't exist in nature but feel but but has function that's similar to proteins that

43:58 have evolved over millions of years but the function isn't that interesting it's not that interesting it's it's not able to do stuff that we care about it's a toy it's a novelty and and we're at gpt2 levels and there's a clear path to go gpt3 GPT 3.5 GPT 4 into that world you know there's a group called profluent that I've invested in who has been one of the small groups

44:30 who have been leading the scaling hypothesis of the transformer into biology and they have now shown that they can have an entirely denovo so brand new protein designed wholly by AI that does crisper Gene editing in a human cell and so rather than writing an essay or a computer code like chat GPT can this same architecture the same model the same algorithm with different

45:00 data is writing human biology and we're at we're at GPT three levels for all of all of our history life sciences medicine has been Advanced based on what we discover in nature crisper was found in a denn yogurt facility because some scientists were curious as to what was happening penicillin was was an accident a happy accident because the researcher had a

45:32 an untidy lab bench and an accident occurred before he went away we have been constrained by what we could discover in nature and we are very quickly going to a future where we can design precisely to our needs right so you you you will rather than telling chat GPT to write me an essay that is five paragraphs long on you know the American Revolution you're going to say I want a I want a

46:03 protein that functions for these purposes and can behave like X Y and Z in the human cell and it's going to go and design that protein for your need this is such a huge topic that we we can't even scratch the surface of but of course like everything in generative AI such Futures can go in magical directions and terrifying directions and what's your best take on that challenge especially as you get into more profound

46:34 Technologies like like AI driven biology I think the the upside in the short term is grossly overestimated because change is just very hard for society but I think long term is going to be profound I I I it feels so silly to even describe like a future where you have have Precision medicine built for everybody tailored to their needs it feels so silly to say that we will

47:06 have software that gets written on demand and and like thrown away after it gets used but that that feels like graspable within a handful of years right but even but even that like I don't even know what the ramifications are if if we all start having disposable software that that gets written on demand like it's so it's so profound how things could go it's just hard it's hard to capture the the positive upside in a way

47:35 that doesn't sound ridiculous to me the idea of the slow hunch is a little bit about the idea but it's also really about the personal journey and you know everything that kind of grabs your attention and your interest along the way leads you to kind of where you are today and when you look backwards it it kind of all makes sense I'm curious for you you like as you look back what do you think is the the

48:02 through line or the consistent idea or the what is the thing you keep coming back to you know is it about ML and AI what's the idea that kind of you keep coming back to I came of age with the internet the thing that is common from then to now is an appreciation that strong technologists who are are wildly icious are able to do like beautiful things and I have spent my life trying to be

48:34 curious about what strong technologists are doing and then going to help them however I can I mean to go full circle you know as Brad said I was sweeping the floor and tidying things up at that moment and I am going to sweep the floor and tidy things up for the rest of my life adjacent to the strongest technologist that I can find because they're the ones who are curious and capable of creating like the future that

49:05 we all should Aspire for beautiful I don't know doesn't that like doesn't that feel app propo for everything that I've done like I'm not I'm not a technologist I haven't built these things I have gone and swept the floor for them who and helped them to do the stuff that they're uniquely capable of doing bring it back to Canada as one does you know it reminds me of you know curling right you got The Pusher you got the stone and then you

49:32 got the sweeper I don't know anything about curling but this the sweeper roll pretty important you know here here's a fun anecdote is when we were all in one office the researcher sat on the third floor and every day the daily Tartine order came to the third floor and they marched it past our floor up to the third floor and I loved it I loved it right because it was a reminder in many ways of why we

50:06 were there and I hope I hope to always remember which floor the Tartine order is going to H and to be the one who happily supports that that is beautiful well Frasier thank you so much this has been a real pleasure and a joy so nice to see you and to get to spend a few minutes you know exploring how you see the world so thank you so so so much for talking to me today thanks Nick