NEAR AI Office Hours #20 with Poolside - Advanced foundational AI model

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   don't stop us now together we me toine the game when near we'll find the flame empowered Minds rise and S use your own AI we explore hold the future hand in hand in this Tech driven hello hello hope everybody's having a great time uh we are joining again for the latest installment of our AI office hours I'm really excited to have Jason Warner here I'll let him introduce himself uh but uh we I'm sure gonna have

2. 0:51

   a very interesting conversation Jason uh introduce yourself and kind of tell us more what you're working on right now all right um thanks I thanks Alex I did not realize you had 20 episodes of these things this is pretty cool um everybody I'm Jason uh I'm one of the co-founders of poolside um poolside. on uh the internet which basically says nothing about us at the moment but give us a day

3. 1:13

   or two and it'll be out there um prior to poolside um I was the CTO at GitHub for four years so two years pre-acquisition by Microsoft and then two years post prior to that kind of the same thing at Heroku which is a platform as a service now owned by Salesforce and before that same thing at canonical people make a bunch of Linux so uh I like to say that uh it's a long-winded way of saying that I am very very

4. 1:38

   useless in a desert island scenario but if it has a keyboard attached to it I'm okay but overall you've been kind of hanging out around Developers for a long time that's pretty much the yeah I mean developers are fun Finance people bad not not my thing so so this will be a fun conversation here yeah so I I mean I don't know how much you want to share about poolside uh given given your background and context but it's probably going to be valuable for the conversation so for the conversation poolside is essentially a

5. 2:12

   continuation of conversations that I was having when I joined GitHub so when I joined GitHub in 2017 uh again two years before the Microsoft acquisition I wanted to turn GitHub from a code host into a software development platform and and infused with intelligence and like lawn you know drawn out plans and all that sort of stuff but the continuation of what pool side is is kind of like going back to

6. 2:35

   that original Vision to a degree which is that software development and neural networks and the marriage of those two things are going to be monumentally beneficial for Humanity and developers and uh we're pushing that boundary forward we're pushing all of those things over there so um you know GitHub co-pilot which was incubated in the office of the CTO while I was at GitHub is a slight Divergence from the original view that I had and my co-founder ISO who I met when I was at GitHub had at

7. 3:07

   the time and PS side is kind of going back to it and very specifically what that means is we're building a foundation model for from scratch for software development um where you have a couple of different proprietary techniques the one that we' become most known for is this thing called reinforcement learning via code execution feedback which we can get into here um but uh the idea being that we're going to advance uh software development

8. 3:30

   via these mechanisms very quickly over the next couple of years yeah so I guess it's relevant within the context obviously of what we've been doing and uh that's been part of the conversation we've been having for past uh a while but yeah I think we we all agree here that AI development is kind of the future yeah and there's going to be a probably pretty rapid transition from um kind of individuals writing code you know and kind of needing big teams to actually facilitate big projects to kind

9. 4:09

   of more and more uh periphery being built by Ai and then kind of architecture orchestration and uh kind of definition going in that direction uh we we see some of this examples already um obviously for us it's been interesting because of web 3 there just just not many developers and so like even uh and the kind of code Bas is in general pretty small so you can do a lot more things from scratch uh I know you guys are working and kind of aiming for

10. 4:44

    work with like larger code bases and really uh giving a leverage there so I'm curious how you see this trade-off of you know helping existing developers working being more productive or kind of leap frogging that and just giving non developers ability to build software from scratch kind of not even trying to uh operate on Legacy code yeah I mean I think like there's a couple of different ways that you could look at this but from a capabilities

11. 5:12

    perspective I think making um let's let's just bucket them as Greenfield Brownfield for now like those two categories Greenfield's always in my view going to be easier to go do and this is kind of like a lot of like what Heroku even started as Heroku was brand new project nothing get it to launch get it to out there um the challenge you have when you're doing Brownfield things is you have a lot of history and crft to deal with in many of these situations so our view has been that we want to affect

12. 5:42

    every developer in the world um every line of code you know that sort of thing um the hard hard version of that in our opinion is to go through Enterprises and Brownfield applications so that's why we're doing that um we think along the way we will advance all of software and there'll be many different versions of advancing all the software as well so like Devon is a good example of

13. 6:05

    something that has started to exist in the world a different view on an interface into software development that could be and um so two things here from our perspective is I think that all of those experiments are worth it you know whether it's going to be how you're going to use agents in Brownfield or agents in Greenfield or you're going to do all that work um and then the other is going to be what do you want to do as a as as a company and our our our view

14. 6:29

    here is going to be that we want to power all those experiments so you know the the devans and cursors and Zeds of the world we want to power those so new editors new new ways that people are going to try interfaces into software development want to power those and then we will have our own opinion for our core customer base which is where we build our own applications on top of our

15. 6:55

    model completely tangential question but what is the Green Field and brown field well it's typically say Greenfield is the uh the brand new applications from scratch and the brown ones that you're you're kind of migrating or helping people navigate um you know a lot a lot of stuff would get history or something like that by the way I do need to call out I find it very funny if ISO was on this call we might have the totality of people in 2016 and 2017 who felt like

16. 7:18

    they cared about soft AI for software development with you guys in near AI at the time and ISO with sourc and me trying to find any in the world that was doing this and it was just just basically near I and sourced at the time I find that to be really amazing there was a couple more there was this Zurich team and then there was Microsoft team but yeah AA was one of the few people whom we ran into 201 not a lot of people

17. 7:43

    in the world cared about it at that point were thinking about it a degree but you guys are like the owest of ogs in a couple different spaces so yeah well I mean this is partially where we bonded with Alex because both of us tried to do this somewhere back back in the day I was trying I was trying to actually do it for my bachelor degree but I realized that I will end up with nothing on my bachelor degree and

18. 8:07

    so end up end up doing something else um yeah so yeah same same story um so but but going back to U uh to uh to pool side right so so you mentioned that you guys training your own foundational models obviously like the first question that everybody want wants to be answered when they hear that is how how do you see it relative to uh you know like find tuning 405b right because you know if you have reinforcement learning from execution which I would also be interested to learn more about but that sounds like something you can do on top of a

19. 8:47

    pre-trained model that is somewhat capable right yeah I mean we the general view that we have is that um let's just take let's just put the whole AGI wrapper on the top of this too is that uh you have to have two different views one is if you're building a business to make money for yourselves as Founders or if you're building something that's going to exist long long term and like the way that we've wrapped our entire company um from

20. 9:10

    the beginning was where're SpaceX we need to build a starlink inside of it so SpaceX is that grand long-term mission and vision and if you have that grand long-term mission and vision you're going to go do certain sets of things that just you need to go do and our view is going to be that you couldn't build an AI company in a in a SpaceX sort of by fine tuning your way to that you had to you had to own your own Foundation model to a degree now that's you know

21. 9:34

    people are going to disagree with that proposition but it would be equivalent of Elon saying like well I don't need to go build Rockets essentially I can use somebody else's Rockets to do that we don't hold that view and I think that you know fine-tuning any of those models is is a perfectly fine um approach to building a business it just wouldn't have been for us um and then from a capabilities perspective you have to ask the question and this is one that I hold

22. 9:58

    really close which is is uh how open is it you know we IL you and I've had this conversation in private many times about what that looks like but I also don't think you can serve Enterprises by building on top of those models because of choices that they have made about their data and how open they're going to be with the data side of the fence as opposed to the weights and again you know I was in the room for a lot of the writing of various open source licenses

23. 10:22

    and I still hold you here that if you're not open opening your data um you're not open source on the model side of the fence and I'm just going to keep saying that and I know it's not the popular opinion and people want to say that you know as long as you have the weights you're open but I don't think that's true I don't think the I think the former is the more popular opinion if you like go to Hacker News after every news is more like us and then Twitter is

24. 10:46

    like maybe a different variety of things so there and I think again incentives rule everything around us so I think the people saying one side of the fence are incented to say that giving the weights are open well I think it's I mean it's pretty much until there's like a fundamental bug in a weights and then you're like uh what do we do now right or or or as IL you and I have talked

25. 11:08

    about the past two I think Alex we even had it over breakfast one day but it was a uh until somebody also has some sort of uh really big security issue or data breach issue with the the pre-training data and I think like you know most people who are who are new to the space don't quite understand that that side of the fence yet and uh because it's not in you know as example meta's best interest to be open about that they're not going to be open about

26. 11:34

    that so so for reasons um that uh that were wrong last year I was involved a lot in training foundational models and uh couple things I learned for myself is that a uh the existing infrastructure that is an open source is not adequate to train them fast yeah right like uh I gu flash attention exists but beyond that everything else is sort of slow right like relative to uh I guess the the the number that everybody

27. 12:06

    has right so like I have some Alpha from my short stinted open area in 2022 uh but but there's also number that is publicly available which is in the Lama paper they say how quickly they were training Lama one 70b right so this is their old code base yeah right and that was very fast right way way faster like like uh several times faster than what you would get with P torch right so I guess to train your own foundational models you either spend more money by using what exists or you need to bring people who can write good curs right yeah so so which of the

28. 12:38

    approaches you guys taking so I will generally say this which is there's there's two ways to look at everything you're doing in Foundation model training you're basically getting more data efficient more compute efficient and our view is you got to get better at both um so we have an entire pre-training team of some world class people who are doing everything on that side of the fence from scratch from the ground up training

29. 13:04

    um so again it's like not using off the shelf anything and Alex as you painfully observe too like if you can't control that side of the stack you're not really control it going back to the rocket techology even so far as to say like you know if you're using somebody else's model using somebody else's rocket but if you're using something off the shelf you might be using component parts of the engine as well well again going back to the SpaceX analogy you got to build the whole thing to a degree now that's

30. 13:26

    that's a dicey proposition for startups it's really difficult ult to go do and you know but we've taken the view that we had to go do that and the first decision example as an example was uh not one of the you know one of the first sets of decisions was we're going to build our own pre-training stack and so we have people who are doing all of that we have people who are working on Cuda kernels exam as an example we have people who doing the entire pre-training sack and yeah it's it's a it's it's

31. 13:50

    annoying and it's having frustrating at times because this stuff you know if you if you know what I'm about to say half the stuff is hiding behind private apis at times you have to like work with vendors on and it's really annoying yeah but but there's no plan to open source those those C there's there's we're going to open source certain things we're just not 100% certain exactly what we're going to open source

32. 14:14

    yet yeah yeah because because one one thing that El and I are are coming back to is that ideally we would have uh you know tier one open source Foundation models be truly open source right and so then we would need the whole training Loop built in the open the whole data preparation Loop so that was the second that would have been my second question but you answered it right that besides the kernals you need data uh and the

33. 14:40

    like cleaning cleaning up data and preparing data is uh very hard especially given that the whole pre-training team that we have is spectacular the whole data team that we have is spectacular and then the first two hires we had in the company um outside of you know I saw myself joining where someone working on the training sack someone working on the data data pipelines and data cleaning and processing and all that sort of stuff um but we we we talk about this I just want

34. 15:05

    to talk about this is where we will and won't open source some stuff um because we have to think about what is a near-term Advantage versus long-term so as an example we've trained two two models at the moment or it's a family of models really but we're going to bring them out to the world as the the large language model the chat model and then the code completion model and the code completion model is um a different architecture uh we've been experimenting

35. 15:29

    with it for about a year now and we're going to bring it live but it's still an advantage for us to keep that proprietary right now so we're going to keep it proprietary the world will catch up to us on this at some point in the next year and then we'll probably open source the foundations of those things there so we've not we've not settled on our strategy just yet what we're going to do here

36. 15:52

    yeah I don't know if now would be a good time to dive deep into uh a how do you created Ace uh yeah so ISO named it and that that would mean that since we're both Engineers it's probably we probably had to put some marketing Cloud behind it but we call rlcf so reinforcement learning via code execution feedback and um the the simple thing here is it's patterned after alphao essentially like all the learnings that we had from

37. 16:22

    alphao on what made that um that so successful we're trying to apply those same things to software and the simplest way to say it is that software for whatever percentage you're about to um attribute to this and I'll say hey it's 95 plus per deterministic and we're going to take advantage of that property in software and use that to our advantage so uh effectively how to describe it in at the highest level is we download um you know tens of thousands um scaling up to hundreds of thousands

38. 16:52

    this year projects open source projects uh we make every git commit in them executable and then we ask our models to complete the tasks um we run them in our environment to reward them um we do it not just once but many times so we can do positive and negative sampling side there and two things come out of that one is obviously that's a learning but two is that becomes synthetic data for us so the whole the whole environment

39. 17:16

    here is obviously decently complex um it's a massive runtime environment but this is how we uh we further train our model as opposed to just taking data off the internet and sending it through and learning from that so so what does so so you download the GitHub reper right uh but but how do you choose like I think the most interesting part here is how to choose the task yeah to evaluate on right because in order to give positive reward you need to know

40. 17:48

    what is the what is the correct output right we're not gonna we're not going to give all of the details on that one because that's also going to be some R Alpha here so but uh I would say that we have a way to identify what are really good candidate tasks here and it's not just like coding tasks so it's not just like add function or change or or write test or whatever it's like it's a host

41. 18:12

    of things under the umbrella of what software looks like um but we have a way to uh to to identify a set of tasks inside of there because you're going to use different heris stics like one is it's got to come from a high quality Source from a high quality source so high quality code based WR by high high quality human it's got to pass a certain set of heuristics all that sort of stuff um from a positive perspective and then

42. 18:35

    we also have to judge our own uh the the as the model produces the tasks uh the samples themselves we got to model those not just does it pass all of the different her istics that would say yes you did the job successfully but you're also going to grade it saying even like let's say you do it 20 times 17 of those fail so obvious on the failure case then you can start grading them but three of

43. 18:59

    them pass you want to know which one is the more optimal of the three so you then have to start doing things on that side of the fence as well so this is kind of like where our secret sauce is but I'll just say that we have a way of grading and it's um and and figuring out which tasks to go do uh I would also say that this is something where we invest a lot of time and energy because this is where we're going to get a lot of our own value in the future and as we scale

44. 19:23

    this up too to doing not just tens of thousands scaling to hundreds of thousands of repos and projects this year but to uh you know let's call them those are millions of tasks let's call them tens of millions to hundreds of millions of tasks in the next year yeah this is go ahead no just maybe relevant comment here is what what this really means is that you're going from kind of some amount of data your synthetic data set is actually way bigger and so you actually like scaling

45. 19:54

    up the amount of like data that you can actually use so um let's just use some like numbers that we've seen and like I again like who knows right like but all of us in the in the AI World effectively have like the Dirty Little Secret being that we effectively have access to the same data we're all pulling from the same public sources um and if unless you're using gray Market data which we know that some are but let's just say

46. 20:19

    that you're not you're just using publicly available data we've got something like let's just call them 30 trillion available tokens I think it's actually 50 if you 50 trillion tokens but you you you um you move down to like high quality data that you should put in there let's call 30 trillion or something like that well of that 30 trillion um about three trillion of that is code related data we project that we will generate on

47. 20:46

    that order of magnitude of code data in less than two years so the equivalent of what is available publicly will generate from this approach that's or you mean on the order of three trillion or 30 trillion three trillion so will double yeah we'll double it now obviously that like P projection and so what what we say very publicly is that somewhere in the order of like one to three trillion but it doesn't really matter if

48. 21:18

    you think about it from that perspective what IL was saying was true which is even if it's a third of that and it's a one trillion like well if there's only three trillion available like this the quality of our data is is high because of the approach that we take but just the availability of the data is is enormous and you know this is the kind of the Holy Grail that everyone's after and as we're all kind of talking here the other side of the fence is everybody's trying to do this everybody's trying to do it um but a lot

49. 21:41

    of folks are trying to do it on language which is one it's like easier to do but harder to judge because it's non-deterministic so it's easier to generate synthetic data but it's harder to judge and you end up with entropy problems inside of that side of the fence pretty quickly on this side you know it doesn't appear we're going to end up with the entry problems it's just harder to build a system that can go do this in a quality way but it's also easier to judge because it's

50. 22:05

    deterministic so we have advantages that we're trying to take um you know advantage of in software to go do this I I find it interesting that U uh there's multiple so so now like you know with your with your uh story of building this AAL CF uh it's it would be the third example in the recent history where someone came you know said You know guys we're doing like interesting ra I I don't know if you guys using any research or like if it's more linear but you know like open AI

51. 22:37

    was like well open didn't want to say that but someone said open AI is doing Q Shar you know and Q sharp everybody like that must be research I'm not gonna say anything but it's it's a it's it's yes I see that is something that's not not anything but yeah yeah and Deep Mind came and they're like hey guys we're doing like Alpha proof and Alpha proof is a Tre search uh in the space of lean programs and then you guys doing Tre search in the space you know

52. 23:07

    like like based on the execution feedback and I find interesting that in all three of them uh you know like as a you know any person who is into machine learning the first thing people do is they go to the Whiteboard and try to you know like how how would I do that right and an all three example like my personal Journey ands uh you know getting stuck with figuring out which data to use and specifically not even data to unroll but the data to for the

53. 23:30

    feed for the actual reward right like yeah like if you do strawberry right I look at GPT 401 I'm like yeah maybe like like I'm pretty certain that that research is very sophisticated and build it by the brightest minds but even before we get to the research how do we how do they choose the reward right like yeah this it feels like um everyone kind of like gravitates to the same set of things like you're in like we're all kind of in the same park if if you will

54. 23:56

    and then we all end up similarly like there's a problem to be there's a a major problem to be solved to to see and be able to explore the next Park it's like a locked gate and we all kind of run up against it um I I'm I typically have started to work backwards and the team that we have here some of the the best on the RL side of the fence kind of in this domain cared about this for quite some time either from Deep Mind or other other places and um the way that

55. 24:22

    we work back is backwards is what would it take to do what I call full program synthesis in the future like if you were to use a system like this to automatically write full programs how would you theoretically use it you could kind of work backwards from that to determine and then you end up at the exact same spot Alex that you were talking about which is like okay so which data set do

56. 24:46

    we use and how do we do this and how do we scale it because it goes back to the goes back to the bitter lesson doesn't and it always does go back to the bitter lesson which is hey you more compute solves a lot of these problems because you can search through these things quicker but I think like yex question is yeah what would be the seed data and the answer is like it's a function of existing data data set data that's already there that

57. 25:13

    you permutate in some ways to create kind of candidates yeah and then the the idea here would be too that you're not creating a singular you would try to be creating multiples and then they would fail spectacular ACC some of them at certain points like yeah that's just wrong right again taking advantage of software you can actually deterministically get to this point quickly in some scenarios because one

58. 25:37

    doesn't compile or syntactically wrong won't run or there's errors and things of that nature plus a host of other things so again it's not just it's not just a theoretical problem you can actually practically put it into an environment and start testing some of those so you know if you you don't necessarily have the right answer you can at least um uh send it through something and say like okay so you know we didn't per have that perfectly mapped out but we can also short circuit what

59. 26:03

    we don't know at the moment I'm curious so one of the kind of interesting things right with scale right we see the capabilities are growing but it's still like clearly missing some conceptual things about you know algorithms about some of the relationships between um entities kind of how kind of how are you looking at that like is it just like let's just scale and build a bigger model and it will hopefully it will figure it out or kind

60. 26:37

    of you are thinking about it in more nuanced way uh I would say that you know we're we're an applied research group so kind of going into computational and data efficiency we're more focused on the data we're making everything computationally more effective for training uh in the current approaches and data from a data efficiency data usage perspective and data generation perspective Ive we kind of went into this um at our current stage I wouldn't

61. 27:02

    say that we are we have a dedicated research group saying what comes next from an algorithmic breakthrough perspective um and you know that's just that's a function of scale and where we are from our own maturity perspective but it is something that we have to invest in in the future and obviously it'll start with let's just say dedicated research more in like the domains that are applicable to us right away and we have already the applied

62. 27:24

    version of that so RL would be a you know a good for us but kind of like the classic breakthroughs that I mean obviously Ilia you know talking about here in that way we're not it uh set up for that at the moment but we do need to in the future it's just that comes with a function of success so the more success that we have from a business perspective the more we're going to invest in there but our primary if you

63. 27:48

    got to think about what what matters from a focus perspective as a business at the moment there's only a couple of things that we need to focus on right now to earn the right to even be able to do that in the future and this goes back to bad SpaceX analogy but it's very simple like I look at this way our Ambitions are X like our Ambitions are SpaceX to explore the universe colonize Mars but right now we also need to build a massive economic engine to fund those

64. 28:11

    Ambitions and so that's our starlink so we're very much focused on the Starling side of the fence like we we need to make enough money to go further the Ambitions um and so you know we we don't have that Mega philanthropic donor giving us 10 billion doar uh to go unbounded compute type of scenario so um but in the future very much so I think it's needed I think it's absolutely needed and I think also you need another

65. 28:37

    thing that we need is we need more more compute diversity you know we have effectively one one Mega company where we can get all compute from and I think that when there's competition on that side of the fence as well we'll see more advancements there now that's not what you were asking about but I think that that will actually benefit us as well because I think we'll see more faster

66. 28:54

    advancements over there so but how however you did mention that you have so so you mentioned in passing that you have two models right one is the chat model one is the code complete and you said that for code complete you spend you were spending like one year of experimenting so so those are not experiments on the foundational side of of things I just think of it as different architectures so um you know we I think I think the world's going to move on um to a set of new techniques as well here

67. 29:27

    and new architectures um but you know ours ours is the version we've been playing with for um that get us some benefit particularly for code completion So when you say new architecture you mean like iterational a Transformer or not a Transformer at all this one happens to be a non- Transformer not a Transformer yeah in 2024 this was fascinating I know yeah I know one I know obviously the irony that talking to Ilia here but I also know some your EXP that too you were looking at that as

68. 29:59

    well weren't you Alex I like I was interested at some point in uh there's a company called symbolica that I was talking to and I was fascinated with what they do but it's uh I I think theirs is a little bit more farfetched at least last time I about but I haven't talked to them in more than a year so it could be that they closer now yeah but I don't think they were in a state to ship something like this year I don't know

69. 30:30

    yeah but when you when you say it's not a Transformer and and yeah it's obviously okay if you cannot talk about it right but it's a is it like a is it still like a differential model like LM or what note or is it completely different it's not so these are not like brand new everything from scratch you know like uh we're going to publish the paper and we'll be touring on the the

70. 30:53

    researcher circuit type of deal these are things that we all kind of know about but we're not sure they scale yet and and and and without saying what they what we're looking at and what we've done it's things that we we said hey that that's a pretty good idea I wonder if it will work at this size I wonder what will happen here and so that's where we spent our time right figuring out what works from a conceptual level on a paper to something that could scale

71. 31:18

    to a certain size yeah yeah that's that's pretty exciting that's uh that I would be looking forward to yeah and that's I think this is one where in a couple of weeks you and ISO and I all four of us should probably catch up on the thing that we're we we've done um and what we're happy with and what we still want to iterate on this one too but butl RCF is it on top of the code completion model model or is it on top of the chat model it's um we're using it in both but primarily the chat model right now um code completion we

72. 31:54

    use it in certain ways uh we do generate data for it is as well we don't use the complete set uh for the code completion side obviously as you would expect we don't use the the complete set over there um uh primarily for the chat model and then what we do for the code completion model I think you guys know our approach too is we're going to release this publicly for the world but we're also releasing this for Enterprises which means that we bring it

73. 32:17

    to them so inside their VPC or data center in that sort of way so when we do a um fine tuning um on this uh on the system on their data it goes into both of them so the fine tuning happens across both of those models for them Al and date of Enterprise fine tuning happens on their machines so you don't see the data we don't see the data that's our promise and we don't need to see the data from because we can

74. 32:42

    generate our way out of the data hole so we can gen and so that's again like part of our major investment here is that we're going to invest in the rlcf because we can generate the data so that we can give the customers what they want so because no customer really want like no customer is ever going to really send their data over to open AI or anthropic to go do this uh they want to have that control but they really want it inside the systems inside the model specifically they don't want to just

75. 33:06

    have it ragged inside there they want to have it so that it's the model itself becomes theirs y so maybe uh zooming out a little bit um kind of would be interesting to talk through how will this start changing the landscape uh in the broader sense like I mean we we've had this conversation before where you said this is this is the last wave of Technology pretty much uh to be built because

76. 33:39

    everything else will be built by this technology um well how how high level you want to go because I love this conversation but it's always fun to to see how how how will high level people want to go because I can go anywhere let's go as as it gets sorry what was that Alex as as high level as it gets so so so here's a couple of things that like I like to net out because I think like you know what I didn't mention in my biography earlier was that I spent a couple of years investing as well um at

77. 34:11

    Red point and doing where sorry for you it was I'm I'm I'm back on the side of the fence just put it that way so um and I think there's a reason why I'm also back over here too because I I I don't I actually feel bad for investors in this moment and I think the the the easiest way I can describe that is because there is a lot of that just doesn't matter anymore like if you like play out 10 years 15 years there's a lot of stuff

78. 34:38

    that doesn't matter and I'll I'll explain why in my view um first I think like if you really want to like distill down what will actually matter in the next 10 years plus type of deal it's basically electricity chips that make and run AI intelligence layers so three things three things like electricity the hardware chips and then intelligence layers and we'll get to intelligence layers in a second but like you can already see this you can see why uh

79. 35:08

    Amazon's buying a nuclear nuclear power plant and Microsoft's doing the same they're investing all these things over like Tri electricity is a real big deal like I think the state of Texas just said we've got no more electricity for data centers like literally today I think I saw it on Twitter somewhere like this is a really big deal all of a sudden um you know in in my view electricity is actually if you think about it from the United States perspective like this is kind of at at

80. 35:32

    the root level like they we need to be thinking about this because we're you know people say hey we have enough electricity like that's a bunch of we don't have enough electricity for what's about to come we just don't um and then the chips chips that make and run AI U you know whether they train and run them and then the intelligence layers now what I think is is going to come and let me start by saying what um my view had always been a

81. 35:54

    GitHub and still is today is that we with GitHub co-pilot which is the most successful in my view application of commercialized AI which and this because it's the domain of software I think it's the most commercializable domain in the moment um is we effectively when we released it we effectively started to save developers time and I think that's the one of the best way to think about this either save time or gain time and the first version of GitHub co-pilot was

82. 36:21

    all about code completion and it was saving developers milliseconds to seconds then chat came along started saving developers minutes to maybe tens of minutes then you know you start to enter a phase where we start to save developers hours uh or days um and then you another way you can think about this is how much time do developers gain in their lives to go do these things and what were was happened is like we you

83. 36:44

    know with the release of GitHub co-pilot we went went from every line of code on the planet Being Human endeavored and then with the the release we went to Dev Le AI assisted which will quickly give way to AI Dev assisted which will quickly just get to way to AI everybody assisted Ilia you alluded to this before like skipping the whole phase of like working in the Enterprises and stuff and

84. 37:07

    that's when you know we go from 100 million developers in the world to two billion people being able to create and all that sort of stuff iterate with these things and you know hey you know in my my Partland hey poolside create me a Twitter clone great here's four running copies which one do you want to iterate on like that's a world that's going to exist for people at some point here soon and the other's going to be

85. 37:29

    you know hey you know poolside do XYZ great here's a couple of running copies of a thing like this one has already passed all your internal check flights from an Enterprise perspective it's all good to go let's iterate and make sure you know this is the one that you want to go do so it's like there's infrastructure underneath these things so you know if you want to say the first 40 Years of computing were the deterministic version of computing we've entered the non-deterministic version of

86. 37:55

    the Computing age and it's all about in my opinion so I rambled a lot but I tried to like put my thoughts into a box so that we can p it well I I think the I mean obviously agree with with all of this but I actually I I always challenge to go one step further because when everybody is a developer or has ability to create software and I kind of view this more as changing how we interact with Computing yeah and so when change

87. 38:27

    how you interact as Computing the the interaction itself like we don't need software anymore right because like software is just an intermediate state of the Computing uh so I mean I have a view that eventually we're going to end up in what I call a femoral software land basically like if you bad analogy I know what I'm about to say isn't quite uh work yet and I'm still trying to sort out how this might

88. 38:52

    look but I could see a world where we just write right software like source code itself on the FL fly and it only persists for the moment in time that's needed to go do a thing and think of it as almost like a Proto buff exchange like you don't care about the protuff itself you care about the messages that are being passed back and forth we already have this with networks with TCP I message passing or UDP message passing and all that sort of stuff I kind of

89. 39:15

    think we're going to end up down a similar path at some point not for all software but and not for All Digital systems but a lot of it doesn't need to persist past a certain usage Point too so you know like it's going to be possible to go do right now we don't do it because the cost to do that is so incredibly High yeah I think I mean I would split this into maybe three categories there's

90. 39:39

    one that's definitely can be completely um kind of temporary and this is application like the user interface part because that can be generated on a fly based on specifically users kind of context and what they're trying to do and we have a demo of that that Alex showed uh kind of last week a token where you just generate on a fly UI for yourself for the actions you want to take or information you you're seeking

91. 40:10

    uh and so like as it gets faster you know more more uh um kind of correct uh like shorter time more context about you as a user and you can imagine that being completely predictive even where you don't even need to like type in prompt because it like it knows you will want to look at whatever stock prices in the morning or like news or something and just generate that then there is a kind of the transformation of AI Services sorry of

92. 40:42

    of API Services right of back ends because they kind of they govern some data they govern some uh logic um and you know some of them Marshall physical bits right like we still want to eat stuff we still want to you know live somewhere Etc so those um those will become more semantic right where you can like talk to it and kind of interact in more natural way but at the same time and this is where kind of what some of the things we're looking at as well like

93. 41:16

    you want them to actually work within specific set of conditions right you don't want it to like accidentally you know spend all your money or something and so you will need some form of verifiability coming in as well uh to ensure that the the interaction you're doing is actually executing correctly and so that's where we've been looking at kind of formal methods uh and specifically lean to kind of provide you lean is in wait sorry which

94. 41:48

    lean the the kind ofing yeah the the the the thing in which I deep mind the the Alp proof yeah okay that one um yeah I go ahead keep going I didn't mean to Sidetrack you on that one no no I mean that's and and then there's like a shared State kind of uh like I I I bundle them under marketplaces because like it's some kind of shared shared things that you want to be ideally AS Global and as open and permission permissionless as possible because you want to get you know either best price or best execution or

95. 42:31

    something uh and those um really belong on kind of blockchain like infrastructure to be able to facilitate that and those will be even more permanent and kind of because like you don't know like the front end can be generated on a fly you know you can talk to it through some AI agents A Services uh but the marketplace at the end is what provides settlement you know you're actually getting the apartment or you know the food is arriving in whatever and there's like

96. 43:02

    some some facility that actually Marshalls like as a physical bit so digital or whatever so that's kind of how I'm thinking like a a stack of three layers that uh we're going to converge to where you have kind of Market marketplaces that have static like not static but like the Govern shared State you have API like AI agent like API services and then you have kind of dynamically generated applications that

97. 43:33

    now just almost like embed into your operating system to have kind of as much context about you as possible and be able to predict ideally like what is the best experience for you to have and that's why to me like if you look at that future as you mentioned like the brown field is seems very not that right because a lot of the software right now is kind of like is about the existing state of you know shipping monolitic apps and kind of creating like long

98. 44:05

    Supply chains of of things because we don't have like Global marketplaces so it is an interesting question of like how's that transition will be playing out as this thing becoming like easier and easier to to create yeah I I always go back to um working with Enterprises and working with individual developers are almost entirely different at the end of the end of the uh end of the day and so you have to treat them as two

99. 44:30

    different organization or two different groups that you're kind of building towards and I always think if you're building towards an individual developer you're doing that green field on you're doing like hey help me get started on something brand new like the side projects and then Enterprises or how do I have maintain my legacy and that sort of thing and they're two different

100. 44:47

     things and um both of them I think advanc but the when you're working with Enterprises the simplest way to say this is like you can't give a spaceship to a caveman you have to help them evolve along a pathway to show them how to use the spaceship and uh but I I think I I mean I share your view of what the future's going to look like you know I think um you know I've talked about this before but probably the the two most

101. 45:11

     influential programming languages for me over the last 20 years and in my youth one was a weird one called Eiffel which not many people know about and the other one was Earline and Eiffel had this Concept in it which was uh designed by contract or programming by contract and you know JavaScript has this now in promises and things of that nature but it's all about hey if you enter a function we're going to guarantee a set of things and if you exit the function we're going to guarantee a set of things we're we're done and I think that what

102. 45:43

     you described in your second layer at the API level is actually a lot of what we actually need to go build inside those things because the way that these things will interact and you know the message passing and the mailboxes and the way that earling flow but the ephemeral interfaces and like things I mean those things are going to be the natural first step I think that's where

103. 45:59

     we're going to start to see these things is you know almost custom uis for various either surfaces whether it be mobile or phone mobile or desktop or whatever but eventually like you can think about the algorithmic feed that people see on social as the first version of this which is it's got a container wrapped around it but the contents of that change on a regular basis but the container itself is going to change at some point too it'll be

104. 46:22

     it'll be ephemeral as well yeah I think well for me I think the container will just become the operating system because right now we yeah switching between Twitter Tik Tok Instagram like it's all it's it's the same mechanism it's just like different pretty much different world guardant of content but um I either one one of the operating systems SL slash one of these big companies do this or we actually going

105. 46:56

     to have an open vers but uh you can kind of Imagine That becoming the operating system and then you can you know from there kind of Spawn into more other activities and and like issue different intents right be that um productive or not but I guess like bringing it back a little bit the I mean that that's kind of one of the things we've been thinking and Alex specifically been thinking about like what are the Innovations we need to to ensure this correctness of the code because the other side of this is like

106. 47:30

     right now we're training all of this foundational models on the code written by people which we know is highly buggy and so as as we're training this models they're going to continue producing you know as buy code um and vulnerable code and you know as we deploy it in more and more critical systems that's going to just continue increasing the overall issues in the kind of system so how how do we how do we circumvent that and kind of make sure it's correct by

107. 48:02

     construction I mean I think this is going to be the I mean I'm glad to hear Alex is working on it um in that way I think this is going to be an area that is really pushes the boundary because again going back to like what exists today from a software perspective people are looking at these things as like again code completion and chat like oh can you do an X for me and I can copy into my project or maybe do a green field but you have no idea in correctness and like in reality we don't

108. 48:27

     have an idea of correctness outside of post Hawk mechanisms today in software anyway the entire stlc basically exists to try to help people guarantee a set of things and behavior shaped and in Enterprises is about Behavior shaping in the real world um outside of Enterprises it's all about like does it compile does it run do I have vulnerabilities can I deploy it all that sort of stuff but no one wants to do that sort of thing like

109. 48:52

     that's like that's an art should be an artifact of a bygone era at some point in the future but it's going to take a lot of effort for us to be able to get to that quote unquote correctness particularly computationally efficient version of that yeah but the uh I think I think we have the the tools to build code that is correct by construction it just it's very tedious to use them and I mean depend depending on how how we see the slope of a it could be that you know Lan proofs will be

110. 49:23

     generated end to end you know in a very short period of time and then it just you know makes no sense to not only write formally verifiable code yeah and I always like using this example I don't know if I ever mentioned it during office hours probably I did but the one of the arguments against formally verified code is that the bugs will just propagate from the code itself into the uh what you call contract in eiel right so uh you know like the code

111. 49:53

     is formally verified against the set of assertions but assertions themselves correct well we're always going to we're always going to end up in a spot where what we asked the system to do was wrong or we had an assumption that was wrong from a human perspective like there's always going to be this human divide to a degree um I mean again you could be 100% correct in asking the wrong question right right but Beauty the beauty is that um uh when you're program

112. 50:19

     in lean and you have a function or like you have a proposition and the proposition is not quite what it seems to be right so you look you look at proposition and it seems to be that it's saying one thing in reality it says another thing the beauty is that is you don't if you don't use the uh the underlying you know code directly but instead try to plug it in into something of yours that has a higher level

113. 50:41

     proposition you will not be able to plug it in if the proposition is not actually fitting into the you know underlying task right so I imagine the way this is going to work in the future too is going to be that it's a cold star boot um cold star problem too like so like as more of these systems like what you just said is true but it's like think of that as an inheritance thing so it's kind of like the same problem or same same way that

114. 51:04

     we've been thinking about writing software for a long time in that way but also as more systems and are written in a way that you can say is is robust enough I'm going to use the word robust I'm gonna not going to say this but like you could say verifiable but the first one's going to be the worst one and then d d d da when enough of those are are are done and they can interact with each other then you have a degree of confidence as well so your the

115. 51:27

     confidence value that you have on these things will obviously increase the more that you have there and the more that they're only interacting with themselves to a degree too this kind of is the same way that if you think about the way that we built deterministic versions of these things all the way from the operating the the assembly all the way up is there's a we no longer have to think about a set of systems anymore because they've become robust enough that we can

116. 51:51

     build upon them and yeah they still have faults but it's to a deg to such a minor degree in this way here I think we'll end up in a very similar spot it'll take years but it like what you're looking at is the very first version of doing that I don't I'm not familiar with lean enough to go into like how it's going to materialize the very first time which is always the the interesting one how does it materialize the very first time that we you know and

117. 52:15

     we boot it up there yeah yeah ly I would highly recommend anyone like both you and everybody listening to play with Linn it's a relatively spiritual experience uh to get a little deeper into the weeds uh and and I think I'm pretty certain I'm pretty certain that when you code in lean you sort of improve uh the the lifetime of the universe because every time you have a proof in lean the entropy decreases of the world but I think I mean you Alex didn't

118. 52:44

     finish the whole point that like this this kind of if you have proofs all the way down you actually as a user now like from a user level you can now instantiate the specific conditions right you can say like hey you cannot spend more than $100 and bugs in my propositions right but this is a very limited scope right like like literally if my only proposition says you know my net worth doesn't go down after executing this transaction right then I know

119. 53:17

     that you know like if if that one is properly written and I verified it that then it can execute you know millions of lines of code underneath but I know that as long as I didn't mess up that one proposition nothing bad will happen so that's a I think I think I'm going to make that one my number one probably to start out then when we're doing these things just well better increase net worth yeah well I mean I guess one's the failure case and one one's the bear

120. 53:41

     one's the bull like so you got the the exit condition is that I did not go down I would have like it if it went up so yeah yeah but but like an example would be the if the smart contct has a bug that that could hypothetically allow someone you know like to withdraw all the money uh then then you know yeah what proof the smart contract has then therefore are incorrect and then I wouldn't be able to plug them in into my you know in into the proof of my

121. 54:05

     statement right and at the end of the day I wouldn't be able to invoke it well again I think like what you're saying is um kind of going back to bad analogy and I got to go read this uh in depth and try it out now is uh uh it's a it's a set of guarantees that someone else is is getting like hey if this exists over here going back to my Eiffel example the way that the reason why that was important if you're designed by contract or programm by contract is that

122. 54:30

     you no longer have to think about certain concerns and now in Eiffel it was all about business concerns like you know X and Y like hey if you entered this function it means that X was was guaranteed like whether it be that the person existed or they had a net worth over a certain thing because you're going to do a credit check and if they exit it it means you approved it's an approved function you approve the credit check um or you know that type of deal uh or or or maybe not that it's approved

123. 54:55

     but it's like a they have a credit score of this they blah blah blah so you don't have to do any of those sorts of checks it would feel similar like okay so we can we build a base set of these things and we can herit them then we can build more Bas set of things and herit them build more Bas sets of things and you can build upon them all the way through y the beauty of Lan is that I believe that I might I might be wrong I think I read it like in passing somewhere in the web but I think the Lan compiler itself

124. 55:20

     is formally verified against a simpler version like the was something very simple and each consecutive step but that that might not be the case that that might have been about something else but even if it is not the case that is something that can become the case in the future you begin like with extremely simple uh very self-contain you know verifier and then you build up from it

125. 55:44

     so that you you you you know that the actual verifier tool is also correct how far are you on this in this exploration at the moment Alex uh with L we're relatively early we don't have anything automated yet and I think I think this step this there there's a couple of interesting steps we're about to take we we we're looking into some synthetic data generation similar to what you were

126. 56:10

     describing uh so Lynn has a beautiful property that unlike many other languages you actually do know if you have the reward like Lin compiler will not allow you to compile the the the proof that does not prove the proposition so so research is easier like in case of deep mind you know like when they publish Alpha proof it wasn't the problem of the reward it was more the problem of how do they see the actual data on which they do the unrolls they start with like one million of

127. 56:40

     human WR problems and they someh canorth them to lean but I think that step that magic step is is missing I mean we were discussing as Alex actually so one of the kind of questions of method mathematics right like which which the theorems are interesting right like which which things to work on which them are interesting given you don't like there's no use case yet for it like what what would be the and like it's really hard to get use cases for theorems before you

128. 57:12

     actually get them and prove them then people can build on them right and so the interesting thing here is the interestingness can be defined by how kind of how straightforward is for the model to prove it first place yeah what was the no no go ahead yeah I wanted to clarify but maybe that's what that's what you were about to say yeah but like if you do like a research right if it's a very trivial theorem you it will very quickly get to the proof and if it's

129. 57:43

     very hard you're just never going to prove neither of those is interesting but for an interesting theorem it will get to the proof eventually but it will explore a lot of possibilities right so the bushiness of the tree is what defines the interestingness of the of the theorem right and and that is something that is very quantifiable right so you can you can do the whole search completely uned by humans by starting you know with a with a set of

130. 58:07

     random simple propositions and then you just you know like your Frontier uh is all the propositions for which the buness is high and then you have model that proposes new uh interesting propositions and another one that is uh you know predicting their bushiness and then that is like you know another enforcement learning is uh training those two uh so that's a you know like that's a

131. 58:32

     fully yeah that that's um uh that makes a ton of sense it's fascinating from that perspective I think like uh what I took away and I really liked what you just said too is that um there's a set of uninteresting ones which are uh which are the binary outcomes like it's not the fact that it did or didn't it's uh it's where there's the possibility and this is like where I think a lot of us spend a lot of our time which is like how do you differentiate when something

132. 58:55

     is um inside that bush I think that's what you're calling it that like that set of branches out there that's that's fascinating okay um so any the audience wants to work on that let us know yeah I was gonna say like uh that's a fascinating problem to be working on what do you and ilot to your point um until you have a problem for this to be Sol so how are you thinking of even like experimenting with this inside a near I assuming you're going to be putting this inside the what we actually

133. 59:30

     have a problem it's called smart contracts okay is it going to be underlying principle for like the smart contract then at some point well ideally it's through the stack right that's that's what I was kind of saying ideally the user def like has some set of PR like conditions that then propagate across all the interactions you have but on in near specifically right now like in and in general in blockchain space

134. 59:53

     you want to have an absolutely correct contracts like you want to make sure that when you know you're deploying something it's not going to spend you know lose people's money it's not going to do things that like it's not expected now the expectations obviously are defined by a person and so you uh like they may have issues but this is where kind of this composability of this preconditions can be really powerful because now other contracts and users

135. 1:00:20

     can Define like hey I will only put my money if it's not GNA lose my money in a future kind of condition and so uh you know if that smart contract doesn't prove that it doesn't lose users money then you know this this conditional fail and so that's kind of the idea is like we can actually put it like on the protocol level almost where your transaction contains your preconditions and smart contracts have the uh kind of

136. 1:00:45

     the you know theorems that they uh postulate and so it will literally fail the transactional fail or part of the transactional fail if those conditions are not met um or viously this is like a little a little bit longer term but like meantime we need to build the facility for being able to prove the things and generate smart contracts from this yeah I mean from a smart contract perspective makes a ton of sense and I would imagine

137. 1:01:10

     there's like a lot of different ways in which you could benefit from this even if it was um just the fact that you want to use a a contract it it can prove a set of things but you can also inherit from another contract or do another contract would say from a security perspective let's just say like one is a financial oriented one but one's a security oriented one and they have to be the composition of the two together for you or maybe one is a precondition for some other postcondition ey I'm just

138. 1:01:34

     trying to think through live but it makes a ton of sense there because what you're trying to do is with smart contract is guarantee things yeah but I think this is I mean this is going to be obviously very useful for you know Enterprises and everywhere else because this is kind of at the core that this is a business logic guarantees that everybody wants yeah I mean again going back to the thing that my again Eiffel which I don't like to reference I got to find a better reference for that for

139. 1:02:00

     that because no one knows Eiffel but it's the reason why is is that was so important is because it made it so that programming became interesting again because you can guarantee things and that I think programming is interesting when you can guarantee PR and posts so yeah no you will have you said you had two languages that that made you experience spiritual you know yeah that was

140. 1:02:23

     so Eiffel and erlang yeah yeah I I I think I think that number is about to bump to three I I feel like another what you're talking about here is like I might be going down a rabbit hole where ISO is just gonna be like stop talking to me about this right now we have to actually serve customers all right I think we over time so I really appreciate you uh coming in hopefully people learned more about pool

141. 1:02:50

     side as well as you know our kind of General conversation on AI developer that and the repercussions of that coming so uh thanks and see you all next time thanks everybody good be