LLMs, OpenAI Dev Day, and the Existential Crisis for Machine Learning Engineering

0:04all right welcome everybody it's Hugo bound Anderson here and I'm very excited to welcome you to a live recording of the vanishing gradients podcast I'm here today with Jeremy Howard from fast.ai among many other things Tre shanka from UC Berkeley and haml Hussein from parin Labs we're going to get started uh in around 60 seconds to talk about llms the exal so I can myself now um the existential crisis for machine learning engineering and open AI Dev day um in the meantime I would love if you could introduce yourself in the chat let us know uh who you are where you're watching from and what your interest is

0:43whether you work in the space professionally or you're just more interested in learning more about these types of things um and we have a question will this be available as a recording it will it'll stay live on on YouTube um in perpetuity all right we'll get started in uh just under a minute

1:32all right well that feels like just under a minute to me um so team perhaps we can all turn our cameras on um and mics on and jump in Jeremy haml Shrea hello hey there great great great to see you all and I mean there's so much happening in in the space I feel like I speak with you all um somewhat regularly both in in these formats and and other formats it's so exciting to have you all um here today as I mentioned just a bit of bookkeeping um this is a live recording of a podcast I do called Vanishing gradients and I'm Hugo B Anderson who works in I work in

2:14AI education what we used to call data Science Education and machine learning education um but now we have many different terms floating around I'm just going to put a link um to the podcast in the YouTube chat if you want to check it out it's available wherever you listen to podcasts as well um and the most recent episode is actually with the one the only Hamil Hussein here who's working on on the front line of deploying llms and sharing all the kind of wild lessons learned from you know the wild west so definitely definitely check that out um if you like you have some kind of wild Applause sound effect

2:49button you can press I was actually thinking of rigging you know like the old radio shows that like you know all of this ding ding ding ding ding um been a huge Roar of approval at that you point I'm I'm I'm on to it we'll have I'll have to have you back and and have have that done I um was going to say if if you enjoy these types of conversations and think your friends or colleagues would would enjoy them please um tell them to check it out and uh if you enjoy it hit subscribe wherever that button may may be but enough about and if you and if you don't enjoy this kind

3:22of conversation reflect on your life choices seriously absolutely um and and don't let us know as well um we don't care I um also if if you have any questions that you think are relevant to the conversation please ask them in in the chat and I'll do my best to weave them in to the conversation um and um but without further Ado I just I'll get everyone to introduce themselves in a bit more detail but I'm just so excited to be here with with SH shanka who is at UC Berkeley but um ex Google brain Facebook and Viaduct Hamil Hussein who's currently um started his own um super

3:58interesting uh cons consultancy pance Labs but X Out of Bounds GitHub and Airbnb as well um and Jeremy Howard fellow Australian co-founder of fast.ai honorary professor at the University or uni of Queensland as we call it former president and chief scientist at kaggle um creator of the UL mfit approach on which all modern language models are based but most importantly a connaisseur of delicious biscuits um which we've discussed many times on podcast the first podcast recording we did people said why wait wait wait Americans don't know what biscuits are there aren't even biscuits in America really the closest they have a cookies but they're not quite the same are they they're not the

4:41biscuits you get in the south of the US exactly I remember um cookies once when I came to San Francisco through Australia I hope people aren't dropping off or like to talk about this type of stuff the whole time but I I I I brought you and your your family um a bunch of timtams which um you know you and and and your daughter were absolutely amazed no people are engaging um any biscuits to recommend is the most important question in the double coat Tim Tims double coat and and the yeah double coat and um I remember yeah mint slices and if if it's a hot day in summer put them in the

5:19freezer just for a bit and then then get into it um but so let's so I think we're all done so we'll it up yeah this has been a great conversation and I appreciate all all of your time but um I'd love to just go around the the proverbial table the the zoom room the YouTube stream and get everyone to introduce themselves um briefly and let me know a bit about why you're so interested in llm so Shrea I thought we could start with you yeah um hi everyone I'm Shrea I am a PhD student at UC Berkeley studying HCI and databases um and I am particularly

5:57interested in data management for machine learning centered around you know how can we improve ml Engineers lives like what's that human CED Focus what are the problems they face um and how can we help more people become ml Engineers because there just aren't as many as the world needs um so what really excites me about llms is that they they really um they they're way they make machine learning more accessible to use in software projects so you don't need to have like very strong um ml expertise or background um but you still run into all the same problems that ml Engineers run into when they deploy their ml models um so I

6:38think it's just a super fun world of seeing you know how do we build better tooling for these people amazing and I'm really excited later on to to dive deeper into what you've seen in the space and how the like the mechanics of how these things have changed and kind of what what you see happening what you see happening in the future haml um tell us a bit about yourself and why why you're wait wait wait wait wait wait wait wait let's not do that I'll introduce haml because Hamill's too humble so I do it properly I love it haml haml for you those of you that

7:11don't know is the guy that led the team that created the first coding large language model product at GitHub code search Cod search net yep it was awesome and so it so impressed the people at GitHub that they shut it down and bought something similar from open AI instead so I don't know what they were thinking but it was really awesome and it set set the agenda in terms of what's possible I don't really remember such a big successful kind of commercial roll out of a product large language model ever actually coding or otherwise So Co so Hamill's been doing this stuff for a while and he's quite good

7:55to thanks Jeremy both of you have worked together right y yeah I mean I first learned about language models through Jeremy and so owing to Jeremy I got started on all all of this stuff and continue to actually learn quite a bit all the time from Jeremy and so perhaps Hamill you could say a bit about why you're so excited about llms now and a bit more context around Hamill I mean Hamill's been working machine learning way before it was it was that that popular as has as Jeremy as well um but been working ml for 20 plus years in in consultancy as well and kind of come full circle to it

8:36now with llms yeah um why is it exciting it's hard to say I mean it's it's a it seems very obvious that they're so uh you know you can do so many things with them they're very capable it's a very easy to use substrate to do so many tasks to interact directly with humans um it opens up a whole new world of products that you can that you can develop and make so I think yeah that is really exciting awesome ham regular text chats where it kind of go something along the lines of like how are you feeling today about lm's existential dread or excitement and enthusiasm and you know

9:21the both of us it seems to go backwards and forwards on a regular basis or both at the same time it definitely depends on the day hey Jeremy we've just got to I can hear you very well we've got a few comments that your audio is a bit lower than than ours I'm wondering you able to if you put a bit closer that would that'll be good um but in the meantime I think it would only be fair to have a sense of symmetry here so haml would you mind introducing Jeremy and I'm just doing an audio check now check check one two three so people can say if

9:51I'm too quiet still it sounds great to me okay yeah yeah I can definitely try it's very uh it's definitely a tall order to introduce someone as impressive as Jeremy um so you know Jeremy has been doing machine learning for a long time he trained his first neural net very very long time ago I don't even know how many years ago that is Jeremy how many years ago was that it's like super long time ago my first neuronet 30 years I think yeah so it was like even before they were cool or even probably a lot of people knew about it um the last they were cool you know they were kind

10:28of somewhat cool then and then they were became uncool um so yeah um you know I guess people know Jeremy for a lot of different reasons um you know Jeremy is a has started many successful businesses um and but then more recently uh you know he's created fast AI which is a education and research lab that has educated many people um on on machine learning and AI including and that's kind of how I got to know Jeremy and sort of you know that's what helped me uh in a lot of ways and um you know Jeremy's actually actually very uh Central to this uh discussion about llms in in a very big way um so

11:22Jeremy and uh several colleagues produced research ULM fit which we may get into in this podcast which was kind of one of the foundations upon which large language models the the fundamental research on that like kind of enabled a lot of these uh one one of the fundamental pieces of research that enabled large language models um particularly around transfer learning which is a really big deal and um and you know right now you know Jeremy is continuing to do research and education and you know is a really big force of nature in the field and does quite a lot for the community thanks so with all of that in

12:05mind what what's so exciting Jeremy for you at the moment about large language models I don't know I'm trying to decide how excited I am still um it's it's quite transformative um I think like like so something I'm really excited about is when my wife Rachel and I started fast. aai the the goal was to make AI more accessible so that was back in like 96 997 I guess um because at that time it was you know not accessible at all there were five labs in the world basically studying deep learning in any serious way and the only people who knew how to train a model that did anything useful were

12:55people who had been at those five Labs that was basically it um and we had some progress in making AI more accessible but only some because in the end we we were really hamstrung by this prerequisite we always had which is if you were going to learn deep lining from us you needed to have had at least a solid a very solid year of coding background you know like an intense year of coding background or a less intense few years of coding background and that rules out most of the world and so that then when I think about language models it's like well maybe that I'm sure that

13:32constraint gone now you know you can do a lot more with no code or not much code the idea of like being able to create useful things with code you can now use stuff like chat GPT and codex to um you know build things where it's doing most of the coding and you only need to understand enough to kind of tweak the logic here and there or paste in a string here or there or whatever so yeah I'm pretty enthusiastic about the ability of this technology to democratize access to this uh you know to coding and to AI um so that lots of people can create the tools that they want in

14:14their lives and that maybe didn't have the skills or expertise to do it before great um and something I want to hear from all of your about um your people have built or researched tools to make machine learning and AI more accessible as we've been talking around how do you think about your role or how have your personal goals changed with the Advent of large language models so maybe we could start with you Jeremy then go back around um I don't think the goals changed at all right so my wife Rachel and I did something quite out there when we started fast at AI which was we quit what we were doing

14:53um and decided we weren't going to try and make any more money uh because we wanted to focus 100% on the mission of of making AI more accessible and we felt like there was a huge danger that AI could result in a world where power is more centralized and that any kind of dependence on any external Financial stuff would make us you know we'd have conflicts um and got then that's happened you know like now we're all talking about it um so at the time nobody else seem to give a about centralization of power due to AI but now it's happening and a lot of people care which is great um I mean

15:42it's a shame it's happening but it's great people care um so the mission has actually suddenly become a lot more serious actually um because the rubbers hit the road and it's clear that dystopian outcomes could absolutely happen um but the way of achieving that mission probably needs to change a lot as well you know um I always created everything at fast AI on the basis that you could genuinely create worldclass models as good as anybody on your laptop or on collab in a day or two um and that's still true somewhat in some areas but it's not true if you want to create a gp4 competitor it's not going to happen uh

16:43or else if you want to create an an imet you know classifier competitor you know we out competed Google and Intel with their fleets of servers um but yeah we're not and that was on a on a a single machine but we can't do that anymore so yeah things are changing yeah and that's of course something you built software and obstruction layers for people to write code on on top of to facilitate that this is something I hope we get to more I've actually just put in the chat a link to your recent post AI safety in the age of dis Enlightenment um you did mention that having tools like chat

17:23GPT um Can democratize a lot of what's happening it democratizes a certain aspect of what's Happening of course something I hope we get to later is how having access to uh Hardware to your own um open source models being able to fine-tune them yourself and not relying on vendors where power is centralized um is something we really need to think a lot more concretely about yeah um increasingly the better models are being trained on the output of models you know so for example Dary 3 which is magnificent 99% I think of the captions are generated by AI um so we're kind of starting to enter into this positive

18:00feedback loop bit where AI is making AI better um which could be great um but it's a um it it creates natural monopolies and monopolies are terrible for society and the more power a monopoly has the more terrible for society it is so it's a bit of a worrying situation um and the more the better AI turns out to be the more problematic having it controlled by what's likely to be an AI Monopoly is going to be somewhere between like um economically a bit sad to uh uh an existential risk for Human Society yeah and I'm glad you gave voice to the positive feedback loops of play

18:55and something I hope we get to at some point is that the regulatory capture with which we're which we're seeing play out in in real time but before we get there I'd love love to um hear from from Shreya um how do you think about your role um changing with the Advent of llms um or how have your personal goals changed to what you're working on I don't know how I'm gonna follow Jeremy's answer I'm sorry I did this to you last time I yeah um I think for for me I feel like I'm I'm always focused on studying like the ML and now ai engineer um Persona so I feel

19:39like I I inevitably started studying it well because the people that I was already studying started to use it um so yeah I don't think I've like changed my goals higher level goals per se but I have started to notice um I'm talking to people from different backgrounds now um the problems that they're running into are a little bit you know interesting a little bit more different um people without the resources without like lots of labeled data are now you know using llms and so um I I think I'm just noticing a few of these things and I I don't quite you know have it all

20:23together in my head and maybe it's worth studying more deeply um but yeah my short answer is just I don't think anything has really changed for me I I have to get a PhD I have to graduate like still doing the things that that I've got to do I never got that absolutely and I I do think though you you already gave voice to you know this emerging role or emergent role of AI engineer and how how that may change maybe not necessarily the the basic I suppose capabilities we all have when when performing these Ro roles but perhaps the different types of skills and where the focus is uh may may

21:03actually change so um hammer was actually very intered on talking about your your paper revisiting prompt engineering via declarative crowdsourcing um and how that relates to the mergence of these new um Technologies and and roles so I I hope to get to that um haml you're you're the one who inspired this panel in a lot lot of ways in this this episode by telling telling me about your existential crisis um that changes on on a daily basis so you know what's what's happening with you man yeah it's a bit of a roller coaster sometimes like um so the perspective I'm dealing with so like there's many

21:46different perspectives so one perspective is as an ml engineer you know in Industry that's your career if your career is somehow tied to helping people um sort of operationalize ml or doing this quote ml engineering stuff um especially with llms it's you know the kind of it it is confusing to look out really far and try to think about what might happen I mean it's of course confusing for almost any profession but it is it seems somewhat acute in ml engineering like okay what do I FOC F on um you know should I with a limited amount of time like okay should I really F be focusing on getting

22:36into the guts of you know like deep learning uh you know should I be training model should I be more focused on something some other aspect um and then it keeps shifting really fast um you know so shifting in many ways so like the models the Technologies um the centralization of resources power and really um and so like what I think about is like okay how do I add the most value to society given my skill set and what should I like what should I do what should I think about and it's actually very confusing like one thought on one hand one thought that goes through my mind is like um you know well

23:19open AI you know automate way enough to truly democratize this in such a way where you know I should be thinking about something else and there's already some hints of that like I experienced in real time that my clients actually go through um you know like the people I work with a lot of them are using open Ai and it's like very sticky um and I'm I'm excited just like everybody else about open source models um in a way I kind of want nothing more than to uh get people off of Open Ai and sort of have their own open models but it's like I'm kind of

24:00swimming Upstream sometimes against that that that force that I can feel and it does I do think about that um but I think there's a lot of you know with that being said it's not all negative like you know there's this it just that's the nature of this like exponential crisis in a way um but then also there's a lot of ways that it can be exciting like you know there's definitely so I think one exciting part of AI and ml has always been the intersection of different fields like if you can be a domain expert in a certain field um and then bring ml to that field

24:37um that's always been an exciting part I I haven't drifted towards that yet but I think um there's even more opportunity for that now like being an expert in two fields and bridging them together very deeply and I so those are like some positive thoughts that I have is like okay if I don't know what to do maybe I will start drifting in that direction or concentrate in that direction um so honestly and also I don't know if I look too far out in the future if I try to think too far on the future that it's like you can it's kind of like staring into an abyss but so I tried to

25:11be a little bit more shortterm but yeah these are the thoughts that are going through my head and be fair the reason that open AI is so sticky is because the models are really good and definitely shape so which is great it's nice to have really good models that we can use definitely yeah and and yeah with like open AI Dev day and stuff like that like you know they're releasing more and more infrastructure around their models not just their models and and people are really eager to take that on um even if you know it may be it may not be as good as something else like you know there's

25:51it's very sticky but isn't that like like that that really surprised me how like you know how bad they are at creating these things around their models like first they tried plug that was a total failure then they tried gpts which aren't a total waste of time but the products really badly made I'm like anybody who thinks that they've achieved AGI internally would have to explain how it is that they've got some kind of super intelligence internally but they can't write a web app that's any good well this is actually really important because when ham and I first talked about the open AI developer day it was like wow this is mindblowing and

26:36then both of us went and used the stuff and was like oh no wait me too right yeah so why is it that you type text over here and then if you type text over here and then you hit this button then all that text disappears with no way to get it back it's like basic stuff I'm sure they'll fix these things but uh yeah they don't have ASI internally that's for sure absolutely and these products are subpar compared to the marketing but I think the original chat gbt release right was because he that wasn't the newest technology right as you well know but the product experience

27:13somehow was transformational with that first chat gbt right it did capture absolutely a lot of us which is because it it was a minimal interface on top of their model and they making models so yeah that you know there's lots of room to improve it um but that in itself was fantastic um so I do want to get a bit more into open a AI Dev day before that I want to kind of think about the role of the ml engineer and how it changes um now we have llms Shreya you actually had um a question for for Jeremy around this I don't know if you have it in front of

27:55you and want to want to ask it is as Tre is the one who knows all about this I don't know why you're asking me a question I don't think I'd know anything about this okay I think all of you know a lot about this not me don't deal with everyone has imposter syndrome at the moment to be clear even Ham's Ham's like no we shouldn't do a podcast because I don't know enough and I'm like dude this is wild so but yeah shre up perhaps you could ask Jeremy your your wonderful question yeah I I don't have it in front of me but I'll ask what I remember sending you

28:29um but Jeremy you worked you've you've helped to teach so many people like machine learning from like you know non-traditional backgrounds they didn't study CS in college for example um what skills or tools like maybe two or three do you see you know people needing to have now for an ml engineer role that you weren't you know teaching five years ago oh okay I can actually kind of answer that question um good thank you um so actually I've been shocked to discover that the answer is almost nothing's changed um and I know this for a few reasons so one is you know I recently redid both parts one and two of the

29:18course after a relaxing Global pandemic Break um and so coming back to it I had to look back over and be like oh what was I talking about two or three years ago um because everything's changed hasn't it it kind of felt that way but then I looked at it and I was like oh no all of these things are equally relevant nothing's changed and I think that's because the stuff I teach is I don't really teach tools you know I don't really teach you know uh high level application platform forms or whatever I teach fundamentals and it's fascinating how they haven't changed at all so you know

30:02I I um I think I've got a bit better at teaching some of it um so I've you know so I've changed like how I teach SGD is now in a much more interactive way with little sliders and things and it used to be like oh here's a picture from my book let me tell you about the picture um I when I teach about production iation even that has hasn't changed much like I do always show a tool and you know I used to use VOA now I use gradio but the fact is in each case it's like oh look you know here's a rest interface we

30:37create and like you know so we show with gradio like oh here's how you create a little JavaScript layer over the rest interface and do whatever you like with it um you know data loaders data sets batching yeah it's all basically the same and the amount of stuff that's actually changed is is pretty pretty minimal for people who are trying to build and train models the thing that's changed is how much you can now do without building and training botels you know like so I recently did a video called something like a hacker's guide to language models which is kind of a 90minut brain dump of all the things I

31:20figure people ought to know and most of it was not about training models at all you know a lot of it was just like here's how to use open AI models or even the training was like oh here's a script you can download called XEL aut and you just type XEL aut space data set name at the command line and it does it um so that stuff that's interesting because it's like it's a it's a whole different audience you know uh it's a whole different way of thinking about it so it's definitely making me think for all these people that maybe don't need to know really how to train

32:02models I still want to teach them how to uh take full advantage of AI and I think it yeah it needs a different approach and although you don't you I mean there is not a focus on training but you do make clear how important well trans we'll call it transfer learning but in the form of fine tuning or using r or or or these these types of things and I think I I can't recall quite correctly but you published a blog post recently or had a conversation called the end of fine tuning so I'm wondering if you could just expand upon that and tell us about

32:39the importance of transfer transfer learning as as a general concept oh yeah it's everything so yeah we didn't know that when Rachel and I started fast at AI we didn't know that um so when we started fast. AI it was totally speculative it was like is it possible to to not be a PhD graduate from one of the five deep learning labs and build something without access to Google brain compute Cloud that's useful and we didn't know the answer to that um but as we started looking into transfer learning that was like nobody was talking about um we were like holy why is nobody talking about

33:28this thing that actually makes the answer to this question yes you can do it um you know you can do it on a CO what coab didn't exist then but you can do it on you know a small AWS instance um or you can do it on your own computer with one Nvidia gaming card on it um and yeah it's basically okay so now you're taking advantage of some big company that has trained a big model on lots of data and lots of computers has been kind enough to release that to the world um and then it's learned like for vision it's learned like how to

34:08recognize objects and shadows and rotations and like that's the hard bit and if you know all that recognizing the difference between a a Golden Retriever and a poodle is easy so that's you know that's the transfer learning bit takes like a couple of minutes on One on One graphics card and uh yeah so then uh the big step I guess I'd then talk was to say like oh obviously that exact same process will work in language I thought it was really obvious so I asked a bunch of NLP people I knew um this is obviously going to work in language as well and they all said no

34:45definitely won't work language is totally different it's way harder there's no way anything that's happening in computer vision or work so for whatever reason I guess because I'm an Arrogant Bastard I just did it anyway um and it did work so ULM fit which you mentioned earlier is transfer learning to language um rather than pre-training on a million imag net images and figuring out which ones are cats and which ones are cricket bats you know I trained on Wikipedia and trained it to figure out the next word of a sentence but it's the same thing right the the basic model you end up with from that isn't any use to

35:24anybody um until you find tuna until you transfer do transfer learning on it and um the frustrating thing was that nobody realized even after ULM fit nobody realized that the phine tuning bit was the important bit [Music] um until you know so I kind felt fast yelling out into the void there and being like it's all about the fine tuning and it wasn't until chat GPT came out which was based on fine tuning of course instruction tuning and RF and everybody's like oh oh wow fine truning is pretty cool isn't it so you know it took a few years from the insight and the demonstration to

36:05people actually understanding what it was that was demonstrated absolutely so Hamil I want to one of the part of the Genesis for this conversation was you know this crisis around ml engineering and also you and I've chatted a lot about the emergence of this new role recently of the AI engineer so I'm wondering if you can speak to ml engineering and what whatever AI engineering is and give us your kind of nent thoughts on this Duality here yeah I mean I don't okay so um there's this term AI engineer I kind of understand what it is but um I don't really focus too much on that sort of popularized definition and

36:48just more focus on like what is my day-to-day life look like when I'm trying to work on llms and like what are the types of things I tend to focus on and the thing that I tend to focus on a lot is looking at lots of data and evaluating and thinking about evaluation and the reason that is is like if you so the key is it's actually something that also learned from Jeremy is like you need to iterate really fast U when you're when you're working on something your iterate your Loop has to be super fast if you change something try something new you should be able to get

37:22feedback within minutes or less or as fast as you can and if you don't have a good evaluation framework um it's really you can't do that and then also um that's the nature of llms if you want to fine-tune them or if you want to discover what your failure modes are you need to look at lots of data to understand what those are and um this is actually a core skill that a lot of data scientists and machine learning Engineers are really good at like that's a kind of a key ke kind of Cornerstone of like being a data scientist even before L learns and that's something that really transfers

38:03really well the things that are a little bit new are being you know I would say um being okay with more software engineering because now the now the like the ml is closer to whatever platform or software that you're working with so just to be really concrete here like in Tradition so like I still see with the people I work with um you know like about half the time I'm still able to use Python and like develop you know the language model component in like a microservice but then the other half of the time in that in the other half I would say is like growing more or I see like more of it or

38:48like you know can't really develop in this microservice and I have to like understand the entire application that might be written in another language like typescript for example um and so that's what I mean by like being more of you know getting more into software engineering that's kind of like let me let me just we've got a great question from Charles that was a question related to something Jeremy was saying but it's very relevant to this Charles Fri asked do the folks Jeremy mentioned who don't need to tr train models need to use Python uh JavaScript seems like a better choice potentially if you're consuming models via apis to

39:25deliver value at the application layer so how how does that type of thing relate to what you're talking about yeah so I mean the reality is is like microservices are okay like a lot of people you know people like really large companies espe Silicon Valley companies have been using those and certainly uh that's the thing but then a lot of people are not using microservices and a lot of modern applications are not built with python like there's no python anywhere in there and so if you want to glue uh like these large language models especially recalling apis to that application and reason about it um you

40:04have to start getting comfortable in these other languages um because it's not it's not you can't just like hand it off and like pretend like it's someone else's problem you have to kind of Tinker with okay how does this actually work and interact with the application that's very much part of the language model is like the the where the the L M meets the ux and reasoning about that you have to kind of think about that not not saying you have to be an expert but it is with the day-to-day kind of work I do find that uh I do find that to be the case so I do see like okay the the AI

40:42engineering definition as like that's popularized right now is like it's kind of a sort of you know it's like this JavaScript developer who learns like a bunch of language model model tools and apis and things like Lang chain and stuff like that and it's like gluing gluing stuff the only kind of little bit of uh the quibble I have with that is okay you need to look at lots of data and a lot of the best data tools are in Python and so that's like a force that's like pulling that skill into this AI engineer that's unavoidable you have to look at lots of data how are you going to do

41:21that um if you're coming from like a front- end development background or something like that I think that's you know I find that to be a blind spot that I constantly am seeing um like people rushing into like creating applications around around large language models without a good evaluation framework or without the like without really looking at any data um and so I think like these two skills like some like match I don't have like the best answer of like how to master all of them and just you know doing it one step at a time and like learning what I need to learn to to do a good job

41:55I guess I'm curious what yeah if anyone else has a perspective yeah well I'm also I'm also interested in with there are a lot of things happening here where we've got a democratization and we've talked about how you know people with low code skills and no code skills can can use large language models and that type of stuff on one side on the other side it's like you need to know some Python and now some application layer stuff and then Lang chain and being able to look at data and it's not only looking at tabular data now it's you know large amounts of unstructured data that we're

42:28not even certain how to version or evaluate so perhaps shreer um I know you think a lot about the democratization of of these skills and who can who can do these types of jobs so how do you square kind of these different different takes for yourself in terms of being able to do stuff with low code but also in some ways when you go one layer below suddenly you need to know all these different things and have all these hacker skills oh I love this question I also love that Palo mentioned um looking at the data as a core skill of ml engineers and what a

43:08lot of people don't realize is that that that looking at the data and building intuition for you know what does materially impact your ml system like changing your data um you know adding something to your output um putting assertions evaluation testing like these if you're a trained ml engineer you know to do these things but if you are not familiar with machine learning or if you know llms is your first fora into having some sort of software with machine learning in it then you are more inclined to you know add instructions to the prompt say like do not output something of this form um you're you're

43:53inclined to do a lot of prompt engineering and hope that at the end of the day the system is just going to work for all failure cases and I I it makes sense because people are always going to do what they're familiar with right like homo mentioned he is looking into like containers for um llms he's looking at different ways to like find tune um to go deep in that way but a lot of other people are spending all their time on llms and they still don't they don't do any training of models they don't do any fine-tuning of models everything is around you know how do we build better

44:26engineering tools um how do we eyeball responses like where do we involve humans in the loop um how do we write better eval specs that use GPT for um I I find this very very fascinating just you know the different ways that people are approaching the same problems um and it's not really clear if there is only one right way to do things everyone EV I I think looking at the data is a core skill set everybody ends up realizing like oh this is something that's important um but not everyone has you know strong opinions on whether they should be you know like using Jupiter notebooks to test their prompts or they

45:04should be serving you know their models with like rest apis or they should be able to use like python or JavaScript or whatever these kinds of questions I I feel like they're just the fact that there's like different right answers to them um means means maybe we shouldn't be looking at them as much as you know the questions that everybody unanimously you know arrives at that same conclusion absolutely and people coming from very different directions that's one of the beautiful things here people coming from a lot of different spaces and there are actually two questions in the chat that I that I want to get to before that I

45:36just Charles has also mentioned looking at data is geni genbutsu which I recognized but I had to look it up it's the Japanese principle of going and directly observing a location and its conditions in order to understand and solve any problems faster and more effectively um so I'm I'm thank you Charles I'm going to use use that time and time again I just wanted to mention something about what said um I think or kind of I'm I'm to hear your guys thoughts about this but um you know I I I agree it can be helpful to to dive into typescript or dive into all kinds of other languages

46:20or whatever um libraries um that kind of thing to scare the hell out of me you know I always like to just one language focus on that for years one web application framework one database back end one of everything um because that's all my brain can handle um and if I had to do something outside my kind of outside things that I've done in the last couple of years and was still kind of in my you know working memory uh I'd probably just skip it not do it do something else or if find somebody else who knew how to do it and get them to do it for me that's changed

47:02with chat GPT and codex and stuff because yeah I find I can um I can dive in now you know I can dive in and do some video processing with FFM Peg or like I don't know the other day I even wanted to just grab some stuff from my um Apple notes and chat gpts like oh here's a python script that uses open scripting architecture to attach to notes to enumerate that I just like okay I wouldn't have written that myself I don't know how to I used open scripting architecture like 25 years ago when I was playing with apple script but that's about it um yeah so I guess I feel like

47:49I kind of want to encourage people to actually not just think of themselves as backend python programmers or frontend JavaScript Engineers or something because actually um I find now yeah if if if chat GPT is there to help me get started on some different back end or library or platform or language I can I can look like I know what I'm doing pretty quickly um and I actually think that's really interesting because for like building a team or something I wonder now like is it better to have a five person team of really like quick Learners openminded kind of people who are really curious about everything versus a 50 person team

48:39of specialists in every area because you know I always hated when I was running companies hiring Specialists because there's all this friction that it creates and the more you create that friction the more you need more Specialists to help with that friction management and um then you suddenly need management hierarchies and stuff so I don't know wondering if there's different ways to do things now you don't necessarily need unicorns but maybe more people can be unicorns now I agree with that I you know like the only reason I'm able to dive into these things is is like with because of co-pilot chat GPT um basically everything I do with regards to code I

49:26am using those heavily like all the time especially now that we're up new knowledge cut off for April you know a lot of these tools didn't exist when it it was trained but now we can actually ask about them yeah I think it's a great Point like like whatever I said shouldn't be that scary it isn't that scary um like I you know like I'm able to dive into these other languages and like get lots of help and like you know with a lot of practice you can actually you know it takes some work so the one thing I want to say is like one thing that's really

50:02surprising is a lot of people aren't using large language models in their daily lives if someone ask me like what should I do to get better with large language models like a developer asked me one like interesting thing I like find quite often is like they're not really using large language models like in their daily lives like all the time um and that is is that's a something that you should change like if you're not using large langage try to use it like as much as you can also at least for now like the fundamental software engineering and computer science skills have suddenly become more important in

50:38my opinion like the fact that I know how to test for edge cases you know the fact that I understand race conditions the fact that I understand how to debug things like some suddenly these are actually the skills that matter you know whether or not I know the details of what parameters this function expects or what module to find it in or whatever you know doesn't matter anymore but like interestingly I find chat GPT you know gp4 incredibly bad at uh almost all coding you know like anything that it hasn't seen before in the training set just it just it's terrible and it's terrible at software

51:23engineering because it says like oh I've see there's a bug I'll fix it and it assumes it knows what the bug is it assumes it knows how to fix it and it'll even write a comment at the top saying like making it work by blah you know it's still broken it's like oh one final fix to do blah and it's still broken and so I find I have to tell it you know I'll say like the recommended way to debug a piece of code you know is to is to write just one or two lines at a time test the output and the types of what

51:51you expected and then don't move on until you've run that piece of code and like I have to teach it how to debug and I have to tell it like create tests that include these things but gee when I do that then I'm 10x faster because it's it's I could just say just do that thing um and it saves my limited brain capacity as well because now it's doing that kind of boring menial stuff but yeah I feel like the foundations I know are really dramatically more than they were a year ago the iterative cycle with chat jpt GPT 4 you've just mentioned Jeremy though is something you've spoken

52:32about publicly ham and I recently discussed it several times on SE several podcasts that the first time you ask at something usually isn't the best response and then you have a conversation with it and that I suppose that's analogous to the first machine learning model you train usually isn't the best model but then you introspect do some eror analysis specific going I was referring to the code interpreter that self reflects like it runs its code and its test doesn't pass or it gets a compilation error or whatever and it says I will fix it and it it never fixes it you know sometimes I let it run takes

53:09like 20 minutes come back and the chat's like got 12 attempts and eventually I'll just stop it um but when I use this trick of like teaching it basic software engineering principles I teach it the way of coding that that haml and I do in kind of notebooks because actually um chat GPT code interpreter is literally a DPR a notebook underneath so I I I teach it like I say like okay here's what I wanted to create now only think of the very first one or two lines of code write those check the outputs and types and stop and come back to me to see if I want to continue

53:55and when I do it that way it can actually solve things just like a human generally you can't just say like write this algorithm and it writes the you know the person writes 50 lines of code and it works that's not that's not how it happens be interesting to see if open AI ever figured this out and actually teach chat GPT software engineering because if they did it would be much more useful I think it's really important because really to write off like oh like you could just any like oh you don't need to learn how to use chat GPT just talking to a thing but no actually like

54:33what Jeremy is saying is like you know being very thoughtful and practicing what works and what doesn't makes it it's just like a completely different world and a lot of people don't do that yet to my surprise and so Jeremy and anyone else do you use system prompts to teach it in these ways and or do you see a future of creating your own gpts so I find system prompts aren't aren't nearly enough yeah it's like it's it's a bit specific to the context I I do want to try yeah I do want to try creating some gpts because I think with some you know kind of think about how to

55:12Cluster the different kinds of things that chap GPT is reg you know predictably bad at and predictably predictably can be taught how to do it well could make for some useful gpts but you know it's also something I feel like you know um maybe this is a better place for like open source models you know because open source models you've got a bit more flexibility and you know the fine tuning and Rag and stuff is a bit cheaper I me I don't know if you guys have looked at um open AI assistance which is the API thing behind gpts it's just insanely expensive like insanely expensive I don't understand

55:53what's going on there um it's not that good either so huh no it's not that good either no so yeah this feels like maybe the edge of you know there's some things open AI is really good at but this is yeah this is maybe not one of them is putting these things together into a into a more so now we're getting there product I would love your impressions and feelings from open AI developer day um and then perhaps how that changed when you started actually playing around with the tools um anyone can feel free to stop I mean at first I had good oh I didn't go I just like watched it um yeah

56:37yeah oh me neither I I didn't go either um I watched it as well um yeah I mean at first I was like oh wow they're getting into like more of like the infrastructure around language models like it seems like they're making a big play um I was like oh of course you know maybe it'll be good like it seems like you know they're putting a lot of yeah it seems like they might have some Advantage there um and then like you know as soon as I got access I played with it as much as I could for like an entire day and I was really disappointed

57:09like I struggled to make like for example gpts work the way at least that they were advertised and I thought wow this is just really bad um and then I had less of an extent crisis to a degree because I thought okay like okay there's like a lot of this the tremendous that white space is like definitely still there um and it's but yeah I mean that's what that's that was was going through my head Rya how about you I I still feel very optimistic about developer day um or just everything that was announced I found it very interesting like how strongly people reacted to it um so

58:05maybe some people have very very high expectations or you know they they um like a whole bunch of people were talking about how like startups are canceled now through due to developer dat and I'm like how how are they cancelled like they're all pre-seed or seed stage or whatever like it's not like they exist they don't have a product that everyone's using um but it it really reinforced in my my mind that like open AI is fundamentally building a consumer product um and so they're going to throw a hundred things at the wall and probably only one of them will stick because that is how consumer products

58:41work and we are the consumers um and it's it's not like you know somebody's like building a technology that organizations that like Enterprises are using and you know they get like $1 million contract and like they're happy um like we're we're the consumers that we're like trying everything and not everything will stick and I I think that's like pretty par for the course um with gpts I at first I was like okay what could I possibly do with a GPT I I don't know what I what I need a GPT for that like chat GPT doesn't already do for me but I found one case one use case I

59:18spent like six hours thinking of use cases and building different things um and this is around making figures for my talks or my slides um and I think it's it's because I have a clear process um like I have a style that I kind of want to make figures in like I want to make it a little bit simple maybe black and white a little cartoony cute and I want that to like persist and I need to make many many figures um I need to make like a figure for every few or visual for every few slides and I know that there's like a broader theme of the

59:56a figure yeah or just like a visual so like an specific like you're talking about like a neural network diagram or a picture of a neural network di diagram could be one of them um I I gave a talk at a lab this morning um and the title figure hold on let me pull it up title slide figure is an I should I share my screen I feel like that's the best way to like show this yeah okay great I'm just going to give you privileges to do so okay you got it got it okay so like like I want cohesive figures or visuals whatever you want to

1:00:45call it and like the best way to do that was a custom GPT because it remembers like the style and all of the things you know wait how does remember the style and all the things this I put it on on the system prompt okay so it doesn't exactly know like you couldn't it's not that you can have exactly the same character appear each time but just that the the general it it knows like the pattern of the queries that I have so my queries are like my slightest title today's road map and you know I want like a long winding road but not too windy and I want it to go top to

1:01:22bottom or left to right I don't know something like that um yeah and and I felt so happy when I got it yeah beautiful I love it thank you it's really nice this is really cool this this is something that like say I wanted to do without a GPT or whatever I kind of would have my stable diffusion prompt um and I'd have to give it the context every single time do you have one of those SL dicks online somewhere that you could share a link I'd love to for sure I'll I'll definitely post it somewhere thank you um yeah but it just it goes to a point

1:02:00of you I at the end of the day I just keep paying attention to the emotions that all of us feel when all these new things come out that's the whole point of a developer day like it's to release tooling that get us all really hyped or like really upset or really demotivated or like in an existential crisis um like that's that's the way that all of this works so yeah I P like I've been trying to decide whether I'm ready to make a call on on on open Ai and whether whether we should start to whether we should say ngmi they're not going to make it

1:02:39because that's my feeling after seeing their recent stuff is I'm feeling like they're not going to make it because like um the basic premise here is like you know exponential growth in capabilities and uh no they're hitting the wall they're closing up signups to chat GPD plus they're you know they're tweaking you know like their new product is the same product as for as before but a bit cheaper and faster and it actually seems a little bit crappier [Music] um they're not releasing stuff for like a year because they can't get enough gpus to run it on um this is not feeling like a f fast

1:03:32take off of super intelligence you know what I mean this is feeling like oh we found a quite useful really quite useful technology that um uh quite quickly surprised a lot of people and how capable it was there was then an assumption that that would result in a rapid feedback loop of improvement um which I kind of expected as well but the thing I had I think not thought enough about is [Music] um that improving capabilities without increasing energy and other resource use is not obvious it's not necessarily clear that that's possible you know um and so we kind of seem to be or at least

1:04:33open I kind of seem to be hitting a bit of a wall here which I think is interesting and so yeah I mean I'm sure they they're going to do the 10 billion training run at some point and then maybe after that they'll do the hundred billion do training run but then I think well what what then right because now you've got like this API that's going to cost like $3 per token or something you know and and two hours later you get two sentences um that are like 30% better than the ones that it would give you now so I've been finding this quite interesting and

1:05:11also yeah also the fact that chat GPT was basically an accident you know it was like oh it's try just chucking something out there to help like people understand a little bit about what gp4 is going to be maybe a handful of people will use it and you know accidentally that was really useful and now they're kind of like oh you know and then Microsoft's like oh you guys create useful products so we would you know that Microsoft's now in basically in control right because open AI is entirely dependent on Microsoft and Microsoft gets 75% of the profits [Music] um so so they got to keep Jing

1:05:56out useful products to Microsoft and it seems they're not actually that good at Jing out useful products you know um so I don't know I think it's really interesting to me I you know this the open AI the result of both open AI Dev day and just their overall trajectory since releasing gp4 is like I'm yeah becoming more and more bearish on on on their future I guess kind of feels like be you know they might I could I could imagine them being the next Google uh which is not bad you know to be sure or off a bit but uh yeah not sure they're going to go past that you

1:06:46know becoming slightly bearish on open AI does that make you more bullish on open source models not sure um I don't know how I feel about open source bottles I mean I I like open source bottles and I want them to be successful [Music] um uh but I I think open AI is very good at what they do you know they're good at building models and they're good at serving models um but you know I think the thing is like if you look at anthropics you know leaked pitch deck where they said like oh whoever's ahead in 2025 2026 they're they're going to win so that's all that matters um yeah

1:07:31I'm not sure that's true actually I used to think that's reasonably likely had a reasonable chance of being true now I don't think it's going to be true um so that does mean that the open source folks are going to have I think time to catch up you know obviously if there was this exponential growth thing going on then that's then that's it for open source unless States get involved and either forcibly slow down the big commercial companies or massively inject something in kind of uh academic collaborations like you know a CERN style project open source was going to lose that would be awesome yeah let's just clear about what

1:08:08anthropic is is and correct I'm going to State some recent historical facts that may I probably need to fact check so you can fact check me um anthropic was formed by a bunch of people who were high up at open AI and perhaps were um expressed dissatisfaction with the money open AI was taking and that it was becoming more closed well they were pissed off about the Codex thing particularly yeah okay they thought that was IR responsible so my understanding is that was the forcing function where they were like okay that's too much we're out that makes sense and then they left to form anthropic but of course after

1:08:46that they recognized that need they needed pretty serious injections of capital and took maybe $4 billion from Amazon or something like that yeah this is what happens to every do gooder company in the world at some point the stakeholders are like oh whatever we're doing we could do it better if only we had more money and so then they start bringing in people who are good at making money and then they become controlled by those people or they themselves get focused on making money because this is a capitalist Society you know so yeah the exact thing that anthropic left because they hated it they've now done the same

1:09:24thing they've released a consumer product they have a sales team they have uh their employees get Equity that is obviously pointless unless they're making it a making a profit and and have some way of making that into liquidity um and every everybody who's got any background in business at all knew that that all of those things were going to happen because of course they do so in your mind what are the odds that one company like like open AI or anthropic create a monopoly capture the entire Market or an oligopoly like a handful of them do either through acquiring increasing resources or regulatory capture which we haven't

1:10:06really talked about yet but I'd be interested in your thoughts on that as well just totally depends right putting aside regulatory capture if these positive feedback loops are a thing and there are a thing even given resource constraints and energy constraints and so forth so if companies are actually able to continually improve capabilities at a growing rate without having an accompanying growing rate in cost um then that there has it basically has to be a monopoly you know whoever's ahead gets ahead by further and further and so why would anybody use the alternative and so they end up with you know all the power in the world because

1:10:50they own the thing which there's like responsible for all the value in the world on the other hand if that as as I think seems likely now if that uh isn't going to happen then there's no particular reason there would be any well I think the the the uh economics look similar to Google and Google is a monopoly to be clear but they only get to stay a monopoly if they keep being reasonably good at what they do you know um there's nothing fundamentally making it impossible for that give they became just crap as being a search engine and somebody else can just come along and be a better search

1:11:31engine they do have a pretty serious Monopoly on search but I suppose on the advertising Market you can think of Facebook or meta and Google I mean they don't have a um it's it's not impossible for others to compete with them um if they stopped being good at what they were doing you know they actually have to keep being a good search engine in order to maintain that Monopoly you know um dto to some extent for Amazon like they've got a lot of monopol practices and they're actually being you know um taken to court I guess by the FTC uh but nonetheless they like they if if they

1:12:14were not genuinely a good place to buy stuff from then people will stop buying stuff from them and in fact in Australia they're not a particularly good place to buy stuff from you know not not that many people use it in Australia um can you tell us a bit about regulatory capture and then I want to go on to kind of talking about how to think about whether to use open source models or vendor based models and that type of stuff but particularly with everything that's happened recently in the US and and the EU I'd have to fire myself for not asking you about regulatory capture

1:12:45I mean it's a great time to talk about regulatory capture because everything is changing so if you had asked me last week I would have been pretty dismal on it it was you know there was this massive shift in the Overton window where suddenly all these World governments and world leaders you know Sam mman was going out doing you know State visits and stuff and um the you know overarching theme I kept hearing from from from governments was uh oh wow you know existential risk is such a problem have to care a lot about AI safety and the way you know we should be doing it

1:13:27like you look at the EU AI act totally changed from earlier versions to be like Oh Well everybody's going to have to submit their model runs ahead of time and have them approved to prove that they're safe you know before they train their model previous versions correct me if I'm wrong were about more like deployment Andis applications yes they they were like if you're going to use AI to decide which insurance claims get paid out then these are all these things you have to do first which makes perfect sense or else if it's like you're not allowed to train a model until you can prove that model can't be used

1:14:10to send a blackmail email like obviously that's impossible because it's a general purpose Computing device so you know things are things you know and that so the regulatory capture there was basically because it's impossible to prove a model is safe and because new competitors couldn't come in until they could prove their models are safe and the existing models were all assumed to be safe because they were def Frontier models were defined as being anything bigger than the current biggest ones it was just totally locking in the entrenched players you know it's fullon 100% regulatory capture um so then interestingly it's all changed in the last few days um basically with in

1:14:53basically in Europe um a few companies in Europe told their representatives this act is going to destroy us and we are the we are the biggest hope in this country to compete with open Ai and anthropic and stuff so if you do this we're dead um so like that was mistol in France it was what's called like Alf Alf or whatever in Germany somebody in Italy was doing something simil I don't know it was but so then their governments came out and said all right we're we're not going to do anything that threatens open source and then the UK who had been the most like into you know Rishi sumac was

1:15:38like all over this existential risk stuff and yeah one of his M ministers came out yesterday and said not we're not going to do anything that's going to threaten open source um so yeah you know it's been this fascinating reaction from politicians of being like oh maybe if we yeah just do all this stuff that the corporate lobbying tells us to do American corporate lobbying tells us to do and in the process destroy the ability of our local companies to compete maybe that's not a good idea so uh and to be seen also a similar thing like suddenly even in VC you know bunch of VCS a couple of

1:16:20days ago all signed this thing saying oh we're to make sure that yeah we're going to make sure that all of our um portfolio companies Ure safety again can't be done so now there are literally Founders putting up screenshots on Twitter of their emails to VC saying I'm pulling out of this round that you're going to do because you signed this um and my I've heard through the great fine that VCS the VCS that signed it are very surprised at I don't think they' thought it through uh but to me I'm just like what were you thinking you know who who decides on behalf of their

1:17:02portfolio companies what their portfolio companies are required to do the VCS they don't understand this as well as their portfolio companies do so why are they telling their portfolio companies what to do it's it's Madness so you know I think uh maybe a bit of Common Sense is starting to appear fingers crossed absolutely so bringing this back to a conversation of what people who are listening and watching can can do to get started with all of these te Technologies I think this you know difference between vendor based stuff and open source models is very interesting of course you mentioned your um hackers guide to large language

1:17:41models on on YouTube recently where you start with um using chat GPT GPT 4 because it's the easiest way to get started but then show people how to how to get up and running with um open source models as well so wonder if Shreya and Hamill what thoughts you have on um the trade-offs between developers using um open source models and closed models um and situations in which it would make sense to use one or the

1:18:12other um I can go first um I actually see them working together very well so like a lot of people who are using GPT you know GPT 4 gpt3 .5 uh and they get started this way um if you're logging your data in a sensible manner like you're you're you're logging your traces from your language model invocations that is a rich source of data that then you can use the to fine-tune and you can like the same data you can use to fine-tune open AI models with you can use that same data to fine tune fine tune open models there's nothing kind of there's nothing preventing you from doing that except

1:18:57some people are a little bit skittish around the terms of service but I find the terms of service fear to be a little bit um I mean so I'm not a lawyer but I find I find the terms of service like the fear of terms of service to not make some that much logical sense because um so the so the terms of service go something like okay you can't use the data that's generated by GPT for to to train a competing model now if you really think about that um what does that mean so that means like and if you have function calling and Rag and all of

1:19:38this stuff uh in your product um basically like any any data generating process that your language model touches is now quote infected by this uh gp4 terms of service and then like if you really think about it carefully like if you have that view of the terms of service maybe you shouldn't be using open AI because like none of your data is yours because there's so many things like when you have Rag and function calling and all this stuff there's so many data generating processes that to think that you will never be able to train your model based on anything it touches that doesn't make doesn't that seems like a really

1:20:23difficult position that even open AI wouldn't want um and so so anyways like this data can be used and is actually what a lot of people are doing uh successfully is like bootstrapping themselves this way now like yes you don't want to train a foundation model that competes with open AI but that's not what most people are trying to do they're just trying to like kind of you know use hybrid between both um or you know use them together and there's a lot of like reasons why people want to use these things together for example for evaluation uh you know you want to use a lot of times the most powerful model you

1:21:02can for certain tasks um and yeah there's like and then also like synthetic data generation and things like that a lot of times you want to use a powerful model there's a lot of like things where you can use them together so I see I see actually that they complement each other if you think about it correctly in a lot of ways and Hamil's being over humble again he he may not be a lawyer but he is trained as a lawyer so it's not he's not totally speaking out of his ass here I mean I guess yeah like um and again it's not legal advice is like if

1:21:36the basic like Common Sense principle is like if a company if someone has like some Rule and that rule a certain reading of that rule doesn't make sense for that company well that's probably not a good reading of that rule because they're the ones who made the rule and so um yeah do I just want to say that so yeah I'll give it over to sha here's what she has say I also think they work hand inand I mean anecdotally gp4 is the best llm we have today so I think a lot of people get value out of trying gp4 to just see what performance you know like what is

1:22:18the best performance you could hope for your test that involves llms um and that's not good enough then you know it you may need to wait a little bit more or like think about how you want to do your data engineering before putting stuff in your prompt um but like just having that kind of ceiling performance or a notion of a ceiling performance can help figure out what you want to do next um the other thing is I mean it's kind of crazy but I I think that open AI has some of the best like API ergonomics for these llms like to get started with llama you need to figure out or an open

1:22:56source model you need to figure out okay like I should probably serve it somewhere and then I need to like talk to that server and this also costs money or I need to pay someone else who's serving llama um and then it kind of feels the same um so yeah I I I feel like I feel like it's one of those things where there's no real right answer and people end up trying a bunch of different models uh what think is really interesting is sometimes people will try a model or try developing their llm chain or Pipeline on top of some model and they would have overfit so

1:23:31hard to that their prompt will overfit so much to that model that when they try it like off the bat on some other model they might be like oh this performance is so bad in comparison I'm not going to consider using an open source model um but a lot of that is just because they spent so much time proms engineering on like GPT 3.5 turbo and well well yeah yeah that that does make sense um like you've added all of these special instructions to your prompt to like kind of figure out the failure modes um or cut those cases yeah so I'm excited for them to all kind of

1:24:07progress at the same to get to the same level absolutely we see that happening in parallel and converging and intersecting I I do wonder if there's a question as to what the term open source even we use open source when we talk about models to mean generally like the weights are are released but that doesn't necessarily mean the code is released or even what the what the training data is right so seems like open source is kind of even potentially a multi-dimensional space now right well the pro I mean the the main problem is the license under which the weights are released so generally the code is available under like a real

1:24:45open source license like Apache but the weights generally come with a license agreement which for example in the case of the E34 B model which is currently the best open source model or bestas model is a no commercial Source sorry no commercial use license um llama 2 comes with a license that does not let you use llama 2 Data to train models other than llama models and also requires you to not use it at all if you're a giant company um so yeah um uh I 2 which is was just released by Microsoft is a no commercial use license yeah um and you know the CEO of Microsoft got

1:25:34up on stage and said we're releasing a new open source model F2 so I use those words but it's it's not open source under any normal definition like in the sense of like commercial you know no commercial use um so there's very few models that are available under like do what you want license La llama 2 I guess is the closest we have so we're gonna have to wrap up in a couple of minutes I I do have a question because I have you know a group of people here who have deep expertise in in this and this is probably a conversation you get them on then

1:26:23yeah I I do think we probably need to have a dayong workshop on this or something but um you know we're all very interested in literate programming and and notebooks to a variety of of extents um how do you think tools like notebooks are affected by llms and how might literate and exploratory programming need to adapt to incorporate um everything that's happening in generative Ai and llms at the moment in 30 seconds Jeremy I think H should start on this one he's been thinking about it more recently than I have I suspect you're muted though haml so it won't work so well sorry um yeah so you

1:27:07know the tools that incorporate llm llm stuff um you know has tended you know has been developed with like IDs in mind first and so um that's one struggle that I feel is like you know kind of like literate programming tools like the ones that that you know Jeremy made like MB Dev okay like I can't use co-pilot as fluidly inside you know notebooks as I can you know in Ides and things like that and I think um I think that Paradigm is really important to bring into things like notebooks as Jeremy said like you know Chad GPT is a notebook and it's like a very native

1:27:52environment to have conversations with not only yourself but with an AI or with anything else so I think it does need to evolve um currently we're at like this point where there's a lot of tension between that because like like even as a person that loves these environments like I'm being pulled away in cases where I want to Avail myself of like LMS in my programming um you know what like cat GPT is a notebook disconnected from the rest of the world just like Jupiter notebooks were pre-bd you know so pre-bd notebooks are a thing you could only basically do like some prototyping or analysis in they were disconnected

1:28:37from the rest of the world no tests no libraries no continuous integration no anything it was its own little world um so yeah chat GPT now feels like that to me which really limits its utility I agree I think it's a big opportunity to to Mel these worlds yeah cool well I I've got a final final question I want to keep it pretty vague but I'm wondering maybe we can start with shre for anyone working in in this space or getting started in this space what's a what's a call to action something that that you'd like to see see more people do oh man H H already said it earlier

1:29:27but just figure out how to use chbt as much as you can in your daily life in a way that's like sustainable so not like try to ask chat PT for everything um but like genuinely well one thing I did um I want to say like in January or February of this year was okay for two weeks I'm going to just try to do everything in chat GPT and wow did that help me on understand what it could be useful for what its failure modes were um yeah with with an open mind honestly awesome how about you haml uh just to add on that look at lot

1:30:08look at data look at lots of data don't just like Vibe check stuff also like try to measure come up with metrics on like if your system is working or not and look at lots of data and if you find that looking at data is painful some small tools for yourself to help you look at that data and I'm talking about the LM data specifically awesome and Jeremy watch my hecka guide to language models and try out some of the things that I show um and also when it comes to trying out making your own tools try making them in like some of these super fast and easy things like gradio you

1:30:48know or um salana the new kind of w thing um MH yeah um I there was a so when I was working with jonno Whitaker on a recent kle competition and llm kle competition you know I thought he he showed a great example he's actually got a video of it maybe we could link to where he was like oh I want to understand this data better so I'll spend 15 minutes or half an hour I don't know exactly creating a little gradio tool you know that's something that he just whipped up in a few minutes to use it for a few hours and I I like that

1:31:23idea you know the idea that we can build particularly now if you use chat GPT you can whip up a a quick uh look through the data tool you know Mark a flag a few duplicate records stuff like that you can create it in a few minutes yeah great and actually um we'll include that link to jono's um video in the show notes Jonah has been here um saying nice things and commenting as well I've already linked to uh your um hackers guide to llm here as well and told people to check it out um and I'll just to come and to troll Hamill as much as possible and I don't

1:32:03see any sign of that happening so he's let me there hasn't been much he he'll troll me afterwards for sure yeah um I'll I'll also just add I do think advice to you know interact with these LMS and chat gbt as much as possible um haml a few weeks ago suggested that I um you know when I when I go for walks talk with chat gbt um you know because it's got whisper and all of that to do um speech to text to text to text to text to speech um now and so chatting with chat gbt while going for for long walks has been incredibly insightful

1:32:39into what is possible um what's possible there and what isn't um so it's time to wrap up but I'd love to thank everyone for for joining um and I'd also love to thank all of you um Shrea Hamil and Jeremy not only for your deep expertise but but your time and um I always enjoy chatting chatting with you all and it was just such a wonderful conversation so thank you all you you go and thank you for doing your best to make me look good which is always a challenge you're too humble you all you all look great all right thanks youo and seeing um