AI Magic: Shipping 1000s of successful products with no managers and a team of 12 — Jeremy Howard of Answer.ai
Disclaimer: We recorded this episode ~1.5 months ago, timing for the FastHTML release. It then got bottlenecked by Llama3.1, Winds of AI Winter, and SAM2 episodes, so we’re a little late. Since then FastHTML was released, swyx is building an app in it for AINews, and Anthropic has also released their prompt caching API.
Remember when Dylan Patel of SemiAnalysis coined the GPU Rich vs GPU Poor war? (if not, see our pod with him). The idea was that if you’re GPU poor you shouldn’t waste your time trying to solve GPU rich problems (i.e. pre-training large models) and are better off working on fine-tuning, optimized inference, etc. Jeremy Howard (see our “End of Finetuning” episode to catchup on his background) and Eric Ries founded Answer.AI to do exactly that: “Practical AI R&D”, which is very in-line with the GPU poor needs. For example, one of their first releases was a system based on FSDP + QLoRA that let anyone train a 70B model on two NVIDIA 4090s. Since then, they have come out with a long list of super useful projects (in no particular order, and non-exhaustive):
* FSDP QDoRA: this is just as memory efficient and scalable as FSDP/QLoRA, and critically is also as accurate for continued pre-training as full weight training.
* Cold Compress: a KV cache compression toolkit that lets you scale sequence length without impacting speed.
* colbert-small: state of the art retriever at only 33M params
* JaColBERTv2.5: a new state-of-the-art retrievers on all Japanese benchmarks.
* gpu.cpp: portable GPU compute for C++ with WebGPU.
* Claudette: a better Anthropic API SDK.
They also recently released FastHTML, a new way to create modern interactive web apps. Jeremy recently released a 1 hour “Getting started” tutorial on YouTube; while this isn’t AI related per se, but it’s close to home for any AI Engineer who are looking to iterate quickly on new products:
In this episode we broke down 1) how they recruit 2) how they organize what to research 3) and how the community comes together.
At the end, Jeremy gave us a sneak peek at something new that he’s working on that he calls dialogue engineering:
So I've created a new approach. It's not called prompt engineering. I'm creating a system for doing dialogue engineering. It's currently called AI magic. I'm doing most of my work in this system and it's making me much more productive than I was before I used it.
He explains it a bit more ~44:53 in the pod, but we’ll just have to wait for the public release to figure out exactly what he means.
Timestamps
* [00:00:00] Intro by Suno AI
* [00:03:02] Continuous Pre-Training is Here
* [00:06:07] Schedule-Free Optimizers and Learning Rate Schedules
* [00:07:08] Governance and Structural Issues within OpenAI and Other AI Labs
* [00:13:01] How Answer.ai works
* [00:23:40] How to Recruit Productive Researchers
* [00:27:45] Building a new BERT
* [00:31:57] FSDP, QLoRA, and QDoRA: Innovations in Fine-Tuning Large Models
* [00:36:36] Research and Development on Model Inference Optimization
* [00:39:49] FastHTML for Web Application Development
* [00:46:53] AI Magic & Dialogue Engineering
* [00:52:19] AI wishlist & predictions
Show Notes
* Jeremy Howard
* Previously on Latent Space: The End of Finetuning, NeurIPS Startups
* Answer.ai
* Fast.ai
* FastHTML
* answerai-colbert-small-v1
* gpu.cpp
* Eric Ries
* Aaron DeFazio
* Yi Tai
* Less Wright
* Benjamin Warner
* Benjamin Clavié
* Jono Whitaker
* Austin Huang
* Eric Gilliam
* Tim Dettmers
* Colin Raffel
* Mark Saroufim
* Sebastian Raschka
* Carson Gross
* Simon Willison
* Sepp Hochreiter
* Llama3.1 episode
* Snowflake Arctic
* Ranger Optimizer
* Gemma.cpp
* HTMX
* UL2
* BERT
* DeBERTa
* Efficient finetuning of Llama 3 with FSDP QDoRA
* xLSTM
Transcript
Alessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.
Swyx [00:00:14]: And today we're back with Jeremy Howard, I think your third appearance on Latent Space. Welcome.
Jeremy [00:00:19]: Wait, third? Second?
Swyx [00:00:21]: Well, I grabbed you at NeurIPS.
Jeremy [00:00:23]: I see.
Swyx [00:00:24]: Very fun, standing outside street episode.
Jeremy [00:00:27]: I never heard that, by the way. You've got to send me a link. I've got to hear what it sounded like.
Swyx [00:00:30]: Yeah. Yeah, it's a NeurIPS podcast.
Alessio [00:00:32]: I think the two episodes are six hours, so there's plenty to listen, we'll make sure to send it over.
Swyx [00:00:37]: Yeah, we're trying this thing where at the major ML conferences, we, you know, do a little audio tour of, give people a sense of what it's like. But the last time you were on, you declared the end of fine tuning. I hope that I sort of editorialized the title a little bit, and I know you were slightly uncomfortable with it, but you just own it anyway. I think you're very good at the hot takes. And we were just discussing in our pre-show that it's really happening, that the continued pre-training is really happening.
Jeremy [00:01:02]: Yeah, absolutely. I think people are starting to understand that treating the three ULM FIT steps of like pre-training, you know, and then the kind of like what people now call instruction tuning, and then, I don't know if we've got a general term for this, DPO, RLHFE step, you know, or the task training, they're not actually as separate as we originally suggested they were in our paper, and when you treat it more as a continuum, and that you make sure that you have, you know, more of kind of the original data set incorporated into the later stages, and that, you know, we've also seen with LLAMA3, this idea that those later stages can be done for a lot longer. These are all of the things I was kind of trying to describe there. It wasn't the end of fine tuning, but more that we should treat it as a continuum, and we should have much higher expectations of how much you can do with an already trained model. You can really add a lot of behavior to it, you can change its behavior, you can do a lot. So a lot of our research has been around trying to figure out how to modify the model by a larger amount rather than starting from random weights, because I get very offended at the idea of starting from random weights.
Swyx [00:02:14]: Yeah, I saw that in ICLR in Vienna, there was an outstanding paper about starting transformers from data-driven piers. I don't know if you saw that one, they called it sort of never trained from scratch, and I think it was kind of rebelling against like the sort of random initialization.
Jeremy [00:02:28]: Yeah, I've, you know, that's been our kind of continuous message since we started Fast AI, is if you're training for random weights, you better have a really good reason, you know, because it seems so unlikely to me that nobody has ever trained on data that has any similarity whatsoever to the general class of data you're working with, and that's the only situation in which I think starting from random weights makes sense.
Swyx [00:02:51]: The other trends since our last pod that I would point people to is I'm seeing a rise in multi-phase pre-training. So Snowflake released a large model called Snowflake Arctic, where they detailed three phases of training where they had like a different mixture of like, there was like 75% web in the first instance, and then they reduced the percentage of the web text by 10% each time and increased the amount of code in each phase. And I feel like multi-phase is being called out in papers more. I feel like it's always been a thing, like changing data mix is not something new, but calling it a distinct phase is new, and I wonder if there's something that you're seeing
Jeremy [00:03:32]: on your end. Well, so they're getting there, right? So the point at which they're doing proper continued pre-training is the point at which that becomes a continuum rather than a phase. So the only difference with what I was describing last time is to say like, oh, there's a function or whatever, which is happening every batch. It's not a huge difference. You know, I always used to get offended when people had learning rates that like jumped. And so one of the things I started doing early on in Fast.ai was to say to people like, no, you should actually have your learning rate schedule should be a function, not a list of numbers. So now I'm trying to give the same idea about training mix.
Swyx [00:04:07]: There's been pretty public work from Meta on schedule-free optimizers. I don't know if you've been following Aaron DeFazio and what he's doing, just because you mentioned learning rate schedules, you know, what if you didn't have a schedule?
Jeremy [00:04:18]: I don't care very much, honestly. I don't think that schedule-free optimizer is that exciting. It's fine. We've had non-scheduled optimizers for ages, like Less Wright, who's now at Meta, who was part of the Fast.ai community there, created something called the Ranger optimizer. I actually like having more hyperparameters. You know, as soon as you say schedule-free, then like, well, now I don't get to choose. And there isn't really a mathematically correct way of, like, I actually try to schedule more parameters rather than less. So like, I like scheduling my epsilon in my atom, for example. I schedule all the things. But then the other thing we always did with the Fast.ai library was make it so you don't have to set any schedules. So Fast.ai always supported, like, you didn't even have to pass a learning rate. Like, it would always just try to have good defaults and do the right thing. But to me, I like to have more parameters I can play with if I want to, but you don't have to.
Alessio [00:05:08]: And then the more less technical side, I guess, of your issue, I guess, with the market was some of the large research labs taking all this innovation kind of behind closed doors and whether or not that's good, which it isn't. And now we could maybe make it more available to people. And then a month after we released the episode, there was the whole Sam Altman drama and like all the OpenAI governance issues. And maybe people started to think more, okay, what happens if some of these kind of labs, you know, start to break from within, so to speak? And the alignment of the humans is probably going to fall before the alignment of the models. So I'm curious, like, if you have any new thoughts and maybe we can also tie in some of the way that we've been building Answer as like a public benefit corp and some of those aspects.
Jeremy [00:05:51]: Sure. So, yeah, I mean, it was kind of uncomfortable because two days before Altman got fired, I did a small public video interview in which I said, I'm quite sure that OpenAI's current governance structure can't continue and that it was definitely going to fall apart. And then it fell apart two days later and a bunch of people were like, what did you know, Jeremy?
Alessio [00:06:13]: What did Jeremy see?
Jeremy [00:06:15]: I didn't see anything. It's just obviously true. Yeah. So my friend Eric Ries and I spoke a lot before that about, you know, Eric's, I think probably most people would agree, the top expert in the world on startup and AI governance. And you know, we could both clearly see that this didn't make sense to have like a so-called non-profit where then there are people working at a company, a commercial company that's owned by or controlled nominally by the non-profit, where the people in the company are being given the equivalent of stock options, like everybody there was working there with expecting to make money largely from their equity. So the idea that then a board could exercise control by saying like, oh, we're worried about safety issues and so we're going to do something that decreases the profit of the company, when every stakeholder in the company, their remuneration pretty much is tied to their profit, it obviously couldn't work. So I mean, that was a huge oversight there by someone. I guess part of the problem is that the kind of people who work at non-profits and in this case the board, you know, who are kind of academics and, you know, people who are kind of true believers. I think it's hard for them to realize that 99.999% of the world is driven very heavily by money, especially huge amounts of money. So yeah, Eric and I had been talking for a long time before that about what could be done differently, because also companies are sociopathic by design and so the alignment problem as it relates to companies has not been solved. Like, companies become huge, they devour their founders, they devour their communities and they do things where even the CEOs, you know, often of big companies tell me like, I wish our company didn't do that thing. You know, I know that if I didn't do it, then I would just get fired and the board would put in somebody else and the board knows if they don't do it, then their shareholders can sue them because they're not maximizing profitability or whatever. So what Eric's spent a lot of time doing is trying to think about how do we make companies less sociopathic, you know, how to, or more, you know, maybe a better way to think of it is like, how do we make it so that the founders of companies can ensure that their companies continue to actually do the things they want them to do? You know, when we started a company, hey, we very explicitly decided we got to start a company, not a academic lab, not a nonprofit, you know, we created a Delaware Seacorp, you know, the most company kind of company. But when we did so, we told everybody, you know, including our first investors, which was you Alessio. They sound great. We are going to run this company on the basis of maximizing long-term value. And in fact, so when we did our second round, which was an angel round, we had everybody invest through a long-term SPV, which we set up where everybody had to agree to vote in line with long-term value principles. So like never enough just to say to people, okay, we're trying to create long-term value here for society as well as for ourselves and everybody's like, oh, yeah, yeah, I totally agree with that. But when it comes to like, okay, well, here's a specific decision we have to make, which will not maximize short-term value, people suddenly change their mind. So you know, it has to be written into the legal documents of everybody so that no question that that's the way the company has to be managed. So then you mentioned the PBC aspect, Public Benefit Corporation, which I never quite understood previously. And turns out it's incredibly simple, like it took, you know, like one paragraph added to our corporate documents to become a PBC. It was cheap, it was easy, but it's got this huge benefit, which is if you're not a public benefit corporation, then somebody can come along and offer to buy you with a stated description of like turning your company into the thing you most hate, right? And if they offer you more than the market value of your company and you don't accept it, then you are not necessarily meeting the kind of your fiduciary responsibilities. So the way like Eric always described it to me is like, if Philip Morris came along and said that you've got great technology for marketing cigarettes to children, so we're going to pivot your company to do that entirely, and we're going to pay you 50% more than the market value, you're going to have to say yes. If you have a PBC, then you are more than welcome to say no, if that offer is not in line with your stated public benefit. So our stated public benefit is to maximize the benefit to society through using AI. So given that more children smoking doesn't do that, then we can say like, no, we're not selling to you.
Alessio [00:11:01]: I was looking back at some of our emails. You sent me an email on November 13th about talking and then on the 14th, I sent you an email working together to free AI was the subject line. And then that was kind of the start of the C round. And then two days later, someone got fired. So you know, you were having these thoughts even before we had like a public example of like why some of the current structures didn't work. So yeah, you were very ahead of the curve, so to speak. You know, people can read your awesome introduction blog and answer and the idea of having a R&D lab versus our lab and then a D lab somewhere else. I think to me, the most interesting thing has been hiring and some of the awesome people that you've been bringing on that maybe don't fit the central casting of Silicon Valley, so to speak. Like sometimes I got it like playing baseball cards, you know, people are like, oh, what teams was this person on, where did they work versus focusing on ability. So I would love for you to give a shout out to some of the awesome folks that you have on the team.
Jeremy [00:11:58]: So, you know, there's like a graphic going around describing like the people at XAI, you know, Elon Musk thing. And like they are all connected to like multiple of Stanford, Meta, DeepMind, OpenAI, Berkeley, Oxford. Look, these are all great institutions and they have good people. And I'm definitely not at all against that, but damn, there's so many other people. And one of the things I found really interesting is almost any time I see something which I think like this is really high quality work and it's something I don't think would have been built if that person hadn't built the thing right now, I nearly always reach out to them and ask to chat. And I tend to dig in to find out like, okay, you know, why did you do that thing? Everybody else has done this other thing, your thing's much better, but it's not what other people are working on. And like 80% of the time, I find out the person has a really unusual background. So like often they'll have like, either they like came from poverty and didn't get an opportunity to go to a good school or had dyslexia and, you know, got kicked out of school in year 11, or they had a health issue that meant they couldn't go to university or something happened in their past and they ended up out of the mainstream. And then they kind of succeeded anyway. Those are the people that throughout my career, I've tended to kind of accidentally hire more of, but it's not exactly accidentally. It's like when I see somebody who's done, two people who have done extremely well, one of them did extremely well in exactly the normal way from the background entirely pointing in that direction and they achieved all the hurdles to get there. And like, okay, that's quite impressive, you know, but another person who did just as well, despite lots of constraints and doing things in really unusual ways and came up with different approaches. That's normally the person I'm likely to find useful to work with because they're often like risk-takers, they're often creative, they're often extremely tenacious, they're often very open-minded. So that's the kind of folks I tend to find myself hiring. So now at Answer.ai, it's a group of people that are strong enough that nearly every one of them has independently come to me in the past few weeks and told me that they have imposter syndrome and they're not convinced that they're good enough to be here. And I kind of heard it at the point where I was like, okay, I don't think it's possible that all of you are so far behind your peers that you shouldn't get to be here. But I think part of the problem is as an R&D lab, the great developers look at the great researchers and they're like, wow, these big-brained, crazy research people with all their math and s**t, they're too cool for me, oh my God. And then the researchers look at the developers and they're like, oh, they're killing it, making all this stuff with all these people using it and talking on Twitter about how great it is. I think they're both a bit intimidated by each other, you know. And so I have to kind of remind them like, okay, there are lots of things in this world where you suck compared to lots of other people in this company, but also vice versa, you know, for all things. And the reason you came here is because you wanted to learn about those other things from those other people and have an opportunity to like bring them all together into a single unit. You know, it's not reasonable to expect you're going to be better at everything than everybody else. I guess the other part of it is for nearly all of the people in the company, to be honest, they have nearly always been better than everybody else at nearly everything they're doing nearly everywhere they've been. So it's kind of weird to be in this situation now where it's like, gee, I can clearly see that I suck at this thing that I'm meant to be able to do compared to these other people where I'm like the worst in the company at this thing for some things. So I think that's a healthy place to be, you know, as long as you keep reminding each other about that's actually why we're here. And like, it's all a bit of an experiment, like we don't have any managers. We don't have any hierarchy from that point of view. So for example, I'm not a manager, which means I don't get to tell people what to do or how to do it or when to do it. Yeah, it's been a bit of an experiment to see how that would work out. And it's been great. So for instance, Ben Clavier, who you might have come across, he's the author of Ragatouille, he's the author of Rerankers, super strong information retrieval guy. And a few weeks ago, you know, this additional channel appeared on Discord, on our private Discord called Bert24. And these people started appearing, as in our collab sections, we have a collab section for like collaborating with outsiders. And these people started appearing, there are all these names that I recognize, like Bert24, and they're all talking about like the next generation of Bert. And I start following along, it's like, okay, Ben decided that I think, quite rightly, we need a new Bert. Because everybody, like so many people are still using Bert, and it's still the best at so many things, but it actually doesn't take advantage of lots of best practices. And so he just went out and found basically everybody who's created better Berts in the last four or five years, brought them all together, suddenly there's this huge collaboration going on. So yeah, I didn't tell him to do that. He didn't ask my permission to do that. And then, like, Benjamin Warner dived in, and he's like, oh, I created a whole transformers from scratch implementation designed to be maximally hackable. He originally did it largely as a teaching exercise to show other people, but he was like, I could, you know, use that to create a really hackable BERT implementation. In fact, he didn't say that. He said, I just did do that, you know, and I created a repo, and then everybody's like starts using it. They're like, oh my god, this is amazing. I can now implement all these other BERT things. And it's not just answer AI guys there, you know, there's lots of folks, you know, who have like contributed new data set mixes and blah, blah, blah. So, I mean, I can help in the same way that other people can help. So like, then Ben Clavier reached out to me at one point and said, can you help me, like, what have you learned over time about how to manage intimidatingly capable and large groups of people who you're nominally meant to be leading? And so, you know, I like to try to help, but I don't direct. Another great example was Kerem, who, after our FSTP QLORA work, decided quite correctly that it didn't really make sense to use LoRa in today's world. You want to use the normalized version, which is called Dora. Like two or three weeks after we did FSTP QLORA, he just popped up and said, okay, I've just converted the whole thing to Dora, and I've also created these VLLM extensions, and I've got all these benchmarks, and, you know, now I've got training of quantized models with adapters that are as fast as LoRa, and as actually better than, weirdly, fine tuning. Just like, okay, that's great, you know. And yeah, so the things we've done to try to help make these things happen as well is we don't have any required meetings, you know, but we do have a meeting for each pair of major time zones that everybody's invited to, and, you know, people see their colleagues doing stuff that looks really cool and say, like, oh, how can I help, you know, or how can I learn or whatever. So another example is Austin, who, you know, amazing background. He ran AI at Fidelity, he ran AI at Pfizer, he ran browsing and retrieval for Google's DeepMind stuff, created Jemma.cpp, and he's been working on a new system to make it easier to do web GPU programming, because, again, he quite correctly identified, yeah, so I said to him, like, okay, I want to learn about that. Not an area that I have much expertise in, so, you know, he's going to show me what he's working on and teach me a bit about it, and hopefully I can help contribute. I think one of the key things that's happened in all of these is everybody understands what Eric Gilliam, who wrote the second blog post in our series, the R&D historian, describes as a large yard with narrow fences. Everybody has total flexibility to do what they want. We all understand kind of roughly why we're here, you know, we agree with the premises around, like, everything's too expensive, everything's too complicated, people are building too many vanity foundation models rather than taking better advantage of fine-tuning, like, there's this kind of general, like, sense of we're all on the same wavelength about, you know, all the ways in which current research is fucked up, and, you know, all the ways in which we're worried about centralization. We all care a lot about not just research for the point of citations, but research that actually wouldn't have happened otherwise, and actually is going to lead to real-world outcomes. And so, yeah, with this kind of, like, shared vision, people understand, like, you know, so when I say, like, oh, well, you know, tell me, Ben, about BERT 24, what's that about? And he's like, you know, like, oh, well, you know, you can see from an accessibility point of view, or you can see from a kind of a actual practical impact point of view, there's far too much focus on decoder-only models, and, you know, like, BERT's used in all of these different places and industry, and so I can see, like, in terms of our basic principles, what we're trying to achieve, this seems like something important. And so I think that's, like, a really helpful that we have that kind of shared perspective, you know?
Alessio [00:21:14]: Yeah. And before we maybe talk about some of the specific research, when you're, like, reaching out to people, interviewing them, what are some of the traits, like, how do these things come out, you know, usually? Is it working on side projects that you, you know, you're already familiar with? Is there anything, like, in the interview process that, like, helps you screen for people that are less pragmatic and more research-driven versus some of these folks that are just gonna do it, you know? They're not waiting for, like, the perfect process.
Jeremy [00:21:40]: Everybody who comes through the recruiting is interviewed by everybody in the company. You know, our goal is 12 people, so it's not an unreasonable amount. So the other thing to say is everybody so far who's come into the recruiting pipeline, everybody bar one, has been hired. So which is to say our original curation has been good. And that's actually pretty easy, because nearly everybody who's come in through the recruiting pipeline are people I know pretty well. So Jono Whitaker and I, you know, he worked on the stable diffusion course we did. He's outrageously creative and talented, and he's super, like, enthusiastic tinkerer, just likes making things. Benjamin was one of the strongest parts of the fast.ai community, which is now the alumni. It's, like, hundreds of thousands of people. And you know, again, like, they're not people who a normal interview process would pick up, right? So Benjamin doesn't have any qualifications in math or computer science. Jono was living in Zimbabwe, you know, he was working on, like, helping some African startups, you know, but not FAANG kind of credentials. But yeah, I mean, when you actually see people doing real work and they stand out above, you know, we've got lots of Stanford graduates and open AI people and whatever in our alumni community as well. You know, when you stand out above all of those people anyway, obviously you've got something going for you. You know, Austin, him and I worked together on the masks study we did in the proceeding at the National Academy of Science. You know, we had worked together, and again, that was a group of, like, basically the 18 or 19 top experts in the world on public health and epidemiology and research design and so forth. And Austin, you know, one of the strongest people in that collaboration. So yeah, you know, like, I've been lucky enough to have had opportunities to work with some people who are great and, you know, I'm a very open-minded person, so I kind of am always happy to try working with pretty much anybody and some people stand out. You know, there have been some exceptions, people I haven't previously known, like Ben Clavier, actually, I didn't know before. But you know, with him, you just read his code, and I'm like, oh, that's really well-written code. And like, it's not written exactly the same way as everybody else's code, and it's not written to do exactly the same thing as everybody else's code. So yeah, and then when I chatted to him, it's just like, I don't know, I felt like we'd known each other for years, like we just were on the same wavelength, but I could pretty much tell that was going to happen just by reading his code. I think you express a lot in the code you choose to write and how you choose to write it, I guess. You know, or another example, a guy named Vic, who was previously the CEO of DataQuest, and like, in that case, you know, he's created a really successful startup. He won the first, basically, Kaggle NLP competition, which was automatic essay grading. He's got the current state-of-the-art OCR system, Surya. Again, he's just a guy who obviously just builds stuff, you know, he doesn't ask for permission, he doesn't need any, like, external resources. Actually, Karim's another great example of this, I mean, I already knew Karim very well because he was my best ever master's student, but it wasn't a surprise to me then when he then went off to create the world's state-of-the-art language model in Turkish on his own, in his spare time, with no budget, from scratch. This is not fine-tuning or whatever, he, like, went back to Common Crawl and did everything. Yeah, it's kind of, I don't know what I'd describe that process as, but it's not at all based on credentials.
Swyx [00:25:17]: Assemble based on talent, yeah. We wanted to dive in a little bit more on, you know, turning from the people side of things into the technical bets that you're making. Just a little bit more on Bert. I was actually, we just did an interview with Yi Tay from Reka, I don't know if you're familiar with his work, but also another encoder-decoder bet, and one of his arguments was actually people kind of over-index on the decoder-only GPT-3 type paradigm. I wonder if you have thoughts there that is maybe non-consensus as well. Yeah, no, absolutely.
Jeremy [00:25:45]: So I think it's a great example. So one of the people we're collaborating with a little bit with BERT24 is Colin Raffle, who is the guy behind, yeah, most of that stuff, you know, between that and UL2, there's a lot of really interesting work. And so one of the things I've been encouraging the BERT group to do, Colin has as well, is to consider using a T5 pre-trained encoder backbone as a thing you fine-tune, which I think would be really cool. You know, Colin was also saying actually just use encoder-decoder as your Bert, you know, why don't you like use that as a baseline, which I also think is a good idea. Yeah, look.
Swyx [00:26:25]: What technical arguments are people under-weighting?
Jeremy [00:26:27]: I mean, Colin would be able to describe this much better than I can, but I'll give my slightly non-expert attempt. Look, I mean, think about like diffusion models, right? Like in stable diffusion, like we use things like UNet. You have this kind of downward path and then in the upward path you have the cross connections, which it's not a tension, but it's like a similar idea, right? You're inputting the original encoding path into your decoding path. It's critical to make it work, right? Because otherwise in the decoding part, the model has to do so much kind of from scratch. So like if you're doing translation, like that's a classic kind of encoder-decoder example. If it's decoder only, you never get the opportunity to find the right, you know, feature engineering, the right feature encoding for the original sentence. And it kind of means then on every token that you generate, you have to recreate the whole thing, you know? So if you have an encoder, it's basically saying like, okay, this is your opportunity model to create a really useful feature representation for your input information. So I think there's really strong arguments for encoder-decoder models anywhere that there is this kind of like context or source thing. And then why encoder only? Well, because so much of the time what we actually care about is a classification, you know? It's like an output. It's like generating an arbitrary length sequence of tokens. So anytime you're not generating an arbitrary length sequence of tokens, decoder models don't seem to make much sense. Now the interesting thing is, you see on like Kaggle competitions, that decoder models still are at least competitive with things like Deberta v3. They have to be way bigger to be competitive with things like Deberta v3. And the only reason they are competitive is because people have put a lot more time and money and effort into training the decoder only ones, you know? There isn't a recent Deberta. There isn't a recent Bert. Yeah, it's a whole part of the world that people have slept on a little bit. And this is just what happens. This is how trends happen rather than like, to me, everybody should be like, oh, let's look at the thing that has shown signs of being useful in the past, but nobody really followed up with properly. That's the more interesting path, you know, where people tend to be like, oh, I need to get citations. So what's everybody else doing? Can I make it 0.1% better, you know, or 0.1% faster? That's what everybody tends to do. Yeah. So I think it's like, Itay's work commercially now is interesting because here's like a whole, here's a whole model that's been trained in a different way. So there's probably a whole lot of tasks it's probably better at than GPT and Gemini and Claude. So that should be a good commercial opportunity for them if they can figure out what those tasks are.
Swyx [00:29:07]: Well, if rumors are to be believed, and he didn't comment on this, but, you know, Snowflake may figure out the commercialization for them. So we'll see.
Jeremy [00:29:14]: Good.
Alessio [00:29:16]: Let's talk about FSDP, Qlora, Qdora, and all of that awesome stuff. One of the things we talked about last time, some of these models are meant to run on systems that nobody can really own, no single person. And then you were like, well, what if you could fine tune a 70B model on like a 4090? And I was like, no, that sounds great, Jeremy, but like, can we actually do it? And then obviously you all figured it out. Can you maybe tell us some of the worst stories behind that, like the idea behind FSDP, which is kind of taking sharded data, parallel computation, and then Qlora, which is do not touch all the weights, just go quantize some of the model, and then within the quantized model only do certain layers instead of doing everything.
Jeremy [00:29:57]: Well, do the adapters. Yeah.
Alessio [00:29:59]: Yeah. Yeah. Do the adapters. Yeah. I will leave the floor to you. I think before you published it, nobody thought this was like a short term thing that we're just going to have. And now it's like, oh, obviously you can do it, but it's not that easy.
Jeremy [00:30:12]: Yeah. I mean, to be honest, it was extremely unpleasant work to do. It's like not at all enjoyable. I kind of did version 0.1 of it myself before we had launched the company, or at least the kind of like the pieces. They're all pieces that are difficult to work with, right? So for the quantization, you know, I chatted to Tim Detmers quite a bit and, you know, he very much encouraged me by saying like, yeah, it's possible. He actually thought it'd be easy. It probably would be easy for him, but I'm not Tim Detmers. And, you know, so he wrote bits and bytes, which is his quantization library. You know, he wrote that for a paper. He didn't write that to be production like code. It's now like everybody's using it, at least the CUDA bits. So like, it's not particularly well structured. There's lots of code paths that never get used. There's multiple versions of the same thing. You have to try to figure it out. So trying to get my head around that was hard. And you know, because the interesting bits are all written in CUDA, it's hard to like to step through it and see what's happening. And then, you know, FSTP is this very complicated library and PyTorch, which not particularly well documented. So the only really, really way to understand it properly is again, just read the code and step through the code. And then like bits and bytes doesn't really work in practice unless it's used with PEF, the HuggingFace library and PEF doesn't really work in practice unless you use it with other things. And there's a lot of coupling in the HuggingFace ecosystem where like none of it works separately. You have to use it all together, which I don't love. So yeah, trying to just get a minimal example that I can play with was really hard. And so I ended up having to rewrite a lot of it myself to kind of create this like minimal script. One thing that helped a lot was Medec had this LlamaRecipes repo that came out just a little bit before I started working on that. And like they had a kind of role model example of like, here's how to train FSTP, LoRa, didn't work with QLoRa on Llama. A lot of the stuff I discovered, the interesting stuff would be put together by Les Wright, who's, he was actually the guy in the Fast.ai community I mentioned who created the Ranger Optimizer. So he's doing a lot of great stuff at Meta now. So yeah, I kind of, that helped get some minimum stuff going and then it was great once Benjamin and Jono joined full time. And so we basically hacked at that together and then Kerim joined like a month later or something. And it was like, gee, it was just a lot of like fiddly detailed engineering on like barely documented bits of obscure internals. So my focus was to see if it kind of could work and I kind of got a bit of a proof of concept working and then the rest of the guys actually did all the work to make it work properly. And, you know, every time we thought we had something, you know, we needed to have good benchmarks, right? So we'd like, it's very easy to convince yourself you've done the work when you haven't, you know, so then we'd actually try lots of things and be like, oh, and these like really important cases, the memory use is higher, you know, or it's actually slower. And we'd go in and we just find like all these things that were nothing to do with our library that just didn't work properly. And nobody had noticed they hadn't worked properly because nobody had really benchmarked it properly. So we ended up, you know, trying to fix a whole lot of different things. And even as we did so, new regressions were appearing in like transformers and stuff that Benjamin then had to go away and figure out like, oh, how come flash attention doesn't work in this version of transformers anymore with this set of models and like, oh, it turns out they accidentally changed this thing, so it doesn't work. You know, there's just, there's not a lot of really good performance type evals going on in the open source ecosystem. So there's an extraordinary amount of like things where people say like, oh, we built this thing and it has this result. And when you actually check it, so yeah, there's a shitload of war stories from getting that thing to work. And it did require a particularly like tenacious group of people and a group of people who don't mind doing a whole lot of kind of like really janitorial work, to be honest, to get the details right, to check them. Yeah.
Alessio [00:34:09]: We had a trade out on the podcast and we talked about how a lot of it is like systems work to make some of these things work. It's not just like beautiful, pure math that you do on a blackboard. It's like, how do you get into the nitty gritty?
Jeremy [00:34:22]: I mean, flash attention is a great example of that. Like it's, it basically is just like, oh, let's just take the attention and just do the tiled version of it, which sounds simple enough, you know, but then implementing that is challenging at lots of levels.
Alessio [00:34:36]: Yeah. What about inference? You know, obviously you've done all this amazing work on fine tuning. Do you have any research you've been doing on the inference side, how to make local inference really fast on these models too?
Jeremy [00:34:47]: We're doing quite a bit on that at the moment. We haven't released too much there yet. But one of the things I've been trying to do is also just to help other people. And one of the nice things that's happened is that a couple of folks at Meta, including Mark Saroufim, have done a nice job of creating this CUDA mode community of people working on like CUDA kernels or learning about that. And I tried to help get that going well as well and did some lessons to help people get into it. So there's a lot going on in both inference and fine tuning performance. And a lot of it's actually happening kind of related to that. So PyTorch team have created this Torch AO project on quantization. And so there's a big overlap now between kind of the FastAI and AnswerAI and CUDA mode communities of people working on stuff for both inference and fine tuning. But we're getting close now. You know, our goal is that nobody should be merging models, nobody should be downloading merged models, everybody should be using basically quantized plus adapters for almost everything and just downloading the adapters. And that should be much faster. So that's kind of the place we're trying to get to. It's difficult, you know, because like Karim's been doing a lot of work with VLM, for example. These inference engines are pretty complex bits of code. They have a whole lot of custom kernel stuff going on as well, as do the quantization libraries. So we've been working on, we're also quite a bit of collaborating with the folks who do HQQ, which is a really great quantization library and works super well. So yeah, there's a lot of other people outside AnswerAI that we're working with a lot who are really helping on all this performance optimization stuff, open source.
Swyx [00:36:27]: Just to follow up on merging models, I picked up there that you said nobody should be merging models. That's interesting because obviously a lot of people are experimenting with this and finding interesting results. I would say in defense of merging models, you can do it without data. That's probably the only thing that's going for it.
Jeremy [00:36:45]: To explain, it's not that you shouldn't merge models. You shouldn't be distributing a merged model. You should distribute a merged adapter 99% of the time. And actually often one of the best things happening in the model merging world is actually that often merging adapters works better anyway. The point is, Sean, that once you've got your new model, if you distribute it as an adapter that sits on top of a quantized model that somebody's already downloaded, then it's a much smaller download for them. And also the inference should be much faster because you're not having to transfer FB16 weights from HPM memory at all or ever load them off disk. You know, all the main weights are quantized and the only floating point weights are in the adapters. So that should make both inference and fine tuning faster. Okay, perfect.
Swyx [00:37:33]: We're moving on a little bit to the rest of the fast universe. I would have thought that, you know, once you started Answer.ai, that the sort of fast universe would be kind of on hold. And then today you just dropped Fastlight and it looks like, you know, there's more activity going on in sort of Fastland.
Jeremy [00:37:49]: Yeah. So Fastland and Answerland are not really distinct things. Answerland is kind of like the Fastland grown up and funded. They both have the same mission, which is to maximize the societal benefit of AI broadly. We want to create thousands of commercially successful products at Answer.ai. And we want to do that with like 12 people. So that means we need a pretty efficient stack, you know, like quite a few orders of magnitude more efficient, not just for creation, but for deployment and maintenance than anything that currently exists. People often forget about the D part of our R&D firm. So we've got to be extremely good at creating, deploying and maintaining applications, not just models. Much to my horror, the story around creating web applications is much worse now than it was 10 or 15 years ago in terms of, if I say to a data scientist, here's how to create and deploy a web application, you know, either you have to learn JavaScript or TypeScript and about all the complex libraries like React and stuff, and all the complex like details around security and web protocol stuff around how you then talk to a backend and then all the details about creating the backend. You know, if that's your job and, you know, you have specialists who work in just one of those areas, it is possible for that to all work. But compared to like, oh, write a PHP script and put it in the home directory that you get when you sign up to this shell provider, which is what it was like in the nineties, you know, here are those 25 lines of code and you're done and now you can pass that URL around to all your friends, or put this, you know, .pl file inside the CGI bin directory that you got when you signed up to this web host. So yeah, the thing I've been mainly working on the last few weeks is fixing all that. And I think I fixed it. I don't know if this is an announcement, but I tell you guys, so yeah, there's this thing called fastHTML, which basically lets you create a complete web application in a single Python file. Unlike excellent projects like Streamlit and Gradio, you're not working on top of a highly abstracted thing. That's got nothing to do with web foundations. You're working with web foundations directly, but you're able to do it by using pure Python. There's no template, there's no ginger, there's no separate like CSS and JavaScript files. It looks and behaves like a modern SPA web application. And you can create components for like daisy UI, or bootstrap, or shoelace, or whatever fancy JavaScript and or CSS tailwind etc library you like, but you can write it all in Python. You can pip install somebody else's set of components and use them entirely from Python. You can develop and prototype it all in a Jupyter notebook if you want to. It all displays correctly, so you can like interactively do that. And then you mentioned Fastlight, so specifically now if you're using SQLite in particular, it's like ridiculously easy to have that persistence, and all of your handlers will be passed database ready objects automatically, that you can just call dot delete dot update dot insert on. Yeah, you get session, you get security, you get all that. So again, like with most everything I do, it's very little code. It's mainly tying together really cool stuff that other people have written. You don't have to use it, but a lot of the best stuff comes from its incorporation of HTMX, which to me is basically the thing that changes your browser to make it work the way it always should have. So it just does four small things, but those four small things are the things that are basically unnecessary constraints that HTML should never have had, so it removes the constraints. It sits on top of Starlet, which is a very nice kind of lower level platform for building these kind of web applications. The actual interface matches as closely as possible to FastAPI, which is a really nice system for creating the kind of classic JavaScript type applications. And Sebastian, who wrote FastAPI, has been kind enough to help me think through some of these design decisions, and so forth. I mean, everybody involved has been super helpful. Actually, I chatted to Carson, who created HTMX, you know, so about it. Some of the folks involved in Django, like everybody in the community I've spoken to definitely realizes there's a big gap to be filled around, like, highly scalable, web foundation-based, pure Python framework with a minimum of fuss. So yeah, I'm getting a lot of support and trying to make sure that FastHTML works well for people.
Swyx [00:42:38]: I would say, when I heard about this, I texted Alexio. I think this is going to be pretty huge. People consider Streamlit and Gradio to be the state of the art, but I think there's so much to improve, and having what you call web foundations and web fundamentals at the core of it, I think, would be really helpful.
Jeremy [00:42:54]: I mean, it's based on 25 years of thinking and work for me. So like, FastML was built on a system much like this one, but that was of hell. And so I spent, you know, 10 years working on that. We had millions of people using that every day, really pushing it hard. And I really always enjoyed working in that. Yeah. So, you know, and obviously lots of other people have done like great stuff, and particularly HTMX. So I've been thinking about like, yeah, how do I pull together the best of the web framework I created for FastML with HTMX? There's also things like PicoCSS, which is the CSS system, which by default, FastHTML comes with. Although, as I say, you can pip install anything you want to, but it makes it like super easy to, you know, so we try to make it so that just out of the box, you don't have any choices to make. Yeah. You can make choices, but for most people, you just, you know, it's like the PHP in your home directory thing. You just start typing and just by default, you'll get something which looks and feels, you know, pretty okay. And if you want to then write a version of Gradio or Streamlit on top of that, you totally can. And then the nice thing is if you then write it in kind of the Gradio equivalent, which will be, you know, I imagine we'll create some kind of pip installable thing for that. Once you've outgrown, or if you outgrow that, it's not like, okay, throw that all away and start again. And this like whole separate language that it's like this kind of smooth, gentle path that you can take step-by-step because it's all just standard web foundations all the way, you know.
Swyx [00:44:29]: Just to wrap up the sort of open source work that you're doing, you're aiming to create thousands of projects with a very, very small team. I haven't heard you mention once AI agents or AI developer tooling or AI code maintenance. I know you're very productive, but you know, what is the role of AI in your own work?
Jeremy [00:44:47]: So I'm making something. I'm not sure how much I want to say just yet.
Swyx [00:44:52]: Give us a nibble.
Jeremy [00:44:53]: All right. I'll give you the key thing. So I've created a new approach. It's not called prompt engineering. It's called dialogue engineering. But I'm creating a system for doing dialogue engineering. It's currently called AI magic. I'm doing most of my work in this system and it's making me much more productive than I was before I used it. So I always just build stuff for myself and hope that it'll be useful for somebody else. Think about chat GPT with code interpreter, right? The basic UX is the same as a 1970s teletype, right? So if you wrote APL on a teletype in the 1970s, you typed onto a thing, your words appeared at the bottom of a sheet of paper and you'd like hit enter and it would scroll up. And then the answer from APL would be printed out, scroll up, and then you would type the next thing. And like, which is also the way, for example, a shell works like bash or ZSH or whatever. It's not terrible, you know, like we all get a lot done in these like very, very basic teletype style REPL environments, but I've never felt like it's optimal and everybody else has just copied chat GPT. So it's also the way BART and Gemini work. It's also the way the Claude web app works. And then you add code interpreter. And the most you can do is to like plead with chat GPT to write the kind of code I want. It's pretty good for very, very, very beginner users who like can't code at all, like by default now the code's even hidden away, so you never even have to see it ever happened. But for somebody who's like wanting to learn to code or who already knows a bit of code or whatever, it's, it seems really not ideal. So okay, that's one end of the spectrum. The other end of the spectrum, which is where Sean's work comes in, is, oh, you want to do more than chat GPT? No worries. Here is Visual Studio Code. I run it. There's an empty screen with a flashing cursor. Okay, start coding, you know, and it's like, okay, you can use systems like Sean's or like cursor or whatever to be like, okay, Apple K in cursors, like a creative form that blah, blah, blah. But in the end, it's like a convenience over the top of this incredibly complicated system that full-time sophisticated software engineers have designed over the past few decades in a totally different environment as a way to build software, you know. And so we're trying to like shoehorn in AI into that. And it's not easy to do. And I think there are like much better ways of thinking about the craft of software development in a language model world to be much more interactive, you know. So the thing that I'm building is neither of those things. It's something between the two. And it's built around this idea of crafting a dialogue, you know, where the outcome of the dialogue is the artifacts that you want, whether it be a piece of analysis or whether it be a Python library or whether it be a technical blog post or whatever. So as part of building that, I've created something called Claudette, which is a library for Claude. I've created something called Cosette, which is a library for OpenAI. They're libraries which are designed to make those APIs much more usable, much easier to use, much more concise. And then I've written AI magic on top of those. And that's been an interesting exercise because I did Claudette first, and I was looking at what Simon Willison did with his fantastic LLM library. And his library is designed around like, let's make something that supports all the LLM inference engines and commercial providers. I thought, okay, what if I did something different, which is like make something that's as Claude friendly as possible and forget everything else. So that's what Claudette was. So for example, one of the really nice things in Claude is prefill. So by telling the assistant that this is what your response started with, there's a lot of powerful things you can take advantage of. So yeah, I created Claudette to be as Claude friendly as possible. And then after I did that, and then particularly with GPT 4.0 coming out, I kind of thought, okay, now let's create something that's as OpenAI friendly as possible. And then I tried to look to see, well, where are the similarities and where are the differences? And now can I make them compatible in places where it makes sense for them to be compatible without losing out on the things that make each one special for what they are. So yeah, those are some of the things I've been working on in that space. And I'm thinking we might launch AI magic via a course called how to solve it with code. The name is based on the classic Polya book, if you know how to solve it, which is, you know, one of the classic math books of all time, where we're basically going to try to show people how to solve challenging problems that they didn't think they could solve without doing a full computer science course, by taking advantage of a bit of AI and a bit of like practical skills, as particularly for this like whole generation of people who are learning to code with and because of ChatGPT. Like I love it, I know a lot of people who didn't really know how to code, but they've created things because they use ChatGPT, but they don't really know how to maintain them or fix them or add things to them that ChatGPT can't do, because they don't really know how to code. And so this course will be designed to show you how you can like either become a developer who can like supercharge their capabilities by using language models, or become a language model first developer who can supercharge their capabilities by understanding a bit about process and fundamentals.
Alessio [00:50:19]: Nice. That's a great spoiler. You know, I guess the fourth time you're going to be on learning space, we're going to talk about AI magic. Jeremy, before we wrap, this was just a great run through everything. What are the things that when you next come on the podcast in nine, 12 months, we're going to be like, man, Jeremy was like really ahead of it. Like, is there anything that you see in the space that maybe people are not talking enough? You know, what's the next company that's going to fall, like have drama internally, anything in your mind?
Jeremy [00:50:47]: You know, hopefully we'll be talking a lot about fast HTML and hopefully the international community that at that point has come up around that. And also about AI magic and about dialogue engineering. Hopefully dialogue engineering catches on because I think it's the right way to think about a lot of this stuff. What else? Just trying to think about all on the research side. Yeah. I think, you know, I mean, we've talked about a lot of it. Like I think encoder decoder architectures, encoder only architectures, hopefully we'll be talking about like the whole re-interest in BERT that BERT 24 stimulated.
Swyx [00:51:17]: There's a safe space model that came out today that might be interesting for this general discussion. One thing that stood out to me with Cartesia's blog posts was that they were talking about real time ingestion, billions and trillions of tokens, and keeping that context, obviously in the state space that they have.
Jeremy [00:51:34]: Yeah.
Swyx [00:51:35]: I'm wondering what your thoughts are because you've been entirely transformers the whole time.
Jeremy [00:51:38]: Yeah. No. So obviously my background is RNNs and LSTMs. Of course. And I'm still a believer in the idea that state is something you can update, you know? So obviously Sepp Hochreiter came up, came out with xLSTM recently. Oh my God. Okay. Another whole thing we haven't talked about, just somewhat related. I've been going crazy for like a long time about like, why can I not pay anybody to save my KV cash? I just ingested the Great Gatsby or the documentation for Starlet or whatever, you know, I'm sending it as my prompt context. Why are you redoing it every time? So Gemini is about to finally come out with KV caching, and this is something that Austin actually in Gemma.cpp had had on his roadmap for years, well not years, months, long time. The idea that the KV cache is like a thing that, it's a third thing, right? So there's RAG, you know, there's in-context learning, you know, and prompt engineering, and there's KV cache creation. I think it creates like a whole new class almost of applications or as techniques where, you know, for me, for example, I very often work with really new libraries or I've created my own library that I'm now writing with rather than on. So I want all the docs in my new library to be there all the time. So I want to upload them once, and then we have a whole discussion about building this application using FastHTML. Well nobody's got FastHTML in their language model yet, I don't want to send all the FastHTML docs across every time. So one of the things I'm looking at doing in AI Magic actually is taking advantage of some of these ideas so that you can have the documentation of the libraries you're working on be kind of always available. Something over the next 12 months people will be spending time thinking about is how to like, where to use RAG, where to use fine-tuning, where to use KV cache storage, you know. And how to use state, because in state models and XLSTM, again, state is something you update. So how do we combine the best of all of these worlds?
Alessio [00:53:46]: And Jeremy, I know before you talked about how some of the autoregressive models are not maybe a great fit for agents. Any other thoughts on like JEPA, diffusion for text, any interesting thing that you've seen pop up?
Jeremy [00:53:58]: In the same way that we probably ought to have state that you can update, i.e. XLSTM and state models, in the same way that a lot of things probably should have an encoder, JEPA and diffusion both seem like the right conceptual mapping for a lot of things we probably want to do. So the idea of like, there should be a piece of the generative pipeline, which is like thinking about the answer and coming up with a sketch of what the answer looks like before you start outputting tokens. That's where it kind of feels like diffusion ought to fit, you know. And diffusion is, because it's not autoregressive, it's like, let's try to like gradually de-blur the picture of how to solve this. So this is also where dialogue engineering fits in, by the way. So with dialogue engineering, one of the reasons it's working so well for me is I use it to kind of like craft the thought process before I generate the code, you know. So yeah, there's a lot of different pieces here and I don't know how they'll all kind of exactly fit together. I don't know if JEPA is going to actually end up working in the text world. I don't know if diffusion will end up working in the text world, but they seem to be like trying to solve a class of problem which is currently unsolved.
Alessio [00:55:13]: Awesome, Jeremy. This was great, as usual. Thanks again for coming back on the pod and thank you all for listening. Yeah, that was fantastic.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe
The End of Finetuning — with Jeremy Howard of Fast.ai
Thanks to the over 17,000 people who have joined the first AI Engineer Summit! A full recap is coming. Last call to fill out the State of AI Engineering survey! See our Community page for upcoming meetups in SF, Paris and NYC.
This episode had good interest on Twitter and was discussed on the Vanishing Gradients podcast.
Fast.ai’s “Practical Deep Learning” courses been watched by over >6,000,000 people, and the fastai library has over 25,000 stars on Github. Jeremy Howard, one of the creators of Fast, is now one of the most prominent and respected voices in the machine learning industry; but that wasn’t always the case.
Being non-consensus and right
In 2018, Jeremy and Sebastian Ruder published a paper on ULMFiT (Universal Language Model Fine-tuning), a 3-step transfer learning technique for NLP tasks:
The paper demonstrated that pre-trained language models could be fine-tuned on a specific task with a relatively small amount of data to achieve state-of-the-art results. They trained a 24M parameters model on WikiText-103 which was beat most benchmarks.
While the paper had great results, the methods behind weren’t taken seriously by the community:
“Everybody hated fine tuning. Everybody hated transfer learning. I literally did tours trying to get people to start doing transfer learning and nobody was interested, particularly after GPT showed such good results with zero shot and few shot learning […] which I was convinced was not the right direction, but who's going to listen to me, cause as you said, I don't have a PhD, not at a university… I don't have a big set of computers to fine tune huge transformer models.”
Five years later, fine-tuning is at the center of most major discussion topics in AI (we covered some like fine tuning vs RAG and small models fine tuning), and we might have gotten here earlier if Jeremy had OpenAI-level access to compute and distribution. At heart, Jeremy has always been “GPU poor”:
“I've always been somebody who does not want to build stuff on lots of big computers because most people don't have lots of big computers and I hate creating stuff that most people can't use.”
This story is a good reminder of how some of the best ideas are hiding in plain sight; we recently covered RWKV and will continue to highlight the most interesting research that isn’t being done in the large labs.
Replacing fine-tuning with continued pre-training
Even though fine-tuning is now mainstream, we still have a lot to learn. The issue of “catastrophic forgetting” and potential solutions have been brought up in many papers: at the fine-tuning stage, the model can forget tasks it previously knew how to solve in favor of new ones.
The other issue is apparent memorization of the dataset even after a single epoch, which Jeremy covered Can LLMs learn from a single example? but we still don’t have the answer to.
Despite being the creator of ULMFiT, Jeremy still professes that there are a lot of open questions on finetuning:
“So I still don't know how to fine tune language models properly and I haven't found anybody who feels like they do.”
He now advocates for "continued pre-training" - maintaining a diversity of data throughout the training process rather than separate pre-training and fine-tuning stages. Mixing instructional data, exercises, code, and other modalities while gradually curating higher quality data can avoid catastrophic forgetting and lead to more robust capabilities (something we covered in Datasets 101).
“Even though I originally created three-step approach that everybody now does, my view is it's actually wrong and we shouldn't use it… the right way to do this is to fine-tune language models, is to actually throw away the idea of fine-tuning. There's no such thing. There's only continued pre-training.
And pre-training is something where from the very start, you try to include all the kinds of data that you care about, all the kinds of problems that you care about, instructions, exercises, code, general purpose document completion, whatever. And then as you train, you gradually curate that, you know, you gradually make that higher and higher quality and more and more specific to the kinds of tasks you want it to do. But you never throw away any data….
So yeah, that's now my view, is I think ULMFiT is the wrong approach. And that's why we're seeing a lot of these so-called alignment tax… I think it's actually because people are training them wrong.
An example of this phenomena is CodeLlama, a LLaMA2 model finetuned on 500B tokens of code: while the model is much better at code, it’s worse on generic tasks that LLaMA2 knew how to solve well before the fine-tuning.
In the episode we also dive into all the places where open source model development and research is happening (academia vs Discords - tracked on our Communities list and on our survey), and how Jeremy recommends getting the most out of these diffuse, pseudonymous communities (similar to the Eleuther AI Mafia).
Show Notes
* Jeremy’s Background
* FastMail
* Optimal Decisions
* Kaggle
* Enlitic
* fast.ai
* Rachel Thomas
* Practical Deep Learning
* fastai for PyTorch
* nbdev
* fastec2 (the underrated library we describe)
* Can LLMs learn from a single example?
* the Kaggle LLM Science Exam competition, which “challenges participants to answer difficult science-based questions written by a Large Language Model”.
* Sebastian Ruder
* Alec Radford
* Sylvain Gugger
* Stephen Merity
* Chris Lattner
* Modular.ai / Mojo
* Jono Whittaker
* Zeiler and Fergus paper
* ULM Fit
* DAWNBench
* Phi-1
* Code Llama
* AlexNet
Timestamps
* [00:00:00] Intros and Jeremy’s background
* [00:05:28] Creating ULM Fit - a breakthrough in NLP using transfer learning
* [00:06:32] The rise of GPT and the appeal of few-shot learning over fine-tuning
* [00:10:00] Starting Fast.ai to distribute AI capabilities beyond elite academics
* [00:14:30] How modern LMs like ChatGPT still follow the ULM Fit 3-step approach
* [00:17:23] Meeting with Chris Lattner on Swift for TensorFlow at Google
* [00:20:00] Continued pre-training as a fine-tuning alternative
* [00:22:16] Fast.ai and looking for impact vs profit maximization
* [00:26:39] Using Fast.ai to create an "army" of AI experts to improve their domains
* [00:29:32] Fast.ai's 3 focus areas - research, software, and courses
* [00:38:42] Fine-tuning memorization and training curve "clunks" before each epoch
* [00:46:47] Poor training and fine-tuning practices may be causing alignment failures
* [00:48:38] Academia vs Discords
* [00:53:41] Jeremy's high hopes for Chris Lattner's Mojo and its potential
* [01:05:00] Adding capabilities like SQL generation through quick fine-tuning
* [01:10:12] Rethinking Fast.ai courses for the AI-assisted coding era
* [01:14:53] Rapid model development has created major technical debt
* [01:17:08] Lightning Round
AI Summary (beta)
This is the first episode we’re trying this. Here’s an overview of the main topics before you dive in the transcript.
* Jeremy's background and philosophies on AI
* Studied philosophy and cognitive science in college
* Focused on ethics and thinking about AI even 30 years ago
* Believes AI should be accessible to more people, not just elite academics/programmers
* Created fast.ai to make deep learning more accessible
* Development of transfer learning and ULMFit
* Idea of transfer learning critical for making deep learning accessible
* ULMFit pioneered transfer learning for NLP
* Proposed training general language models on large corpora then fine-tuning - this became standard practice
* Faced skepticism that this approach would work from NLP community
* Showed state-of-the-art results on text classification soon after trying it
* Current open questions around fine-tuning LLMs
* Models appear to memorize training data extremely quickly (after 1 epoch)
* This may hurt training dynamics and cause catastrophic forgetting
* Unclear how best to fine-tune models to incorporate new information/capabilities
* Need more research on model training dynamics and ideal data mixing
* Exciting new developments
* Mojo and new programming languages like Swift could enable faster model innovation
* Still lots of room for improvements in computer vision-like innovations in transformers
* Small models with fine-tuning may be surprisingly capable for many real-world tasks
* Prompting strategies enable models like GPT-3 to achieve new skills like playing chess at superhuman levels
* LLMs are like computer vision in 2013 - on the cusp of huge new breakthroughs in capabilities
* Access to AI research
* Many key convos happen in private Discord channels and forums
* Becoming part of these communities can provide great learning opportunities
* Being willing to do real work, not just talk about ideas, is key to gaining access
* The future of practical AI
* Coding becoming more accessible to non-programmers through AI assistance
* Pre-requisite programming experience for learning AI may no longer be needed
* Huge open questions remain about how to best train, fine-tune, and prompt LLMs
Transcript
Alessio: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO at Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI. [00:00:21]
Swyx: Hey, and today we have in the remote studio, Jeremy Howard all the way from Australia. Good morning. [00:00:27]
Jeremy: The remote studio, also known as my house. Good morning. Nice to see you. [00:00:32]
Swyx: Nice to see you too. I'm actually very used to seeing you in your mask as a message to people, but today we're mostly audio. But thank you for doing the very important public service of COVID awareness. It was a pleasure. [00:00:46]
Jeremy: It was all very annoying and frustrating and tedious, but somebody had to do it. [00:00:52]
Swyx: Somebody had to do it, especially somebody with your profile. I think it really drives home the message. So we tend to introduce people for them and then ask people to fill in the blanks on the personal side. Something I did not know about you was that you graduated with a BA in philosophy from the University of Melbourne. I assumed you had a PhD. [00:01:14]
Jeremy: No, I mean, I barely got through my BA because I was working 80 to 100 hour weeks at McKinsey and Company from 19 years old onwards. So I actually didn't attend any lectures in second and third year university. [00:01:35]
Swyx: Well, I guess you didn't need it or you're very sort of self-driven and self-motivated. [00:01:39]
Jeremy: I took two weeks off before each exam period when I was working at McKinsey. And then, I mean, I can't believe I got away with this in hindsight, I would go to all my professors and say, oh, I was meant to be in your class this semester and I didn't quite turn up. Were there any assignments I was meant to have done, whatever. I can't believe all of them let me basically have it. They basically always would say like, okay, well, if you can have this written by tomorrow, I'll accept it. So yeah, stressful way to get through university, but. [00:02:12]
Swyx: Well, it shows that, I guess, you min-maxed the opportunities. That definitely was a precursor. [00:02:18]
Jeremy: I mean, funnily, like in as much as I, you know, in philosophy, the things I found interesting and focused on in the little bit of time I did spend on it was ethics and cognitive science. And it's kind of really amazing that it's now come back around and those are actually genuinely useful things to know about, which I never thought would happen. [00:02:38]
Swyx: A lot of, yeah, a lot of relevant conversations there. So you were a consultant for a while and then in the magical month of June 1989, you founded both Optimal Decisions and Fastmeal, which I also briefly used. So thank you for that. [00:02:53]
Jeremy: Oh, good for you. Yeah. Cause I had read the statistics, which is that like 90% or something of small businesses fail. So I thought if I start two businesses, I have a higher chance. In hindsight, I was thinking of it as some kind of stochastic thing I didn't have control over, but it's a bit odd, but anyway. [00:03:10]
Swyx: And then you were president and chief scientist at Kaggle, which obviously is the sort of composition platform of machine learning. And then Enlitic, where you were working on using deep learning to improve medical diagnostics and clinical decisions. Yeah. [00:03:28]
Jeremy: I was actually the first company to use deep learning in medicine, so I kind of founded the field. [00:03:33]
Swyx: And even now that's still like a pretty early phase. And I actually heard you on your new podcast with Tanish, where you went very, very deep into the stuff, the kind of work that he's doing, such a young prodigy at his age. [00:03:47]
Jeremy: Maybe he's too old to be called a prodigy now, ex-prodigy. No, no. [00:03:51]
Swyx: I think he still counts. And anyway, just to round out the bio, you have a lot more other credentials, obviously, but most recently you started Fast.ai, which is still, I guess, your primary identity with Rachel Thomas. So welcome. [00:04:05]
Jeremy: Yep. [00:04:06]
Swyx: Thanks to my wife. Thank you. Yeah. Doing a lot of public service there with getting people involved in AI, and I can't imagine a better way to describe it than fast, fast.ai. You teach people from nothing to stable diffusion in seven weeks or something, and that's amazing. Yeah, yeah. [00:04:22]
Jeremy: I mean, it's funny, you know, when we started that, what was that, like 2016 or something, the idea that deep learning was something that you could make more accessible was generally considered stupid. Everybody knew that deep learning was a thing that you got a math or a computer science PhD, you know, there was one of five labs that could give you the appropriate skills and that you would join, yeah, basically from one of those labs, you might be able to write some papers. So yeah, the idea that normal people could use that technology to do good work was considered kind of ridiculous when we started it. And we weren't sure if it was possible either, but we kind of felt like we had to give it a go because the alternative was we were pretty sure that deep learning was on its way to becoming, you know, the most or one of the most, you know, important technologies in human history. And if the only people that could use it were a handful of computer science PhDs, that seemed like A, a big waste and B, kind of dangerous. [00:05:28]
Swyx: Yeah. [00:05:29]
Alessio: And, you know, well, I just wanted to know one thing on your bio that at Kaggle, you were also the top rank participant in both 2010 and 2011. So sometimes you see a lot of founders running companies that are not really in touch with the problem, but you were clearly building something that you knew a lot about, which is awesome. Talking about deep learning, you created, published a paper on ULM fit, which was kind of the predecessor to multitask learning and a lot of the groundwork that then went to into Transformers. I've read back on the paper and you turned this model, AWD LSTM, which I did the math and it was like 24 to 33 million parameters, depending on what training data set you use today. That's kind of like not even small, it's like super small. What were some of the kind of like contrarian takes that you had at the time and maybe set the stage a little bit for the rest of the audience on what was kind of like the state of the art, so to speak, at the time and what people were working towards? [00:06:32]
Jeremy: Yeah, the whole thing was a contrarian take, you know. So okay, so we started Fast.ai, my wife and I, and we thought, yeah, so we're trying to think, okay, how do we make it more accessible? So when we started thinking about it, it was probably 2015 and then 2016, we started doing something about it. Why is it inaccessible? Okay, well, A, no one knows how to do it other than a few number of people. And then when we asked those few number of people, well, how do you actually get good results? They would say like, oh, it's like, you know, a box of tricks that aren't published. So you have to join one of the labs and learn the tricks. So a bunch of unpublished tricks, not much software around, but thankfully there was Theano and rappers and particularly Lasagna, the rapper, but yeah, not much software around, not much in the way of data sets, you know, very hard to get started in terms of the compute. Like how do you get that set up? So yeah, no, everything was kind of inaccessible. And you know, as we started looking into it, we had a key insight, which was like, you know what, most of the compute and data for image recognition, for example, we don't need to do it. You know, there's this thing which nobody knows about, nobody talks about called transfer learning, where you take somebody else's model, where they already figured out like how to detect edges and gradients and corners and text and whatever else, and then you can fine tune it to do the thing you want to do. And we thought that's the key. That's the key to becoming more accessible in terms of compute and data requirements. So when we started Fast.ai, we focused from day one on transfer learning. Lesson one, in fact, was transfer learning, literally lesson one, something not normally even mentioned in, I mean, there wasn't much in the way of courses, you know, the courses out there were PhD programs that had happened to have recorded their lessons and they would rarely mention it at all. We wanted to show how to do four things that seemed really useful. You know, work with vision, work with tables of data, work with kind of recommendation systems and collaborative filtering and work with text, because we felt like those four kind of modalities covered a lot of the stuff that, you know, are useful in real life. And no one was doing anything much useful with text. Everybody was talking about word2vec, you know, like king plus queen minus woman and blah, blah, blah. It was like cool experiments, but nobody's doing anything like useful with it. NLP was all like lemmatization and stop words and topic models and bigrams and SPMs. And it was really academic and not practical. But I mean, to be honest, I've been thinking about this crazy idea for nearly 30 years since I had done cognitive science at university, where we talked a lot about the CELS Chinese room experiment. This idea of like, what if there was somebody that could kind of like, knew all of the symbolic manipulations required to answer questions in Chinese, but they didn't speak Chinese and they were kind of inside a room with no other way to talk to the outside world other than taking in slips of paper with Chinese written on them and then they do all their rules and then they pass back a piece of paper with Chinese back. And this room with a person in is actually fantastically good at answering any question you give them written in Chinese. You know, do they understand Chinese? And is this, you know, something that's intelligently working with Chinese? Ever since that time, I'd say the most thought, to me, the most thoughtful and compelling philosophical response is yes. You know, intuitively it feels like no, because that's just because we can't imagine such a large kind of system. But you know, if it looks like a duck and acts like a duck, it's a duck, you know, or to all intents and purposes. And so I always kind of thought, you know, so this is basically a kind of analysis of the limits of text. And I kind of felt like, yeah, if something could ingest enough text and could use the patterns it saw to then generate text in response to text, it could appear to be intelligent, you know. And whether that means it is intelligent or not is a different discussion and not one I find very interesting. Yeah. And then when I came across neural nets when I was about 20, you know, what I learned about the universal approximation theorem and stuff, and I started thinking like, oh, I wonder if like a neural net could ever get big enough and take in enough data to be a Chinese room experiment. You know, with that background and this kind of like interest in transfer learning, you know, I'd been thinking about this thing for kind of 30 years and I thought like, oh, I wonder if we're there yet, you know, because we have a lot of text. Like I can literally download Wikipedia, which is a lot of text. And I thought, you know, how would something learn to kind of answer questions or, you know, respond to text? And I thought, well, what if we used a language model? So language models are already a thing, you know, they were not a popular or well-known thing, but they were a thing. But language models exist to this idea that you could train a model to fill in the gaps. Or actually in those days it wasn't fill in the gaps, it was finish a string. And in fact, Andrej Karpathy did his fantastic RNN demonstration from this at a similar time where he showed like you can have it ingest Shakespeare and it will generate something that looks a bit like Shakespeare. I thought, okay, so if I do this at a much bigger scale, using all of Wikipedia, what would it need to be able to do to finish a sentence in Wikipedia effectively, to do it quite accurately quite often? I thought, geez, it would actually have to know a lot about the world, you know, it'd have to know that there is a world and that there are objects and that objects relate to each other through time and cause each other to react in ways and that causes proceed effects and that, you know, when there are animals and there are people and that people can be in certain positions during certain timeframes and then you could, you know, all that together, you can then finish a sentence like this was signed into law in 2016 by US President X and it would fill in the gap, you know. So that's why I tried to create what in those days was considered a big language model trained on the entirety on Wikipedia, which is that was, you know, a bit unheard of. And my interest was not in, you know, just having a language model. My interest was in like, what latent capabilities would such a system have that would allow it to finish those kind of sentences? Because I was pretty sure, based on our work with transfer learning and vision, that I could then suck out those latent capabilities by transfer learning, you know, by fine-tuning it on a task data set or whatever. So we generated this three-step system. So step one was train a language model on a big corpus. Step two was fine-tune a language model on a more curated corpus. And step three was further fine-tune that model on a task. And of course, that's what everybody still does today, right? That's what ChatGPT is. And so the first time I tried it within hours, I had a new state-of-the-art academic result on IMDB. And I was like, holy s**t, it does work. And so you asked, to what degree was this kind of like pushing against the established wisdom? You know, every way. Like the reason it took me so long to try it was because I asked all my friends in NLP if this could work. And everybody said, no, it definitely won't work. It wasn't like, oh, maybe. Everybody was like, it definitely won't work. NLP is much more complicated than vision. Language is a much more vastly complicated domain. You know, and you've got problems like the grounding problem. We know from like philosophy and theory of mind that it's actually impossible for it to work. So yeah, so don't waste your time. [00:15:10]
Alessio: Jeremy, had people not tried because it was like too complicated to actually get the data and like set up the training? Or like, were people just lazy and kind of like, hey, this is just not going to work? [00:15:20]
Jeremy: No, everybody wasn't lazy. So like, so the person I thought at that time who, you know, there were two people I thought at that time, actually, who were the strongest at language models were Stephen Merity and Alec Radford. And at the time I didn't know Alec, but I, after we had both, after I'd released ULM Fit and he had released GPT, I organized a chat for both of us with Kate Metz in the New York Times. And Kate Metz answered, sorry, and Alec answered this question for Kate. And Kate was like, so how did, you know, GPT come about? And he said, well, I was pretty sure that pre-training on a general large corpus wouldn't work. So I hadn't tried it. And then I read ULM Fit and turns out it did work. And so I did it, you know, bigger and it worked even better. And similar with, with Stephen, you know, I asked Stephen Merity, like, why don't we just find, you know, take your AWD-ASTLM and like train it on all of Wikipedia and fine tune it? And he's kind of like, well, I don't think that's going to really lie. Like two years before I did a very popular talk at KDD, the conference where everybody in NLP was in the audience. I recognized half the faces, you know, and I told them all this, I'm sure transfer learning is the key. I'm sure ImageNet, you know, is going to be an NLP thing as well. And, you know, everybody was interested and people asked me questions afterwards and, but not just, yeah, nobody followed up because everybody knew that it didn't work. I mean, even like, so we were scooped a little bit by Dai and Lee, Kwok Lee at Google. They had, they had, I already, I didn't even realize this, which is a bit embarrassing. They had already done a large language model and fine tuned it. But again, they didn't create a general purpose, large language model on a general purpose corpus. They only ever tested a domain specific corpus. And I haven't spoken to Kwok actually about that, but I assume that the reason was the same. It probably just didn't occur to them that the general approach could work. So maybe it was that kind of 30 years of mulling over the, the cell Chinese room experiment that had convinced me that it probably would work. I don't know. Yeah. [00:17:48]
Alessio: Interesting. I just dug up Alec announcement tweet from 2018. He said, inspired by Cobe, Elmo, and Yola, I'm fit. We should have a single transformer language model can be fine tuned to a wide variety. It's interesting because, you know, today people think of AI as the leader, kind of kind of like the research lab pushing forward the field. What was that at the time? You know, like kind of like going back five years, people think of it as an overnight success, but obviously it took a while. [00:18:16]
Swyx: Yeah. Yeah. [00:18:17]
Jeremy: No, I mean, absolutely. And I'll say like, you know, it's interesting that it mentioned Elmo because in some ways that was kind of diametrically opposed to, to ULM fit. You know, there was these kind of like, so there was a lot of, there was a lot of activity at the same time as ULM fits released. So there was, um, so before it, as Brian McCann, I think at Salesforce had come out with this neat model that did a kind of multitask learning, but again, they didn't create a general fine tune language model first. There was Elmo, um, which I think was a lip, you know, actually quite a few months after the first ULM fit example, I think. Um, but yeah, there was a bit of this stuff going on. And the problem was everybody was doing, and particularly after GPT came out, then everybody wanted to focus on zero shot and few shot learning. You know, everybody hated fine tuning. Everybody hated transfer learning. And like, I literally did tours trying to get people to start doing transfer learning and people, you know, nobody was interested, particularly after GPT showed such good results with zero shot and few shot learning. And so I actually feel like we kind of went backwards for years and, and not to be honest, I mean, I'm a bit sad about this now, but I kind of got so disappointed and dissuaded by like, it felt like these bigger lab, much bigger labs, you know, like fast AI had only ever been just me and Rachel were getting all of this attention for an approach I thought was the wrong way to do it. You know, I was convinced was the wrong way to do it. And so, yeah, for years people were really focused on getting better at zero shot and few shots and it wasn't until, you know, this key idea of like, well, let's take the ULM fit approach, but for step two, rather than fine tuning on a kind of a domain corpus, let's fine tune on an instruction corpus. And then in step three, rather than fine tuning on a reasonably specific task classification, let's fine tune on a, on a RLHF task classification. And so that was really, that was really key, you know, so I was kind of like out of the NLP field for a few years there because yeah, it just felt like, I don't know, pushing uphill against this vast tide, which I was convinced was not the right direction, but who's going to listen to me, you know, cause I, as you said, I don't have a PhD, not at a university, or at least I wasn't then. I don't have a big set of computers to fine tune huge transformer models. So yeah, it was definitely difficult. It's always been hard. You know, it's always been hard. Like I've always been somebody who does not want to build stuff on lots of big computers because most people don't have lots of big computers and I hate creating stuff that most people can't use, you know, and also stuff that's created on lots of big computers has always been like much more media friendly. So like, it might seem like a recent thing, but actually throughout my 30 years in data science, the attention's always been on, you know, the big iron results. So when I first started, everybody was talking about data warehouses and it was all about Teradata and it'd be like, oh, this big bank has this huge room full of computers and they have like terabytes of data available, you know, at the press of a button. And yeah, that's always what people want to talk about, what people want to write about. And then of course, students coming out of their PhDs and stuff, that's where they want to go work because that's where they read about. And to me, it's a huge distraction, you know, because like I say, most people don't have unlimited compute and I want to help most people, not the small subset of the most well-off people. [00:22:16]
Alessio: That's awesome. And it's great to hear, you do such a great job educating that a lot of times you're not telling your own story, you know? So I love this conversation. And the other thing before we jump into Fast.AI, actually, a lot of people that I know, they run across a new architecture and whatnot, they're like, I got to start a company and raise a bunch of money and do all of this stuff. And say, you were like, I want everybody to have access to this. Why was that the case for you? Was it because you already had a successful venture in like FastMail and you were more interested in that? What was the reasoning? [00:22:52]
Jeremy: It's a really good question. So I guess the answer is yes, that's the reason why. So when I was a teenager, I thought it would be really cool to like have my own company. You know, I didn't know the word startup. I didn't know the word entrepreneur. I didn't know the word VC. And I didn't really know what any of those things were really until after we started Kaggle, to be honest. Even the way it started to what we now call startups. I just thought they were just small businesses. You know, they were just companies. So yeah, so those two companies were FastMail and Optimal Decisions. FastMail was the first kind of synchronized email provider for non-businesses. So something you can get your same email at home, on your laptop, at work, on your phone, whatever. And then Optimal Decisions invented a new approach to insurance pricing. Something called profit-optimized insurance pricing. So I saw both of those companies, you know, after 10 years. And at that point, I had achieved the thing that as a teenager I had wanted to do. You know, it took a lot longer than it should have because I spent way longer in management consulting than I should have because I got caught up in that stupid rat race. But, you know, eventually I got there and I remember my mom saying to me, you must be so proud. You know, because she remembered my dream. She's like, you've done it. And I kind of reflected and I was like, I'm not proud at all. You know, like people quite liked FastMail. You know, it's quite nice to have synchronized email. It probably would have happened anyway. Yeah, I'm certainly not proud that I've helped some insurance companies suck more money out of their customers. Yeah, no, I'm not proud. You know, it's actually, I haven't really helped the world very much. You know, maybe in the insurance case I've made it a little bit worse. I don't know. So, yeah, I was determined to not waste more years of my life doing things, working hard to do things which I could not be reasonably sure would have a lot of value. So, you know, I took some time off. I wasn't sure if I'd ever work again, actually. I didn't particularly want to, because it felt like, yeah, it felt like such a disappointment. And, but, you know, and I didn't need to. I had enough money. Like, I wasn't super rich, but I had enough money. I didn't need to work. And I certainly recognized that amongst the other people I knew who had enough money that they didn't need to work, they all worked ridiculously hard, you know, and constantly put themselves in extremely stressful situations. And I thought, I don't want to be one of those idiots who's tied to, you know, buying a bigger plane than the next guy or whatever. You know, Kaggle came along and I mainly kind of did that just because it was fun and interesting to hang out with interesting people. But, you know, with Fast.ai in particular, you know, Rachel and I had a very explicit, you know, long series of conversations over a long period of time about like, well, how can we be the most helpful to society as a whole, and particularly to those people who maybe need more help, you know? And so we definitely saw the world going in a potentially pretty dystopian direction if the world's most powerful technology was controlled by a small group of elites. So we thought, yeah, we should focus on trying to help that not happen. You know, sadly, it looks like it still is likely to happen. But I mean, I feel like we've helped make it a little bit less likely. So we've done our bit. [00:26:39]
Swyx: You've shown that it's possible. And I think your constant advocacy, your courses, your research that you publish, you know, just the other day you published a finding on, you know, learning that I think is still something that people are still talking about quite a lot. I think that that is the origin story of a lot of people who are going to be, you know, little Jeremy Howards, furthering your mission with, you know, you don't have to do everything by yourself is what I'm saying. No, definitely. Definitely. [00:27:10]
Jeremy: You know, that was a big takeaway from like, analytic was analytic. It definitely felt like we had to do everything ourselves. And I kind of, I wanted to solve medicine. I'll say, yeah, okay, solving medicine is actually quite difficult. And I can't do it on my own. And there's a lot of other things I'd like to solve, and I can't do those either. So that was definitely the other piece was like, yeah, you know, can we create an army of passionate domain experts who can change their little part of the world? And that's definitely happened. Like I find nowadays, at least half the time, probably quite a bit more that I get in contact with somebody who's done really interesting work in some domain. Most of the time I'd say, they say, yeah, I got my start with fast.ai. So it's definitely, I can see that. And I also know from talking to folks at places like Amazon and Adobe and stuff, which, you know, there's lots of alumni there. And they say, oh my God, I got here. And like half of the people are fast.ai alumni. So it's fantastic. [00:28:13]
Swyx: Yeah. [00:28:14]
Jeremy: Actually, Andre Kapathy grabbed me when I saw him at NeurIPS a few years ago. And he was like, I have to tell you, thanks for the fast.ai courses. When people come to Tesla and they need to know more about deep learning, we always send them to your course. And the OpenAI Scholars Program was doing the same thing. So it's kind of like, yeah, it's had a surprising impact, you know, that's just one of like three things we do is the course, you know. [00:28:40]
Swyx: Yes. [00:28:40]
Jeremy: And it's only ever been at most two people, either me and Rachel or me and Sylvia nowadays, it's just me. So yeah, I think it shows you don't necessarily need a huge amount of money and a huge team of people to make an impact. [00:28:56]
Swyx: Yeah. So just to reintroduce fast.ai for people who may not have dived into it much, there is the courses that you do. There is the library that is very well loved. And I kind of think of it as a nicer layer on top of PyTorch that people should start with by default and use it as the basis for a lot of your courses. And then you have like NBDev, which I don't know, is that the third one? [00:29:27]
Jeremy: Oh, so the three areas were research, software, and courses. [00:29:32]
Swyx: Oh, sorry. [00:29:32]
Jeremy: So then in software, you know, fast.ai is the main thing, but NBDev is not far behind. But then there's also things like FastCore, GHAPI, I mean, dozens of open source projects that I've created and some of them have been pretty popular and some of them are still a little bit hidden, actually. Some of them I should try to do a better job of telling people about. [00:30:01]
Swyx: What are you thinking about? Yeah, what's on the course of my way? Oh, I don't know, just like little things. [00:30:04]
Jeremy: Like, for example, for working with EC2 and AWS, I created a FastEC2 library, which I think is like way more convenient and nice to use than anything else out there. And it's literally got a whole autocomplete, dynamic autocomplete that works both on the command line and in notebooks that'll like auto-complete your instance names and everything like that. You know, just little things like that. I try to make like, when I work with some domain, I try to make it like, I want to make it as enjoyable as possible for me to do that. So I always try to kind of like, like with GHAPI, for example, I think that GitHub API is incredibly powerful, but I didn't find it good to work with because I didn't particularly like the libraries that are out there. So like GHAPI, like FastEC2, it like autocompletes both at the command line or in a notebook or whatever, like literally the entire GitHub API. The entire thing is like, I think it's like less than 100K of code because it actually, as far as I know, the only one that grabs it directly from the official open API spec that GitHub produces. And like if you're in GitHub and you just type an API, you know, autocomplete API method and hit enter, it prints out the docs with brief docs and then gives you a link to the actual documentation page. You know, GitHub Actions, I can write now in Python, which is just so much easier than writing them in TypeScript and stuff. So, you know, just little things like that. [00:31:40]
Swyx: I think that's an approach which more developers took to publish some of their work along the way. You described the third arm of FastAI as research. It's not something I see often. Obviously, you do do some research. And how do you run your research? What are your research interests? [00:31:59]
Jeremy: Yeah, so research is what I spend the vast majority of my time on. And the artifacts that come out of that are largely software and courses. You know, so to me, the main artifact shouldn't be papers because papers are things read by a small exclusive group of people. You know, to me, the main artifacts should be like something teaching people, here's how to use this insight and here's software you can use that builds it in. So I think I've only ever done three first-person papers in my life, you know, and none of those are ones I wanted to do. You know, they were all ones that, like, so one was ULM Fit, where Sebastian Ruder reached out to me after seeing the course and said, like, you have to publish this as a paper, you know. And he said, I'll write it. He said, I want to write it because if I do, I can put it on my PhD and that would be great. And it's like, okay, well, I want to help you with your PhD. And that sounds great. So like, you know, one was the masks paper, which just had to exist and nobody else was writing it. And then the third was the Fast.ai library paper, which again, somebody reached out and said, please, please write this. We will waive the fee for the journal and everything and actually help you get it through publishing and stuff. So yeah, so I don't, other than that, I've never written a first author paper. So the research is like, well, so for example, you know, Dawn Bench was a competition, which Stanford ran a few years ago. It was kind of the first big competition of like, who can train neural nets the fastest rather than the most accurate. And specifically it was who can train ImageNet the fastest. And again, this was like one of these things where it was created by necessity. So Google had just released their TPUs. And so I heard from my friends at Google that they had put together this big team to smash Dawn Bench so that they could prove to people that they had to use Google Cloud and use their TPUs and show how good their TPUs were. And we kind of thought, oh s**t, this would be a disaster if they do that, because then everybody's going to be like, oh, deep learning is not accessible. [00:34:20]
Swyx: You know, to actually be good at it, [00:34:21]
Jeremy: you have to be Google and you have to use special silicon. And so, you know, we only found out about this 10 days before the competition finished. But, you know, we basically got together an emergency bunch of our students and Rachel and I and sat for the next 10 days and just tried to crunch through and try to use all of our best ideas that had come from our research. And so particularly progressive resizing, just basically train mainly on small things, train on non-square things, you know, stuff like that. And so, yeah, we ended up winning, thank God. And so, you know, we turned it around from being like, like, oh s**t, you know, this is going to show that you have to be Google and have TPUs to being like, oh my God, even the little guy can do deep learning. So that's an example of the kind of like research artifacts we do. And yeah, so all of my research is always, how do we do more with less, you know? So how do we get better results with less data, with less compute, with less complexity, with less education, you know, stuff like that. So ULM fits obviously a good example of that. [00:35:37]
Swyx: And most recently you published, can LLMs learn from a single example? Maybe could you tell the story a little bit behind that? And maybe that goes a little bit too far into the learning of very low resource, the literature. [00:35:52]
Jeremy: Yeah, yeah. So me and my friend, Jono Whittaker, basically had been playing around with this fun Kaggle competition, which is actually still running as we speak, which is, can you create a model which can answer multiple choice questions about anything that's in Wikipedia? And the thing that makes it interesting is that your model has to run on Kaggle within nine hours. And Kaggle's very, very limited. So you've only got 14 gig RAM, only two CPUs, and a small, very old GPU. So this is cool, you know, if you can do well at this, then this is a good example of like, oh, you can do more with less. So yeah, Jono and I were playing around with fine tuning, of course, transfer learning, pre-trained language models. And we saw this, like, so we always, you know, plot our losses as we go. So here's another thing we created. Actually, Sylvain Guuger, when he worked with us, created called fast progress, which is kind of like TQEDM, but we think a lot better. So we look at our fast progress curves, and they kind of go down, down, down, down, down, down, down, a little bit, little bit, little bit. And then suddenly go clunk, and they drop. And then down, down, down, down, down a little bit, and then suddenly clunk, they drop. We're like, what the hell? These clunks are occurring at the end of each epoch. So normally in deep learning, this would be, this is, you know, I've seen this before. It's always been a bug. It's always turned out that like, oh, we accidentally forgot to turn on eval mode during the validation set. So I was actually learning then, or, oh, we accidentally were calculating moving average statistics throughout the epoch. So, you know, so it's recently moving average or whatever. And so we were using Hugging Face Trainer. So, you know, I did not give my friends at Hugging Face the benefit of the doubt. I thought, oh, they've fucked up Hugging Face Trainer, you know, idiots. Well, you'll use the Fast AI Trainer instead. So we switched over to Learner. We still saw the clunks and, you know, that's, yeah, it shouldn't really happen because semantically speaking in the epoch, isn't like, it's not a thing, you know, like nothing happens. Well, nothing's meant to happen when you go from ending one epoch to starting the next one. So there shouldn't be a clunk, you know. So I kind of asked around on the open source discords. That's like, what's going on here? And everybody was just like, oh, that's just what, that's just what these training curves look like. Those all look like that. Don't worry about it. And I was like, oh, are you all using Trainer? Yes. Oh, well, there must be some bug with Trainer. And I was like, well, we also saw it in Learner [00:38:42]
Swyx: and somebody else is like, [00:38:42]
Jeremy: no, we've got our own Trainer. We get it as well. They're just like, don't worry about it. It's just something we see. It's just normal. [00:38:48]
Swyx: I can't do that. [00:38:49]
Jeremy: I can't just be like, here's something that's like in the previous 30 years of neural networks, nobody ever saw it. And now suddenly we see it. [00:38:57]
Swyx: So don't worry about it. [00:38:59]
Jeremy: I just, I have to know why. [00:39:01]
Swyx: Can I clarify? This is, was everyone that you're talking to, were they all seeing it for the same dataset or in different datasets? [00:39:08]
Jeremy: Different datasets, different Trainers. They're just like, no, this is just, this is just what it looks like when you fine tune language models. Don't worry about it. You know, I hadn't seen it before, but I'd been kind of like, as I say, I, you know, I kept working on them for a couple of years after ULM fit. And then I kind of moved on to other things, partly out of frustration. So I hadn't been fine tuning, you know, I mean, Lama's only been out for a few months, right? But I wasn't one of those people who jumped straight into it, you know? So I was relatively new to the kind of Lama fine tuning world, where else these guys had been, you know, doing it since day one. [00:39:49]
Swyx: It was only a few months ago, [00:39:51]
Jeremy: but it's still quite a bit of time. So, so yeah, they're just like, no, this is all what we see. [00:39:56]
Swyx: Don't worry about it. [00:39:56]
Jeremy: So yeah, I, I've got a very kind of like, I don't know, I've just got this brain where I have to know why things are. And so I kind of, I ask people like, well, why, why do you think it's happening? And they'd be like, oh, it would pretty obviously, cause it's like memorize the data set. It's just like, that can't be right. It's only seen it once. Like, look at this, the loss has dropped by 0.3, 0.3, which is like, basically it knows the answer. And like, no, no, it's just, it is, it's just memorize the data set. So yeah. So look, Jono and I did not discover this and Jono and I did not come up with a hypothesis. You know, I guess we were just the ones, I guess, who had been around for long enough to recognize that like, this, this isn't how it's meant to work. And so we, we, you know, and so we went back and like, okay, let's just run some experiments, you know, cause nobody seems to have actually published anything about this. [00:40:51]
Well, not quite true.
Some people had published things, but nobody ever actually stepped back and said like, what the hell, you know, how can this be possible? Is it possible? Is this what's happening? And so, yeah, we created a bunch of experiments where we basically predicted ahead of time. It's like, okay, if this hypothesis is correct, that it's memorized in the training set, then we ought to see blah, under conditions, blah, but not under these conditions. And so we ran a bunch of experiments and all of them supported the hypothesis that it was memorizing the data set in a single thing at once. And it's a pretty big data set, you know, which in hindsight, it's not totally surprising because the theory, remember, of the ULMFiT theory was like, well, it's kind of creating all these latent capabilities to make it easier for it to predict the next token. So if it's got all this kind of latent capability, it ought to also be really good at compressing new tokens because it can immediately recognize it as like, oh, that's just a version of this. So it's not so crazy, you know, but it is, it requires us to rethink everything because like, and nobody knows like, okay, so how do we fine tune these things? Because like, it doesn't even matter. Like maybe it's fine. Like maybe it's fine that it's memorized the data set after one go and you do a second go and okay, the validation loss is terrible because it's now really overconfident. [00:42:20]
Swyx: That's fine. [00:42:22]
Jeremy: Don't, you know, don't, I keep telling people, don't track validation loss, track validation accuracy because at least that will still be useful. Just another thing that's got lost since ULMFiT, nobody tracks accuracy of language models anymore. But you know, it'll still keep learning and it does, it does keep improving. But is it worse? You know, like, is it like, now that it's kind of memorized it, it's probably getting a less strong signal, you know, I don't know. So I still don't know how to fine tune language models properly and I haven't found anybody who feels like they do, like nobody really knows whether this memorization thing is, it's probably a feature in some ways. It's probably some things that you can do usefully with it. It's probably, yeah, I have a feeling it's messing up training dynamics as well. [00:43:13]
Swyx: And does it come at the cost of catastrophic forgetting as well, right? Like, which is the other side of the coin. [00:43:18]
Jeremy: It does to some extent, like we know it does, like look at Code Llama, for example. So Code Llama was a, I think it was like a 500 billion token fine tuning of Llama 2 using code. And also pros about code that Meta did. And honestly, they kind of blew it because Code Llama is good at coding, but it's bad at everything else, you know, and it used to be good. Yeah, I was pretty sure it was like, before they released it, me and lots of people in the open source discords were like, oh my God, you know, we know this is coming, Jan Lukinsk saying it's coming. I hope they kept at least like 50% non-code data because otherwise it's going to forget everything else. And they didn't, only like 0.3% of their epochs were non-code data. So it did, it forgot everything else. So now it's good at code and it's bad at everything else. So we definitely have catastrophic forgetting. It's fixable, just somebody has to do, you know, somebody has to spend their time training a model on a good mix of data. Like, so, okay, so here's the thing. Even though I originally created three-step approach that everybody now does, my view is it's actually wrong and we shouldn't use it. [00:44:36]
Jeremy: And that's because people are using it in a way different to why I created it. You know, I created it thinking the task-specific models would be more specific. You know, it's like, oh, this is like a sentiment classifier as an example of a task, you know, but the tasks now are like a, you know, RLHF, which is basically like answer questions that make people feel happy about your answer. So that's a much more general task and it's a really cool approach. And so we see, for example, RLHF also breaks models like, you know, like GPT-4, RLHDEFT, we know from kind of the work that Microsoft did, you know, the pre, the earlier, less aligned version was better. And these are all kind of examples of catastrophic forgetting. And so to me, the right way to do this is to fine-tune language models, is to actually throw away the idea of fine-tuning. There's no such thing. There's only continued pre-training. And pre-training is something where from the very start, you try to include all the kinds of data that you care about, all the kinds of problems that you care about, instructions, exercises, code, general purpose document completion, whatever. And then as you train, you gradually curate that, you know, you gradually make that higher and higher quality and more and more specific to the kinds of tasks you want it to do. But you never throw away any data. You always keep all of the data types there in reasonably high quantities. You know, maybe the quality filter, you stop training on low quality data, because that's probably fine to forget how to write badly, maybe. So yeah, that's now my view, is I think ULM fit is the wrong approach. And that's why we're seeing a lot of these, you know, so-called alignment tacks and this view of like, oh, a model can't both code and do other things. And, you know, I think it's actually because people are training them wrong. [00:46:47]
Swyx: Yeah, well, I think you have a clear [00:46:51]
Alessio: anti-laziness approach. I think other people are not as good hearted, you know, they're like, [00:46:57]
Swyx: hey, they told me this thing works. [00:46:59]
Alessio: And if I release a model this way, people will appreciate it, I'll get promoted and I'll kind of make more money. [00:47:06]
Jeremy: Yeah, and it's not just money. It's like, this is how citations work most badly, you know, so if you want to get cited, you need to write a paper that people in your field recognize as an advancement on things that we know are good. And so we've seen this happen again and again. So like I say, like zero shot and few shot learning, everybody was writing about that. Or, you know, with image generation, everybody just was writing about GANs, you know, and I was trying to say like, no, GANs are not the right approach. You know, and I showed again through research that we demonstrated in our videos that you can do better than GANs, much faster and with much less data. And nobody cared because again, like if you want to get published, you write a GAN paper that slightly improves this part of GANs and this tiny field, you'll get published, you know. So it's, yeah, it's not set up for real innovation. It's, you know, again, it's really helpful for me, you know, I have my own research lab with nobody telling me what to do and I don't even publish. So it doesn't matter if I get citations. And so I just write what I think actually matters. I wish there was, and, you know, and actually places like OpenAI, you know, the researchers there can do that as well. It's a shame, you know, I wish there was more academic, open venues in which people can focus on like genuine innovation. [00:48:38]
Swyx: Twitter, which is unironically has become a little bit of that forum. I wanted to follow up on one thing that you mentioned, which is that you checked around the open source discords. I don't know if it's too, I don't know if it's a pusher to ask like what discords are lively or useful right now. I think that something I definitely felt like I missed out on was the early days of Luther AI, which is a very hard bit. And, you know, like what is the new Luther? And you actually shouted out the alignment lab AI discord in your blog post. And that was the first time I even knew, like I saw them on Twitter, never knew they had a discord, never knew that there was actually substantive discussions going on in there and that you were an active member of it. Okay, yeah. [00:49:23]
Jeremy: And then even then, if you do know about that and you go there, it'll look like it's totally dead. And that's because unfortunately, nearly all the discords, nearly all of the conversation happens in private channels. You know, and that's, I guess. [00:49:35]
Swyx: How does someone get into that world? Because it's obviously very, very instructive, right? [00:49:42]
Jeremy: You could just come to the first AI discord, which I'll be honest with you, it's less bustling than some of the others, but it's not terrible. And so like, at least, to be fair, one of Emma's bustling channels is private. [00:49:57]
Swyx: I guess. [00:49:59]
Jeremy: So I'm just thinking. [00:50:01]
Swyx: It's just the nature of quality discussion, right? Yeah, I guess when I think about it, [00:50:05]
Jeremy: I didn't have any private discussions on our discord for years, but there was a lot of people who came in with like, oh, I just had this amazing idea for AGI. If you just thought about like, if you imagine that AI is a brain, then we, you know, this just, I don't want to talk about it. You know, I don't want to like, you don't want to be dismissive or whatever. And it's like, oh, well, that's an interesting comment, but maybe you should like, try training some models first to see if that aligns with your intuition. Like, oh, but how could I possibly learn? It's like, well, we have a course, just actually spend time learning. Like, you know, anyway. And there's like, okay, I know the people who always have good answers there. And so I created a private channel and put them all in it. And I got to admit, that's where I post more often because there's much less, you know, flight of fancy views about how we could solve AGI, blah, blah, blah. So there is a bit of that. But having said that, like, I think the bar is pretty low. Like if you join a Discord and you can hit the like participants or community or whatever button, you can see who's in it. And then you'll see at the top, who the admins or moderators or people in the dev role are. And just DM one of them and say like, oh, here's my GitHub. Well, here's some blog posts I wrote. You know, I'm interested in talking about this, you know, can I join the private channels? And I've never heard of anybody saying no. I will say, you know, Alutha's all pretty open. So you can do the Alutha Discord still. You know, one problem with the Alutha Discord is it's been going on for so long that it's like, it's very inside baseball. It's quite hard to get started. Yeah. Carpa AI looks, I think it's all open. That's just less stability. That's more accessible. [00:52:03]
Swyx: Yeah. [00:52:04]
Jeremy: There's also just recently, now it's research that does like the Hermes models and data set just opened. They've got some private channels, but it's pretty open, I think. You mentioned Alignment Lab, that one it's all the interesting stuff is on private channels. So just ask. If you know me, ask me, cause I've got admin on that one. There's also, yeah, OS Skunkworks, OS Skunkworks AI is a good Discord, which I think it's open. So yeah, they're all pretty good. [00:52:40]
Swyx: I don't want you to leak any, you know, Discords that don't want any publicity, but this is all helpful. [00:52:46]
Jeremy: We all want people, like we all want people. [00:52:49]
Swyx: We just want people who like, [00:52:51]
Jeremy: want to build stuff, rather than people who, and like, it's fine to not know anything as well, but if you don't know anything, but you want to tell everybody else what to do and how to do it, that's annoying. If you don't know anything and want to be told like, here's a really small kind of task that as somebody who doesn't know anything is going to take you a really long time to do, but it would still be helpful. Then, and then you go and do it. That would be great. The truth is, yeah, [00:53:19]
Swyx: like, I don't know, [00:53:20]
Jeremy: maybe 5% of people who come in with great enthusiasm and saying that they want to learn and they'll do anything. [00:53:25]
Swyx: And then somebody says like, [00:53:25]
Jeremy: okay, here's some work you can do. Almost nobody does that work. So if you're somebody who actually does the work and follows up, you will massively stand out. That's an extreme rarity. And everybody will then want to help you do more work. [00:53:41]
Swyx: So yeah. [00:53:41]
Jeremy: So just, yeah, just do work and people will want to support you. [00:53:47]
Alessio: Our Discord used to be referral only for a long time. We didn't have a public invite and then we opened it and they're kind of like channel gating. Yeah. A lot of people just want to do, I remember it used to be like, you know, a forum moderator. [00:54:00]
Swyx: It's like people just want to do [00:54:01]
Alessio: like drive-by posting, [00:54:03]
Swyx: you know, and like, [00:54:03]
Alessio: they don't want to help the community. They just want to get their question answered. [00:54:07]
Jeremy: I mean, the funny thing is our forum community does not have any of that garbage. You know, there's something specific about the low latency thing where people like expect an instant answer. And yeah, we're all somehow in a forum thread where they know it's like there forever. People are a bit more thoughtful, but then the forums are less active than they used to be because Discord has got more popular, you know? So it's all a bit of a compromise, you know, running a healthy community is, yeah, it's always a bit of a challenge. All right, we got so many more things [00:54:47]
Alessio: we want to dive in, but I don't want to keep you here for hours. [00:54:50]
Swyx: This is not the Lex Friedman podcast [00:54:52]
Alessio: we always like to say. One topic I would love to maybe chat a bit about is Mojo, modular, you know, CrystalLiner, not many of you on the podcast. So we want to spend a little time there. You recently did a hacker's guide to language models and you ran through everything from quantized model to like smaller models, larger models, and all of that. But obviously modular is taking its own approach. Yeah, what got you excited? I know you and Chris have been talking about this for like years and a lot of the ideas you had, so. [00:55:23]
Jeremy: Yeah, yeah, yeah, yeah, no, absolutely. So I met Chris, I think it was at the first TensorFlow Dev Summit. And I don't think he had even like, I'm not sure if he'd even officially started his employment with Google at that point. So I don't know, you know, certainly nothing had been mentioned. So I, you know, I admired him from afar with LLVM and Swift and whatever. And so I saw him walk into the courtyard at Google. It's just like, oh s**t, man, that's Chris Latner. I wonder if he would lower his standards enough to talk to me. Well, worth a try. So I caught up my courage because like nobody was talking to him. He looked a bit lost and I wandered over and it's like, oh, you're Chris Latner, right? It's like, what are you doing here? What are you doing here? And I was like, yeah, yeah, yeah. It's like, oh, I'm Jeremy Howard. It's like, oh, do you do some of this AI stuff? And I was like, yeah, yeah, I like this AI stuff. Are you doing AI stuff? It's like, well, I'm thinking about starting to do some AI stuff. Yeah, I think it's going to be cool. And it's like, wow. So like, I spent the next half hour just basically brain dumping all the ways in which AI was stupid to him. And he listened patiently. And I thought he probably wasn't even remember or care or whatever. But yeah, then I kind of like, I guess I re-caught up with him a few months later. And it's like, I've been thinking about everything you said in that conversation. And he like narrated back his response to every part of it, projects he was planning to do. And it's just like, oh, this dude follows up. Holy s**t. And I was like, wow, okay. And he was like, yeah, so we're going to create this new thing called Swift for TensorFlow. And it's going to be like, it's going to be a compiler with auto differentiation built in. And blah, blah, blah. And I was like, why would that help? [00:57:10]
Swyx: You know, why would you? [00:57:10]
Jeremy: And he was like, okay, with a compiler during the forward pass, you don't have to worry about saving context, you know, because a lot will be optimized in the backward. But I was like, oh my God. Because I didn't really know much about compilers. You know, I spent enough to kind of like, understand the ideas, but it hadn't occurred to me that a compiler basically solves a lot of the problems we have as end users. I was like, wow, that's amazing. Okay, you do know, right, that nobody's going to use this unless it's like usable. It's like, yeah, I know, right. So I was thinking you should create like a fast AI for this. So, okay, but I don't even know Swift. And he was like, well, why don't you start learning it? And if you have any questions, ask me. It's just like, holy s**t. Like, not only has Chris Latner lowered his standards enough to talk to me, but he's offering me personal tutoring on the programming language that he made. So I was just like, I'm not going to let him down. So I spent like the next two months, like just nerding out on Swift. And it was just before Christmas that I kind of like started writing down what I'd learned. And so I wrote a couple of blog posts on like, okay, this is like my attempt to do numeric programming in Swift. And these are all the challenges I had. And these are some of the issues I had with like making things properly performant. And here are some libraries I wrote. And I sent it to Chris and was like, I hope he's not too disappointed with me, you know, because that would be the worst. It's like, you know, and I was also like, I was like, I hope he doesn't dislike the fact that I, you know, didn't love everything. [00:58:46]
Jeremy: And yeah, he was like, oh, thanks for sending me that. Let's get on a call and talk about it. And we spoke and he was like, this is amazing. I can't believe that you made this. This is exactly what Swift needs. And he was like, and so like somebody set up like a new Swift, what they call them, the equivalent of a pep, you know, kind of RFC thing of like, oh, you know, let's look at how we can implement Jeremy's ideas and the language. And so it's like, oh, wow. And so, yeah, you know, and then we ended up like literally teaching some lessons together about Swift for TensorFlow. And we built a fast AI kind of equivalent with him and his team. It was so much fun. Then in the end, you know, Google didn't follow through, which is fair enough, like asking everybody to learn a new programming language is going to be tough. But like, it was very obvious, very, very obvious at that time that TensorFlow 2 is going to be a failure, you know, and so it's felt like, okay, I, you know, well, you know, what are you going to do? Like, you can't focus on TensorFlow 2 because it's not going to, like, it's not working. It's never going to work. You know, nobody at Google's using it. Internally. So, you know, in the end, Chris left, you know, Swift for TensorFlow got archived. [01:00:13]
Swyx: There was no backup plan. [01:00:15]
Jeremy: So it kind of felt like Google was kind of screwed, you know, and Chris went and did something else. But we kept talking and I was like, look, Chris, you know, you've got to be your own boss, man. It's like, you know, you've got the ideas, you know, like only you've got the ideas, you know, and if your ideas are implemented, we'd all be so much better off because like Python's the best of a whole bunch of s**t, you know, like I would, it's amazing, but it's awful, you know, compared to what it could be. And anyway, so eventually a few years later, he called me up and he was like, Jeremy, I've taken your advice. I've started a company. And I was like, oh my God. It's like, we've got to create a new language. We're going to create a new infrastructure. It's going to build, it's going to have all the stuff we've talked about. And it's like, oh wow. So that's what Mojo is. And so Mojo is like, you know, building on all the stuff that Chris has figured out over, I mean, really from when he did his PhD thesis, which developed LLVM onwards, you know, in Swift and MLIR, you know, the TensorFlow runtime engine, which is very good. You know, that was something that he built and has lasted. So yeah, I'm pumped about that. I mean, it's very speculative. Creating a whole new language is tough. I mean, Chris has done it before and he's created a whole C++ compiler amongst other things. Looking pretty hopeful. I mean, I hope it works because, you know, [01:01:53]
Alessio: You told them to quit his job. [01:01:55]
Swyx: So I mean, in the meantime, I will say, you know, [01:02:00]
Jeremy: Google now does have a backup plan, you know, they have Jax, which was never a strategy. It was just a bunch of people who also recognized TensorFlow 2 as s**t and they just decided to build something else. And for years, my friends in that team were like, don't tell anybody about us because we don't want to be anything but a research project. So now these poor guys, suddenly they're the great white hope for Google's future. And so Jax is, you know, also not terrible, but it's still written in Python. Like it would be cool if we had all the benefits of Jax, but in a language that was designed for those kinds of purposes. So, you know, fingers crossed that, yeah, that Mojo turns out great. [01:02:45]
Swyx: Yeah. [01:02:47]
Alessio: Any other thoughts on when, where people should be spending their time? So that's more the kind of language framework level. Then you have the, you know, GGML, some of these other like quantization focused kind of model level things. Then you got the hardware people. It's like a whole other bucket. Yeah. What are some of the exciting stuff that you're excited about? [01:03:08]
Jeremy: Well, you won't be surprised to hear me say this, but I think fine tuning transfer learning is still a hugely underappreciated area. So today's zero shot, few shot learning equivalent is retrieval augmented generation, you know, RAC, which is like, just like few shot learning is a thing. Like it's a real thing. It's a useful thing. It's not a thing anybody would want to ignore. Why are people not spending at least as much effort on fine tuning? You know, cause you know, RAG is like such a inefficient hack really, [01:03:45]
Swyx: isn't it? [01:03:45]
Jeremy: It's like, you know, segment up my data in some somewhat arbitrary way, embed it, ask questions about that, you know, hope that my embedding, you know, model embeds questions in the same bedding space as the paragraphs, which obviously is not going to, if your question is like, if I've got a whole bunch of archive papers embeddings, and I asked like, what are all the ways in which we can make inference more efficient? Like the only paragraphs it'll find is like if there's a review paper, here's a list of ways to make, you know, inference more efficient. Doesn't have any of the specifics. No, it's not going to be like, oh, here's one way, here's one way, here's a different way in different papers, [01:04:33]
Swyx: you know? Yeah. [01:04:35]
Jeremy: If you fine tune a model, then all of that information is getting directly incorporated into the weights of your model in a much more efficient and nuanced way. And then you can use RAG on top of that. So I think that that's one area that's definitely like underappreciated. And also the kind of like the confluence or like, okay, how do you combine RAG and fine tuning, for example. [01:05:00]
Swyx: Something that I think a lot of people are uncertain about, and I don't expect you to know either, is that whether or not you can fine tune new information in, and I think that that is the focus of some of your open questions. And of course you can, right? [01:05:17]
Jeremy: Like, obviously you can, because there's no such thing as fine, there's no such thing as fine tuning. There's only continued pre-training. So fine tuning is pre-training, like they're literally the same thing. So the knowledge got in there in the first place through pre-training. So how could like continuing to pre-train not put more knowledge in? Like it's the same thing. The problem is just we're really bad at it because everybody's doing it dumb ways. So, you know, it's a good question. And it's not just new knowledge, but like new capabilities. You know, I think like in my Packers Guide to LLM, into Packers Guide to LLM's talk, I show a simple, I mean, it's a funny, that's a simple example, because it doesn't sound it, but like taking a pre-trained based model and getting it to generate SQL. And it took 15 minutes to train on a single GPU. You know, I think that might surprise people that that capability is at your fingertips. And, you know, because it was already there, it was just latent in the base model. Really pushing the boundaries of what you can do with small models, I think is a really interesting question. Like what can you do with a, like, I mean, there isn't much in the way of good small models. A really underappreciated one is a BTLM 3B, which is a like kind of 7B quality 3B model. There's not much at the 1 to 2B range sadly, there are some code ones, but like the fact that there are some really good code ones in that 1 to 2B range shows you that that's a great size for doing complex tasks well. [01:06:56]
Swyx: There was PHY 1 recently, which has been the subject of a little bit of discussion about whether to train on benchmarks. [01:07:04]
Jeremy: PHY 1.5 as well. So that's not a good model yet. [01:07:09]
Swyx: Why not? [01:07:11]
Jeremy: It's good at doing, so PHY 1 in particular is good at doing a very specific thing, which is creating very small Python snippets. [01:07:19]
Swyx: The thing, okay, [01:07:21]
Jeremy: so like PHY 1.5 has never read Wikipedia, for example, so it doesn't know who Tom Cruise is, you know, it doesn't know who anybody is, it doesn't know about any movies, it doesn't really know anything about anything, like, because it's never read anything, you know, it was trained on a nearly entirely synthetic data set, which is designed for it to learn reasoning, and so it was a research project, and a really good one, and it definitely shows us a powerful direction in terms of what you can do with synthetic data, and wow, gosh, even these tiny models can get pretty good reasoning skills, pretty good math skills, pretty good coding skills, [01:08:04]
Jeremy: but I don't know if it's a model you could necessarily build on. Some people have tried to do some fine tunes of it, and again, they're like surprisingly good in some ways for a 1.5b model, but not sure you'd find it useful for anything. [01:08:24]
Swyx: I think that's the struggle of pitching small models, because small is great, you know, you don't need a lot of resources to run them, but the performance evaluation is always so iffy, it's always just like, yeah, it works on some things, and we don't trust it for others. [01:08:41]
Jeremy: Yeah, so that's why we're back to fine tuning. So Microsoft did create a 5.1.5 web, but they didn't release it, unfortunately. I would say a 5.1.5 web with fine tuning for your task, you know, might quite, you know, might solve a lot of tasks that people have in their kind of day-to-day lives. You know, particularly in kind of an enterprise setting, I think there's a lot of like repetitive kind of processing that has to be done. It's a useful thing for coders to know about, because I think quite often you can like replace some thousands and thousands of lines of complex buggy code, maybe with a fine tune, you know. [01:09:24]
Swyx: Got it. Yeah. [01:09:27]
Alessio: And Jeremy, before we let you go, I think one question on top of a lot of people's minds. So you've done practical deep learning for coders in 2018, 19, 21, 22. I feel like the more time goes by, the more the GPUs get concentrated. If you're somebody who's interested in deep learning today and you don't want to go join OpenAI, you don't want to join Anthropic, what's like the best use of their time? Should they focus on, yeah, small model development? Should they focus on fine tuning math and all of that? Should they just like focus on making Ragnar a hack and coming up with a better solution? Yeah, what's a practical deep learning for coders 2024 kind of look like? [01:10:10]
Jeremy: Yeah. [01:10:11]
Swyx: I mean, good question. [01:10:12]
Jeremy: I'm trying to figure that out for myself. You know, like what should I teach? Because I definitely feel like things have changed a bit. You know, one of the ways in which things have changed is that coding is much more accessible now. So if you look at a lot of the folks in the kind of open source LLM community, they're folks who really hadn't coded before a year ago. And they're using these models to help them build stuff they couldn't build before, which is just fantastic, you know? So one thing I kind of think is like, okay, well, we need a lot more material to help these people use this newfound skill they have because they don't really know what they're doing, you know, and they don't claim to, but they're doing it anyway. And I think that's fantastic, you know? So like, are there things we could do to help people, [01:10:58]
Swyx: you know, bridge this gap? [01:11:00]
Jeremy: Because previously, you know, I know folks who were, you know, doing manual jobs a year ago, and now they're training language models thanks to the help of Codex and Copilot and whatever. So, you know, yeah, what does it look like to like really grab this opportunity? You know, maybe Fast.ai's goals can be dramatically expanded now to being like, let's make coding more accessible, you know, kind of AI-oriented coding more accessible. If so, our course should probably look very different, you know, and we'd have to throw away that like, oh, you have to have at least a year of full-time programming, you know, as a prerequisite. Yeah, what would happen if we got rid of that? So that's kind of one thought that's in my head. You know, as to what should other people do? Honestly, I don't think anybody has any idea, like, the more I look at it, what's going on. I know I don't, you know, like, we don't really know how to do anything very well. Clearly OpenAI do, like, they seem to be quite good at some things, or they're talking to folks at, or who have recently left OpenAI. [01:12:17]
Swyx: Even there, it's clear there's a lot of stuff [01:12:19]
Jeremy: they haven't really figured out, and they're just kind of like using recipes that they've noticed have been okay, so, yeah, we don't really know how to train these models well, we don't know how to fine-tune them well, we don't know how to do React well, we don't know what they can do, we don't know what they can't do, we don't know how big a model you need to solve different kinds of problems, we don't know what kind of problems they can't do, we don't know what good prompting strategies are for particular problems, you know. Like, somebody sent me a message the other day saying they've written something that is a prompting strategy for GPT-4, for GPT-4, they've written, like, 6,000 lines of Python code, and it's to help it play chess. And then they've said they've had it play against other chess engines, including the best Stockfish engines, and it's got an ELO of 3,400, [01:13:11]
Swyx: which would make it close to [01:13:13]
Jeremy: the best chess engine in existence. And I think this is a good example of, like, people were saying, like, GPT-4 can't play chess. I mean, I was sure that was wrong. I mean, obviously, it can play chess. But the difference between, like, with no prompting strategy, it can't even make legal moves, with good prompting strategies, it might be just about the best chess engine in the world, far better than any human player. So, yeah, I mean, we don't really know what the capabilities are yet. So I feel like it's all blue sky at this point. It feels like computer vision in 2013 to me, which was, like, in 2013, computer vision was, like, OK, OK. [01:13:51]
Swyx: We just had the AlexNet. [01:13:52]
Jeremy: We've had AlexNet. We've had VGGNet. It's around the time Zyler and Fergus, like, no, it's probably before that. So we hadn't yet had the Zyler and Fergus, like, oh, this is actually what's going on inside the layers. So, you know, we don't actually know what's happening inside these transformers. We don't know how to create good training dynamics. We don't really know anything much. And there's a reason for that, right? And the reason for that is language models suddenly got really useful. And so the kind of economically rational thing to do, like, this is not criticism. This is true. The economic rational thing to do is to, like, OK, like, build that as fast as possible. You know, make something work, get it out there. And that's what, you know, OpenAI in particular did and Anthropic kind of did. But there's a whole lot of technical debt everywhere. You know, nobody's really figured this stuff out because everybody's been so busy [01:14:53]
Swyx: building what we know works as quickly as possible. [01:14:57]
Jeremy: So, yeah, I think there's a huge amount of opportunity to, you know, I think we'll find things can be made to work a lot faster, a lot less memory. I got a whole bunch of ideas I want to try, you know, every time I look at something closely, like really closely, I'm always like, oh, it turns out this person actually had no idea what they're doing, you know, [01:15:21]
Swyx: which is fine. [01:15:23]
Jeremy: Like, none of us know what we're doing. We should experiment with that. As we had a trade out on the podcast [01:15:32]
Alessio: who created FlashAttention. Yeah. And I asked him, did nobody think of using SRAM before you? Like, were people just like, no. And he was like, yeah, people just didn't think of it. They didn't try. They didn't come from like a systems background. [01:15:48]
Swyx: Yeah. [01:15:48]
Jeremy: I mean, the thing about FlashAttention is, I mean, lots of people absolutely had thought of that. So had I, right? But I mean, the honest truth is, particularly before Triton, like everybody knew that tiling is the right way to solve anything. And everybody knew that attention, fused attention wasn't tiled. That was stupid. But not everybody's got his ability to like, be like, oh, well, I am confident enough in CUDA and or Triton to use that insight to write something better, you know? And this is where, like, I'm super excited about Mojo, right? And I always talk to Chris about FlashAttention because I'm like, you know, there is a thousand FlashAttentions out there for us to build. You just got to make it easy for us to build them. Like Triton definitely helps, but it's still not easy. You know, it still requires kind of really understanding the GPU architecture and writing it in that kind of very CUDA-ish way. So yeah, I think, you know, if Mojo or something equivalent can really work well, we're going to see a lot more FlashAttentions popping up. [01:17:06]
Swyx: Great, Jerry. [01:17:08]
Alessio: And before we wrap, we usually do a quick lightning round. [01:17:10]
Swyx: We're going to have three simple questions. [01:17:13]
Alessio: So the first one is around acceleration. And you've been in this field a long time. What's something that it's already here today in AI that you thought would take much longer? I don't think anything. [01:17:24]
Jeremy: So I've actually been slightly too bullish. So in my 2014 TED talk, I had a graph and I said, like, this is like the slope of human capabilities and this is the slope of AI capabilities. And I said, oh, and I put a dot saying we are here. It was just before they passed. And I looked back at the transcript the other day and I said, in five years, I think we'll, you know, we might have crossed that threshold in which computers will be better at most human tasks than most humans or most average humans. And so that might be almost true now for non-physical tasks. So I was like, took, you know, took that twice as long as I thought it might. [01:18:11]
Jeremy: Yeah, no, I wouldn't say anything surprised me too much. It's still like, definitely like, I got to admit, you know, I had a very visceral reaction using GPT-4 for the first time. Not because I found it surprising, but actually doing it, like something I was pretty sure would exist by about now, maybe a bit earlier. But actually using it definitely is different to just feeling like it's probably on its way, you know, and yeah, whatever GPT-5 looks like. I'm sure, I imagine I'll have the same visceral reaction, you know. [01:18:56]
Swyx: It's really amazing to watch develop. We also have an exploration question. So what do you think is the most interesting unsolved question in AI? [01:19:07]
Jeremy: How do language models learn? You know, what are the training dynamics? Like I want to see, there was a great paper about ResNets a few years ago that showed how, that was able to like plot a kind of projected three-dimensional loss surface for a ConvNet with and without skip connections. And you know, you could very clearly see without the skip connections, it was bumpy, and with the skip connections, it was super smooth. That's the kind of work we need. Like, so there was actually an interesting blog post that came out just today from the PyTorch team where some of them have created this like 3D matrix product visualization thing. [01:19:56]
Swyx: The MatMul Visualizer. [01:19:58]
Jeremy: Yeah, and they actually showed some nice examples of like a GPT-2 attention layer and like showed an animation and said, like, if you look at this, we can actually see a bit about what it's doing. You know, so again, it reminds me of the Zeiler and Fergus, you know, ConvNet paper that was the first one to do these reverse convolutions to show what's actually being learned in each layer in a ConvNet. Yeah, we need a lot more of this, like, what is going on inside these models? How do they actually learn? And then how can we use those insights to help them to learn better? So I think that would be one. The other exploration I'd really like to see is a much more rigorous analysis of what kind of data do they need at what level? And when do they need it? And how often? So that kind of like dataset mixing, curation, so forth. [01:20:52]
Swyx: Right. In order to get the best capabilities. Yeah. How much is Wikipedia? Yeah. [01:20:58]
Jeremy: Yeah. [01:20:59]
Swyx: Very uncertain. [01:20:59]
Jeremy: Fine-tune what, you know, what kind of mix do you need for it to keep its capabilities? And what are the kind of underlying capabilities that it most needs to keep? And if it loses those, it would lose all these other ones. And what data do you need to keep those? And, you know, other things we can do to change the loss function, to help it to not forget to do things, stuff like that. [01:21:20]
Swyx: Awesome. [01:21:21]
Alessio: And yeah, before wrapping, what's one message, one idea you want everyone to remember and think about? [01:21:27]
Jeremy: You know, I guess the main thing I want everybody to remember is that, you know, there's a lot of people in the world. And they have a lot of, you know, diverse experiences and capabilities. And they all matter. And now that we have a, you know, newly powerful technology in our lives, we could think of that one of two ways. One would be, gee, that's really scary. What would happen if all of these people in the world had access to this technology? Some of them might be bad people. Let's make sure they can't have it. Or one might be, wow, of all those people in the world, I bet a lot of them could really improve the lives of a lot of humanity if they had this tool. This has always been the case, you know, from the invention of writing, to the invention of the printing press, to the, you know, development of education. And it's been a constant battle between people who think that the distributed power is unsafe and it should be held on to by an elite few. And people who think that humanity on net, you know, is a marvelous species, particularly when part of a society and a civilization. And we should do everything we can to enable more of them to contribute. This is a really big conversation right now. And, you know, I want to see more and more people showing up and showing what, you know, what the great unwashed masses out there can actually achieve. You know, that actually, you know, regular people are going to do a lot of really valuable work and actually help us be, you know, more safe and also flourishing in our lives and providing a future for our children to flourish in. You know, if we lock things down to the people that we think, you know, the elites that we think can be trusted to run it for us, yeah, I think all bets are off about where that leaves us as a society, you know. [01:24:00]
Alessio: Yep. Now that's an important message. And yeah, that's why we've been promoting a lot of open source developers, open source communities, I think, letting the builders build and explore. That's always a good idea. Thank you so much for coming on, Jeremy. This was great. [01:24:20]
Jeremy: Thank you for having me. [01:24:22]
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe