Answer.ai & AI Magic with Jeremy Howard

Key Takeaways

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 Smalley AI. And today we're back uh with Jeremy Howard, I think your third appearance on Latent Space. Welcome. Wait, third? Second. Well, I grabbed you in NeurIPS and I see. Very fun uh standing outside street heard that, by the way. You got to send me a link. I got to hear what it sounded like. Yeah, yeah, it I think the two hours are four or six hours. So there's there's plenty to to listen. We'll make sure to send it over. Yeah, we're trying this thing where uh the major ML conferences we, you know, do a little audio tour of of the conference and uh give people a sense of what it's like. Um but the last time you were on, you declared the end of fine-tuning. Uh I hope that I I know that I, you know, I I I sort of editorialized the title a little bit, and I know you were slightly uncomfortable with it, but you you just owned it anyway. Uh I think you're very good at the hot takes. Um and we were just discussing in our pre-show that um things have it's really happening that uh the continued pre-training is is really happening. Yeah, absolutely. Um 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 are now calling instruction tuning, and then I don't know if we've got a general term for this DPO RLHF step, you know, but you know, 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 like Llama 3, 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 like yeah, wasn't the end of pre-training. Sorry, it wasn't the end of fine-tuning, but more that we should treat it as a continuum and we should we should have much higher expectations of how much you can do with a already trained model. You can really add a lot of behavior to it. You can change its behavior. You can, you know, 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, cuz I I get very offended at the idea of starting from random weights. Yeah, I saw that um in ICLR in Vienna, there was a there was a outstanding paper about starting transformers from data-driven priors. I don't know if you saw that one. They they called it sort of never trained from scratch. And I think it was kind of rebelling against like the the sort of random initialization of it. Yeah, I've You know, that's been our kind of continuous message since we started fast.ai is if you're training from random weights, you better have a really good reason. You know, cuz 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. Yeah. 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 the 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 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 and 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 on 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, they should, you know, there's a a function or whatever which is happening every batch. And it doesn't like it's not a huge difference, but it's like we're back, 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 um, training mix. There's been a 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. Uh, it's just because you mentioned learning rate schedules. Uh, you know, what if you didn't have a schedule? I mean, I don't I don't care very much, honestly. Like, I don't think that schedule-free optimizers that exciting. It's fine. Um, you we've had non-scheduled optimizers for ages like um, less right who's now at Meta who was part of the fast AI community there created something called the Ranger optimizer. Um, you know, the um, I actually like having more hyper parameters. 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 in my atom, for example. I schedule all the things. Um, so, um, but then the other thing we always did with the fast.ai library was make it so you don't have to you don't have to set any schedules. So, fast.ai always supported like not 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. Um but to me I would like to have more parameters I can play with if I want to, but that you don't have to. And then the more less technical side I guess of your um issue I guess uh with the 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. Um and how we could maybe make it more available to people. And then after a month a month after we released the the episode there was the whole Sam Altman drama and like all the OpenAI governance issues. Um and maybe people started to think more okay, what happens if some of these kind of labs uh you know, start to break from within so to speak and the the alignment of the humans is probably going to fall before the alignment of the models. Um so I'm curious like if you have any new thoughts and maybe we can also tie in some of the the way that we've been building Answer as like a public benefit corp and um some of those aspects. 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 uh quite sure that OpenAI's current governance structure can't continue. Um 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?" I What did Jeremy see? I didn't see anything. It's just obviously true. Um and so yeah, so my friend Eric Reese 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 kind of start-up and AI governance and you know, we could both clearly see that this didn't make sense to have like a so-called nonprofit where then there are people working at a commercial company that's owned by or controlled nominally by the nonprofit 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 and I guess it's like I guess part of the problem is that the kind of people who work at nonprofits, you know, and in this case the board, you know, who are kind of academics and you know, people who 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, so yeah, Eric and I had been talking for a long time before that about like, "Well, what could be done differently?" Because also companies are sociopathic like 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 I wish our company didn't do that thing. Uh but you know, I know that if I didn't do it then I would just get fired and the board would have put in somebody else and the board knows if they don't do it then their shareholders can sue them cuz they're not maximizing profitability or whatever. So um what Eric's spent a lot of time doing is trying to think about like how do we make companies less sociopathic, you know? How do all more you know, maybe a better way to think of it is like how do we make it so that the you know the founders of companies can ensure that their companies continue to actually do the things they want them to do. Um so you know, when we started a company we you know, like well A, we very explicitly decided we're going to start a company. Not a academic lab, not a nonprofit, you know, we created a Delaware C corp, you know, the most company kind of company. Um but when we did so we told everybody, you know, including our first investors which was you LEO. It does sound great. We are going to run this company on the basis of maximizing long-term value. You know. Um so you know uh and in fact so when we did our our second round which was an angel round, we had everybody invest through a long-term SBV which we set up where everybody had to agree to vote in line with long-term value principles. Um so like it's not just it's it's never enough just to say to people like, "Okay, we're trying to create long-term value here for society as well as for ourselves." And everybody says like, "Good. 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 it it there's 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 uh 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 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, you know, is like uh if Philip Morris came along and said to you, "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 this you know, 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." Yep. And I I was looking back at um some of our emails. Um you you sent me an email on November 13th uh about talking and then on on the 14th I sent you an email uh working together to free AI was the the subject line. Um and then that was kind of the the start of the the seed round. And then 2 days later Sam Altman got fired. So, this was like not even 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, um yeah, you were very ahead of the ahead of the curve so to speak. I would love just to you know, people people can read your awesome introduction blog and answering the idea of having a R&D lab versus um our lab and then a D lab uh somewhere else. Uh 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 they don't fit the central casting of Silicon Valley so to speak. Like sometimes I got to 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 to for you to give a shout out to to some of the awesome folks that you have on the team. So, you know, there's a like a graphic going around describing like the people at xAI, you know, the Elon Musk thing. And like they're all connected to like, you know, multiple of Stanford, Meta, DeepMind, OpenAI, Berkeley, Oxford. It's just 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 another thing I found really interesting is um kind of anytime I almost anytime I see something which I think like this is really high quality work and it's like 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 like either they like came from poverty and like didn't get an opportunity to go to good school or they like, you know, had dyslexia and, you know, got kicked out of school in year 11 or you know, 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. And those are the people that throughout my career I've tended to kind of accidentally hire more of, but I like it's not exactly accidentally. It's like when I see somebody who's done two people who've done extremely well. One of them did extremely well in exactly the normal way from the background that 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, like that's normally the person I'm likely to find useful to work with. Cuz they're often like risk takers, they're often creative, they're often extremely tenacious. Um they're often very open-minded. So, that's the kind of folks we you know I tend to find myself hiring and I think like so now at Answer AI um 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 said and told me that they have imposter syndrome and they're not convinced that they're good enough to be here. You know, and I kind of got to 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 like 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 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. Um so you know, it's not reasonable to expect you're going to be better at everything than everybody else. Even though like 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 the 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. Um and it's been really nice to see like it's all a bit of an experiment like um we don't have any managers. Uh we don't have any hierarchy from that point of view. So, for example, I'm not a manager, which means I I don't get to tell people what to do or how to do it or when to do it. Um and it's been a Yeah, it's been a been an experiment to see how that would work out and it's been great like um So, for instance, um 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 he has he you know, this additional channel appeared on Discord on our on our private Discord called BERT 24. Like 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 and they're all these names that I recognize like BERT 24 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. Um cuz 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. Um and then like Benjamin Warner dived in and he's like, "Oh, I created a the whole Transformers from scratch implementation designed to be maximally hackable. Um he originally did it largely as a teaching exercise to show other people because like I could, you know, use that to create a really hackable bird implementation. Um In fact, he didn't say that. He said, "I just did do that." You know, uh 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 bird things, you know." Um 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 LaViolette reached out to me at one point and said like, "Okay, can you help me like what have you learned over time about how to manage, you know, intimidatingly capable and large groups of people who you're nominally meant to be leading?" Um And so, I you know, I like I try to help, but I don't direct. Um another great example was Karam, um uh who um after our FSDP 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. And like two or three weeks after we did FSDP QLoRA, he just popped up and said, "Okay, I've just converted the whole thing to DoRA, and I've also created these VL and extensions, and I've got all these benchmarks, and you know, now I've got um training of quantized models with adapters that are as fast as Laura and as actually better than weirdly fine-tuning. It's just like, "Okay, that's great, you know?" Um and yeah, so the things we've done to try to help make these things happen as well is like we have So, 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 uh Austin, who you know, amazing background. He ran AI at Fidelity, he ran AI at Pfizer, he ran browsing and retrieval for for Google's deep mind stuff, um credit gemini.cpp, and he's been working on a new uh system making it easier to do WebGPU programming cuz again, he quite correctly identified like, you know, this is a way that not everybody has to use CUDA, not everybody has to use NVIDIA, you can do stuff on your own computer optionally through the browser. We need to make this easier to do. And so, I yeah, so I said to him like, "Okay, I I want to learn about that." Not an area they have much expertise in, so yeah, 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 the um what Eric Gilliam, who wrote the second blog post in our series, the the R&D historian describes as everybody has total flexibility to do what they want, but we all understand like kind of roughly why we're here. You know, we all have the same You know, we agree with the premises around like you know, everything's too expensive, everything's too complicated, you know, 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 like we're all on the same wavelength about you know, all the ways in which current research is up and you know, all the ways in which you know, we kind of try you know, worried about centralization and we you know, um 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, but there's this kind of like shared vision. People understand like you know, so when they when I said like, "Oh, well, you know, tell me about about about 24. What's that about?" And he's like, you know, like, "Oh, well, you know, you can see from 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 um decoder-only models and you know, like BERT's used in all of these different places in industry. And so I can see like in terms of our basic principles or 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. 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 some of the traits? Like how do these things come out, you know, usually? Is it working on side projects that you you you you're already familiar with? Is there anything like in the interview process that like helps you screen for people that are more uh less pragmatic and more research-driven versus some of these folks that are like are just going to do it. You know, they're not waiting for like the perfect process. Everybody who comes through the recruiting is interviewed by everybody in the company. Um you know, our our our goal is 12 people. So, it's not an unreasonable amount. And like the way I So, the other thing to say is everybody so far who's come into the recruiting pipeline, everybody by one has been hired. So, which is to say our original curation has been good. Um And that's actually pretty easy cuz nearly everybody who's coming through the recruiting pipeline are people I I know pretty well. So, you know, Jonah I Whittaker and I you know, he worked on the stable diffusion course we did. Um he's outrageously creative and talented and he's super like enthusiastic tinkerer. Um just likes making things and um yep, Benjamin was one of the strongest parts of the fast.ai community, which is now the alumni is like hundreds of thousands of people. Um 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. Um Jonah was living in Zimbabwe. He was not you know, he was working on like helping some African startups, you know, but not Fang kind of credentials. Um But yeah, I mean, when you actually see people doing real work and they stand out above you know, the the We've got lots of Stanford graduates and OpenAI people and whatever in our alumni community as well. You know, when you stand out above all of above all of those people anyway, obviously, you've got something going for you. Um you know, Austen uh him and I worked together on the um mask study we did in the proceeding of the National Academy of Science. Uh so, 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 um uh research design and so forth and Austen was 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 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 Klavier actually, I didn't know before. But, you know, with him like I I just read his code and I'm like, "Oh, that's really well-written code." Like I 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. Felt like we'd known each other for years. Like we just were on the same wavelength and 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. Um Yeah, another example uh this guy named Vic who was previously the CEO of Dataquest. Um And like in that case like he's you know, he's created a really successful startup. He's like he won the the the first basically Kaggle NLP competition which was automatic essay grading. Um he's got the current state of the art OCR system, Syria. Um And 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. Um actually Kerem's another great example of this. I mean, I already knew Kerem very well because he was my best ever master student. But it wasn't my 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. You know, from scratch. This is not fine-tuning or whatever. He like went back to Common Crawl and did everything. So, 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. Yeah. Assemble based on talent. Yeah. Um we wanted to uh dive in a little bit more uh you know, turn turning from the people side of things into the technical bets that you're making. Uh just a little bit more on birds. Uh I I was actually we just did an interview with Etai from Reka. Uh I don't know if you're familiar with uh his work, but also another encoder-decoder bet. And um one of his arguments was actually people kind of over-index on the decoder-only GPT-3 type uh paradigm. I wonder if you have if you have thoughts there that is maybe not consensus as well. Yeah, no, absolutely. So, I think it's a great example. So, one of the people we're collaborating with a little bit with BERT 24 is um Colin Raffel, who is Original T5. Yeah. Yeah, most of that stuff. Um you know, between that and you all too, there's a lot of really interesting work. And so, one of the things I've been encouraging the BERT group to do, and Colin has as well, is to consider using a T5 pre-trained uh encoder backbone as a thing you fine-tune, which I think would be really cool. Um But uh he was saying, you know, Colin's also saying actually just using encoder-decoder as your BERT. You know, why don't you you like use that as a baseline, which I also think's a good idea. Yeah, it look Like, you know, what technical arguments are, you know, are people under under-weighting? I mean, Colin would be able to describe this much better than I can, but I'll I'll 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 U-Net. We you know, you you you have this kind of downward path, and then in the upward path you have the cross-connections, which you it's not attention, but it's like a similar idea, right? You you you're you're you're inputting the original encoding path into your decoding path. It's it's critical to make it work, right? Cuz otherwise, in the decoding part, the model has to like do so much kind of from scratch, right? 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 uh engineering, the right feature encoding for the original sentence. Um And it kind of means then on every token that you generate you have to recreate the whole 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 for your input information. Um So, I think there's really strong uh arguments for encoder-decoder models anywhere that there is this kind of like context or source thing, you know. Um And then why encoder only? Well, because like so much of the time what we actually care about is like, you know, a classification, you know, it's like an output. It's like we're not generating an arbitrary length sequence of tokens. So, anytime you're not generating an arbitrary length sequence of tokens uh decoder models don't seem to make much sense to me. Now, the interesting thing is you see on like Kaggle competitions that decoder models still are at least competitive with things like the BERT of V3. Um but they have to be way bigger to be competitive with things like the BERT of V3. Um and the only reason they are competitive is cuz people have put a lot more time and money and effort into training the decoder only ones, you know, there there isn't a recent de BERT uh there isn't a recent BERT. So, 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 that's the more interesting path, you know, but people tend to be like, oh, I I need to get citations. So, what's everybody else doing? Can I make it 0.1% better? You know, 0.1% faster. That's what everybody tends to do. Yeah, so I think it's like it tastes work. Commercially now is interesting cuz 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 um you know, GPT and Gemini and Claude. Um there should be a good commercial opportunity for them if they can figure out what those tasks are. Well, if rumors are to be believed, and he didn't comment on this, but you know, Snowflake may be figure out the commercialization for them, so we'll see. Good day. Let's talk about FSDP, QLoRA, Q-DorA, and all of that awesome stuff. What 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. Um 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 I can can we actually do it? Um and then obviously you all figured it out. Um can you maybe tell some of the war stories behind that? Like the the idea behind FSDP, which is kind of taking, you know, sharded um data parallel computation, and then QLoRA, which is do not touch all the weights, just go at the quantize some of the model, and then within the quantized model only do certain layers instead of doing everything. Well, do the adapters, yeah. Yeah, yeah, do the to the adapters. Um yeah, I I would 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. Yeah. I mean, to be honest, it It extremely unpleasant work to do. Um it's like not at all enjoyable. It's Um so I I kind of did version 0.1 of it myself before we had launched the company. Um or at least the kind of like the the the pieces, which is I just they're just they're all pieces that are difficult to work with, right? So for for the quantization, you know, I chatted to Tim Dettmers 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 Dettmers. You know, so so he wrote bits and bytes, which is his quantization library. And um you know, he wrote that for a paper. Um he didn't write that to be production like code. It's now like we're using it. Yeah. Yeah. So like it's not particularly well structured. There's lots of code paths that never get used. There's lots of you know, multiple versions of the same thing and you have to try to figure it out. So trying to get my head around that was hard. And you know, because it like the interesting bits are all written in CUDA, it's hard to like just step through it and see what's happening. Um and then, you know, FSDP is this very complicated library in PyTorch, which not particularly well documented. So the only really really way to understand it properly is again to just read the code and step through the code. And then um like bits and bytes doesn't really work in practice unless it's used with PEFT, the Hugging Face library, and PEFT doesn't really work in practice unless you use it with other things. And there's a lot of coupling in the Hugging Face ecosystem where like none of it works separately. You have to use it all together, which I don't love. Um 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. Um to kind of create this like minimal script. One thing that helped a lot was um Meta had this Llama recipes 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 FSDP LoRA didn't work with QLoRA on Llama. Actually, a lot of that had been put together. Like a lot of the stuff I discovered, the interesting stuff had been put together by Less Wright, who's uh he was actually the 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. Um So, yeah, I kind of that helped get some minimum stuff going. And then it was great once Benjamin and Johno joined full-time. And so, we basically hacked at that together. And then Kerem joined like a month later or something. Um but gee, it was just a lot of like fiddly, detailed engineering on like barely documented bits of obscure internals. Um 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 with you know, we needed to have good benchmarks, right? So, we'd like we'd we'd 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, in these like really important cases that the memory use is higher." You know, or it's actually slower. And we'd go in and we'd just find like all these things that had nothing to do with our library that just didn't work properly. And nobody had noticed they hadn't worked properly cuz nobody had really benchmarked it properly. So, we ended up you know, try 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 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, it doesn't. So, yeah, there's a load of war stories from 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, um to get the details right, to check them. Yeah. Yeah, we had Tri Dao 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 the nitty-gritty of it? 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. Yeah. Uh what about inference? You know, obviously you've done all this amazing work on fine-tuning. Um do you have any research you've been doing on the inference side, how to make local inference really fast on these models, too? We're doing quite a bit on that at the moment. We haven't released too much there yet, um but uh one of the things I've been trying to do is also just to help other people. Um and one of the nice things that's happened is that um couple of folks at at Meta including Max Serafin have done a nice job of creating this CUDA mode community of people working on like CUDA kernels or 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. Um 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. Also, the PyTorch team have created this torchao project on quantization. Um and so, there's a yeah, big overlap now between kind of the fastai and answer_ai and CUDA mode communities of people um working on stuff for both inference and fine-tuning, but um 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. Um 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 Karen's been doing a lot of work with with vllm for example. The these 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 hqq, which is a really great quantization library and works super well. Um So, yeah, there's a lot of other people outside answer_ai that we're working with a lot who are who are really helping on on all this performance optimization stuff open source. Just to follow up on on merging models, I I picked up there that you said nobody should be merging models. That's interesting because you know, I 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. Um to explain it's not that you shouldn't merge models, it should that you shouldn't be distributing a merged model. You should distribute it a merged adapter. Um 99% of the time. And actually often one of the best things happening in the model merging world is actually that off often merging adapters works better. The point is shown that 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 FP16 weights from FP from HBM memory at all or or ever load them off disk. Um 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. Got it. Got Okay, perfect. Um we're moving on a little bit to the rest of the fast universe. Um I had I would have thought that uh you know, once you started answering AI, that the sort of fast universe would be kind of on hold. Uh and then today you just dropped fast light and it looks like uh you know, there's there's more activity going on in sort of fast land. Yeah. So fast land and answer land are not really distinct things. Answer land is kind of like the fast-land 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. Uh and we want to do that with like 12 people. So, that's 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. Um People often forget about the D part of our R&D firm. So, we've got to be extremely good at, you know, creating, deploying, and maintaining applications, not just models. Much to my, you know, horror, the story around creating web applications is much worse now than it was 10 or 15 years ago in terms of like 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 like libraries like React and stuff and all the complex like details around security and web stuff around how you then talk to a back end and then all the details about creating the back end. You know, if that's your job, you know, and you know, you know, you have specialists who work in just one of those areas, it is possible to for that to all work, but compared to like I'll write a PHP script and put it in the home directory that you get when you sign up to this shell provider. Just what it was like in the '90s. You know, here are the 25 lines of code. Um and you're done. And now you can pass that URL around to all your friends. You know, I'll put this, you know, .py file inside the CGI-bin directory that you got when you signed up to this web host. Um So, yeah, the thing I've been mainly working on the last few weeks is fixing all that. And I I think I fixed it. Um So, I've created this thing called FastHTML. Yeah, sorry. Uh I don't know if this is an announcement, but I can I tell you guys. So, yeah, there's this thing called FastHTML, um which basically lets you create a complete web application in a single Python file. Um 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 Jinja, there's no separate like CSS and JavaScript files. Um it looks and behaves like a modern SPA web application. Um and you can create components for like DaisyUI or Bootstrap or Shoelace or whatever fancy JavaScript and or CSS, Tailwind, etc. library you like. Um 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. Um so you can like interactively do that. And then you mentioned Fast Light, so um specifically now if you're using SQL Light in particular, it's like ridiculously easy to have that persistence, you know, uh and you can basically all of your handlers will be passed database ready objects automatically um that you can just call {dot} delete, {dot} update, {dot} insert on. Um yeah, you get session, you get security, you get all that. So, it's it's it's again, like with the most of everything I do, it's very little code, it's mainly tying together really cool stuff that other people have written, so um um you don't have to use it, but a lot of the best stuff comes from its incorporation of HTMX, um which to me is basically the thing that changes your browser to make it work the way it always should have. So, it's a it's just does four small things, but those four small things are the things that are basically un- unnecessary constraints that HTML should never have had. So, it removes the constraints. Um uh it sits on top of Starlette, which is a very nice, you know, kind of lower-level platform for building these kind of web applications. The um the actual interface matches as closely as possible to Fast API, which is a really nice system for creating the kind of classic Java- JavaScript type applications. And uh Sebastian, who wrote Fast API, has been kind enough to help me think through some of these design decisions and so forth. Um I mean, everybody involved has been super helpful, actually. I chatted to Carson, who created HTMX, you know, also about it. Chatted to some of the folks involved in Django. Um 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 you know, pure Python framework. Um with a minimum of fast. So, yeah, I'm getting a lot of support trying to make sure that fast fast HTML works well for people. Uh yeah, I would say when I heard about this I I I just I just I texted Alex you I think this is going to be pretty huge. You know, like um people consider Streamlit and Gradio to be the state of the art, but I think there's so much to improve in uh you know, having a having sort of what do you say what do you call web phone web foundations or web fundamentals at the core of it I think it would be really helpful. I mean, it's based on 25 years of thinking and work for me. So, like fast mail was built on a system much like this one. Um 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. So, you know, and obviously lots of other people have done like great stuff and particularly HTMX, you know. So, I've been thinking about like, yeah, how do I pull together the best of the frame web framework I created for fast mail with HTMX. There's also things like Pico CSS which um is the CSS system which by default fast HTML comes with. Otherwise, I say you can pip install anything you want to, but it makes it like super easy to, you know, so try to make it so So, just out of the box you don't have any choices to make, you know, if you don't want to. You can make choices, but if 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 would 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 in this like whole separate language, but it's like this kind of smooth, gentle path that you can take step-by-step, cuz it's all just standard web foundations all the way, you know. Yeah, got it. Um well, so you know, just just to wrap up the sort of open source um work that you're doing, um you know, you're you're you're aiming to create thousands of projects with a with a very very small team. Um and I haven't heard you mention once AI agents or AI developer tooling or AI code maintenance. Um you know, please I I know you're very productive, but you know, what is the role of AI in your own work? So, I'm making something. Ooh. I'm not sure how much I want to say just yet. Uh Okay. All right, I'll give you the key thing. So, I've created a new uh approach. It's not called prompt engineering, it's called dialogue engineering. Um and I'm creating a system for doing dialogue engineering. Um um It's called AI magic. Um 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. Um Think about chat GPT with code and code interpreter, right? Um The 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 like hit enter and it would scroll up and then the answer from APL would would be printed out and would scroll up and then you'd type the next thing and like uh which is also the way for example um a shell works like bash or zsh whatever. Um It's it's not terrible you know? Like we all get a lot done in these like very very basic teletype style REPL environments. Um but I've never felt like it's optimal, you know? And to me um you know, so and and everybody else has just copied chat GPT. So it's also the way Bard 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 um Oh, you want to do more than ChatGPT? 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 and Cursor's like, uh create a form that blah blah blah. But it's 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 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 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 you know, the 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 for Claude. I've created something called Cosette, which is a library for OpenAI. Um they're libraries which are designed to make those APIs much more usable, much easier to use, much more concise, um and then I've written AI magic on top of those. Um And that's been an interesting exercise because uh I did Claude at first. And rather than try to like uh 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 Claude at was. So, for example, one of the really nice things in Claude is prefill. Uh 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. Um so, yeah, I kind of Claude at to be as Claude-friendly as possible. And then after I did that, um and then with Claude with particularly with GPT-4o 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 how can I make 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?" Um 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. Uh the name is based on the classic Polya book, if you know how to solve it, which is um you know, one of the classic math books of all time. Um 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. Um and it's particularly for this like whole generation of people who are learning to code with and because of ChatGPT. Like I know a lot of 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 cuz they don't really know how to code. So, this course will be designed to show you how you can like you know, 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. Um so, yeah. Nice. That's a That's a great spoiler. You know, I guess the fourth time you're going to be on Latent Space, we're going to talk about AI magic. Uh Jeremy, before before we wrap, um this was a just a great run through everything. Um what are the things that when you next come on the podcast in 9 to 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? Um you know, what's the next company that's going to fall like a drama internally? Anything? hopefully we'll be talking a lot about Fasts HTML and hopefully the international community that at that point has come up around that. And also about AI magic and about dialogue engineering. Um hopefully dialogue engineering catches on cuz I think it's the right way to think about a lot of this stuff. What else? Um just trying to think about it 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. Um There's a state space model that came out today that uh might be interesting for just general discussion. Um one thing that stood out to me with Cartesia's blog post was that they were talking about real-time ingestion of billions and trillions of tokens. Um and keeping that context obviously in the in the state space that they have. Uh Yeah. I'm wondering what your thoughts are because you you've been entirely focused the whole time. Yeah, no. So obviously my background is RNNs and LSTMs. Um and I'm still a believer in the idea that state is something you can update, you know? So obviously um Set Pokraka came up kind out with uh xLSTMs recently. Um Oh my god. Okay, another whole thing we haven't talked about just somewhat related. Uh I've gone I've been going crazy for like a long time about like why can I not pay anybody to save my KV cache, you know, for like 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, you know? So Gemini is about to finally come out with um KV cache and this is something that Austin actually in Gemma.cpp had on his road map for years. Um well, not years, months, long time. Um is is is that the idea that the KV cache is like a thing that like it's 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 uh KV cache creation. Um I think it creates like a whole new class almost of applications or of of of of of techniques where you know, for me, for example, I very often like 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 of my new library to be there all the time. So, yeah, I want to upload them once and then all of have a whole discussion about building this application using fastHTML. Well, nobody's got fastHTML in there in there uh 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 um have the documentation of the libraries you're working on where you can have always available. So, there'll be ways to uh you know, 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 and how to use state. Um cuz in state models and XLSTM, again, state is something you you update. So, how do we combine the best of all of these worlds? And Jeremy, I know before you talked about how some of the other aggressive 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? In the same way that like we probably ought to have state that you can update, i.e. XLSTM and state models, in the same way um a lot of things probably should have an encoder. Jepa and diffusion both seem like the right conceptual for a lot of things we properly want to do. So, the idea of like there there should be a a a a piece of the generative pipeline which is like thinking about the answer and coming up with a sketch of to what the answer looks like before you start outputting tokens. That's where I kind of feels like diffusion ought to fit. You know, and diffusion is because it's not autoregressive Mhm. it's like let's try to like gradually deblur 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. Mhm. Um so, yeah, there's a lot of different pieces here and uh 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. Awesome, Jeremy. This was great, um as usual. Um thanks again for coming back on the pod. Thank you, Leslie. And thank you for for listening. Yeah, that was fantastic.