Interconnects

Ai2 released OLMoE, which is probably our "best" model yet relative to its peers, but not much has changed in the process.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/olmoe-and-building-better-llms

00:00 OLMoE and the hidden simplicity in training better foundation models
02:04 Frontier model team compute allocations
04:19 De-risking training complexity
06:40 On organizational complexity
09:05 Compounding improvements -- the key to building better language models

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_005.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_007.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_009.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_011.png
Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_028.png
Fig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_030.png
Fig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmoe/img_032.png

The Open Source Initiative is working towards a definition.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/defining-open-source-ai

0:00 On the current definitions of open-source AI and the state of the data commons
3:17 Reasons to not mandate fully released data
4:24 Sufficient but not exhaustive data docs
5:22 Frustration with the data commons
7:04 We need more examples to define the definition

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/defining-open-source/img_005.png

The latest model from one of the most popular fine-tuning labs makes us question how a model should be identified as a "frontier model."
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/nous-hermes-3

0:00 Nous Hermes 3 and exploiting underspecified evaluations
5:29 Parsing training lessons from Hermes 3

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_005.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_010.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_012.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_020.png
Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_027.png
Fig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_030.png
Fig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_032.png
Fig 8: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_036.png

I had the pleasure of Talking with Ross Taylor, who has a great spectrum of unique experiences in the language modeling space — evaluation experience, Galactica lead author, Llama post training, etc. This is a really great conversation on the frontier of language model (LM) reasoning, LM deployments and demos, LM’s for science, RLHF, and other topics. I’ve been trying to get Ross to come on for a bit. He’s one of those people in the LM space that doesn’t speak too much, but when you do, you listen.

Ross Taylor was previously an LLM lead at Meta AI, heading up the reasoning team. Previously he led the early work on LLM agents, and was the research lead on the Galactica project. Before that, he was a co-founder of Papers with Code, which was acquired by Meta in 2019. Before that, he has worked as a quant in sports betting and finance, and before that a policy advisor for the UK Government. He is currently working on a new startup.

Listen on Apple Podcasts, Spotify, and where ever you get your podcasts. For other Interconnects interviews, go here.

YouTube

Chapters

* [00:00:00] Introduction of Ross Taylor and his background

* [00:02:12] Papers with Code

* [00:09:58] Galactica, goals, controversy, legacy

* [00:18:12] Technical details of the Galactica model

* [00:23:18] Potential for language models to make scientific discoveries

* [00:25:21] Defining and improving reasoning in language models

* [00:32:38] Process-based reward models and their potential applications

* [00:35:00] Generating synthetic data for SFT

* [00:40:23] Evaluating the effectiveness of language models as judges for human preference data

* [00:42:43] Considerations for creating base models that are easy to fine-tune

* [00:46:45] Balancing SFT and RLHF

* [00:54:13] Characteristics of successful post-training teams

* [00:58:26] Future directions for language model development

We mention

* Galactica

* Papers with Code

* Rob Stojnic (co-founder of Papers with Code)

* DPO, PPO

* Armen Aghajanyan (Chameleon)

* Tom Scialom on Latent Space

* Soumith Chintala (PyTorch)

* Alex Graves

* Llama 3 paper

* Process Reward Models / Let’s Verify Step by Step

Transcript

Built with smol-podcaster and with love of Latent Space.

Nathan Lambert [00:01:07]: Today, we're here with Ross. This is a really exciting one. I've been trying to get Ross on the show for a while. Ross has done a lot of interesting work. And also the path to where you ended up with working on state-of-the-art LLaMA work at Meta is very interesting to me. So we're going to start with some of that, but then there are a few people that want to know more about reasoning and some of the RLHF stuff. We won't cover the secretive new start-up - I don't know what it is, but that's how it goes these days. I'm sure it'll be great. So welcome to the show!

Ross Taylor [00:01:41]: Thanks for having me.

Nathan Lambert [00:01:44]: So I wanted to start with Papers with Code. For people that don't know, Papers with Code is one of these platforms - I never was a heavy user of it - but it collates papers, people can upvote them, popular papers, attaching code and dataset and evaluations to papers, which is great - it was like sort of ahead of its time. It fits into a lot of these open ecosystem things. So I'm kind of curious, like, how you ended up there and why you all started this startup that ended up building this thing that got acquired by Meta?

Ross Taylor [00:02:12]: Yeah, that was a weird one. This was like back in 2018. So I was at an incubator, I just quit my previous job and I was like, okay, I want to do a startup. And I met Rob, my co-founder, who came along with me for the journey. We both came from different backgrounds. I was from a sports betting / quant finance kind of background, which is a whole other episode I guess. And Rob was in various startups, like applying ML to things like hate speech detection, that kind of stuff. And the cool thing was, we both resonated on similar kinds of problems within the ML space, even though we came from different domains. So we spent a lot of time doing various experiments, trying to make new kinds of ML tooling, thinking of these stupid questions like “what is the Git equivalent for ML?” - that kind of stuff. One of those experiments was hacking around on this little website to solve a really basic problem: I'm trying to reproduce this paper, but I can't find the code. That was the thing that really blew up beyond our expectations. It was weird because we thought it was fairly trivial at first.

Nathan Lambert [00:03:16]: What year was this? 2018?

Ross Taylor [00:03:18]: Yeah.

Nathan Lambert [00:03:19]: This makes sense. I think this was like, I was starting Deep RL then, but Deep RL was so hot, which was like the worst evaluation has ever been probably for ML. Like people complain about it today, but like Deep RL evaluation was like, every single person was just lying to make themselves look better.

Ross Taylor [00:03:38]: The interesting thing now is that the open ecosystem has shifted to focus more on weights as a central artifact rather than code. I think there's an interesting debate there. Would it be more useful to have the LLaMA-3 8B model weights or all the code for training LLaMA-3? I think there's still interesting debates to be had about what's actually useful.

Nathan Lambert [00:03:56]: I think the code would be more useful. Like OpenAI released their rules-based reward models, but it's like code washing because it's like just a bunch of people just released like eval code now. And it's like, that's a whole another tier is like actual training code versus eval code. But yeah, I guess I'll just skip ahead.

Ross Taylor [00:04:12]: So essentially Papers with Code was the thing that didn't die for us. We always thought we were going to make something else and Papers with Code was more of a marketing thing. But eventually we were like: okay, our users are telling us this is what we should be working on. And we expanded from that very simple use case of finding code towards indexing various artifacts in ML.

Another big problem was trying to find the state of the art in something like ImageNet and all these different benchmarks. There just wasn't a central place to find this information…So we had this quite good Christmas - me and Robert - where we hacked for the whole month, indexing every leaderboard we could and all the related papers. I didn't want to do any annotation again after that! But that took things to the next tier, and that's when things really started to blow up.

Nathan Lambert [00:05:03]: Because this is like the first round of leaderboards, because now it's really popular with Hugging Face again. And I was like, yeah, is that just because it became like a Meta thing and it's just kind of a thing that existed? You're like the first leaderboard company in a way, which I don't think many people think about. Yeah, which is weird.

Ross Taylor [00:05:19]: Yeah. And the interesting thing about us was that we never had to do any marketing because everything was from organic traffic. So you would type in “state of the art ImageNet” and we would come to the top as the most useful site. That was really the source of our growth, and we grew to a million MAU fairly quickly. And as for Meta, we were in touch with the PyTorch folks at the time who we really liked. You know - Soumith, Joe - those folks, and they had a shared interest in promoting the open source ecosystem back in 2018/19. And while it was like a tough decision, we were just like “we really like working with these people, we want to work more closely with them”, and that got us into Meta.

And then within Meta, we originally continued to develop the platform. But the big shift for us was that, even then, we saw we were moving to a world where compute was the currency. And we saw that, if we wanted to be well positioned in five years time, we needed to be building these large-scale systems. Even for our own platform, we had lots of ML in the backend and we saw we were using fewer and fewer models to do more and more tasks. So that kind of shifted us into research, into Galactica, and then eventually LLaMA and that kind of stuff.

It was a weird shift because we were product people who ended up doing hardcore research! But I guess it was natural to us that we were within a research org with these amazing people, lots of resources. It was just the best use of our time to conduct this shift.

Nathan Lambert [00:06:43]: Do you think there should have been more integration between Hugging Face and Papers with Code? It would have been wonderful if it had happened.

Ross Taylor [00:06:54]: The backstory is that we saw them as competitors, to be honest, because we had the same vision originally. We were going to do model hosting, that kind of stuff. But we never got into it because we hit friction with leadership - who was not onboard with that as a goal. Because from their point of view, it's like, okay, if we host these things, this might expose Facebook to some kind of legal risk. It wasn't in the perceived interest of the company.

Nathan Lambert [00:07:17]: This is a classic story of tech, really. They can't take the risk. They can't expose themselves.

Ross Taylor [00:07:23]: If you're a startup and it's your number one priority, then yeah, your attitude on risk is different. But I think it was a blessing in disguise for us because clearly the bigger wave was going to be large language models - we saw that incredibly early. And our mission was fundamentally not infrastructure, but something closer to: how do you organize information? It was a Google-y type of mission. And while we were focused on ML, we were more broadly thinking about science: how do we reduce friction for finding out about new advances and, I guess, lots of small tasks that when added up lead to a lot of progress in science.

Nathan Lambert [00:07:59]: I should have probably looked this up. Did you have another scientific background? Did you have a hard science background or what about Rob? Stojnic?

Ross Taylor [00:08:10]: Yeah, [Robert] Stojnic, my co-founder, he was from a bio background. So he's actually-

Nathan Lambert [00:08:15]: That makes sense.

Ross Taylor [00:08:16]: Well, he also had a computer science background. He was one of the original developers of Wikipedia, so he has his own crazy story…

Nathan Lambert [00:08:22]: Yesterday I was talking to somebody that was one of the original arXiv moderators. So we're digging all these things up…

Ross Taylor [00:08:29]: It is interesting because we both had this background, I would say, in building useful “utilities” [on the internet] at some point in our lives. I think Papers with Code is one of those things which is easy to forget, but if it went away, everyone would go crazy.

As for me, my background is more statistics and econometrics. My first job was in the Government, which I kind of hated. But I did a Master's degree, which I thought was going to be in economics, but the thing I ended up loving was time series and statistics. So I did all this research on state space models - before it was cool, I guess! - and then that got me into sports betting. And then eventually, we were using more and more deep learning [in the 2010s], and that’s how I got into AI. So a fairly nonlinear path. But -

Nathan Lambert [00:09:09]: Yeah. Well back to what you were saying on the scientific stuff, I think the Galactica story has many angles, and you led on this.

I think if people go look at the paper, it's a very interesting paper, like you cite Galileo in the first sentence, and it really has a lot of early modern language model features and quirks. It's something that people don't remember that well.

I'm very on the record saying the backlash was overblown. I think that was before there were clear habits and community norms around what language model demos should look like. So it was kind of in that teething phase.

But what was the actual goal that you wanted? You mentioned organizing the world's information. What was the goal and how close do you think the model came to accomplishing it?

Ross Taylor [00:09:58]: So there were several different things at once.

There were immediate product integrations we had in mind. We actually had an agreement at the time with Overleaf to be a “co-pilot for writing papers”. We'd have a really good LaTeX model in Overleaf, and whenever you wanted to include a citation, you could simply prompt for one.

More broadly, we imagined the future would be instead of..using more classical ways to find and extract information, if you wanted to learn about something like DPO, you would just prompt a language model to find out about it. Or if you wanted to ask “What's the state-of-the-art on SWE-Bench?” or something like that, you would just prompt the model and it would find the relevant information and answer the question.

Nathan Lambert [00:10:46]: So this is something that language models are so bad at. One of my challenge questions - I've been doing this for 6-12 months - is to ask models about DPO, and none of the models without internet access have yet done it right. You would think that it would start to kick in. And I don't just ask “what is DPO?”, I ask “What is DPO for language model fine tuning”, and they still just make up nonsense.

Ross Taylor [00:11:06]: Yeah, which actually relates to an interesting debate about LLM creativity. If you want to solve something like LLM creativity, you want to be confident about the frontier of knowledge, but frontier knowledge is where you have the most token scarcity.

But anyway, just to finish that thought. Bear in mind, we were developing Galactica while the whole Web 3.0 boom was happening. And we were in this weird state where we were like “All everyone is talking about is Web 3.0, but clearly generative AI is going to be the thing that powers the next generation of the web!”. So I guess that was our primary motivation.

Now, in terms of the [Galactica] launch, I think there's two aspects.

First, like you said, the paper. Now we were a small team of 7-8 people. We had so much fun developing these new ideas at the time: internal reasoning tokens, how do language models cite, training for multiple epochs…

Nathan Lambert [00:12:00]: What's that? A citation token? Did you have a special token for citations?

Ross Taylor [00:12:04]: Yeah. So we had a start citation token [START_REF], and we used two methods. The first was: we'd put the title of the paper within the citation tags. And the other one was: we'd have an alphanumeric ID.

The interesting thing was, it actually worked really well - but in the demo interface, it had a tendency to hallucinate - or “hallucitate”. The backstory is that, while the model was really good, for the demo we turned up the temperature to 0.7 so the text generation was better [at the expense of citation accuracy]. So generative citations were something that people thought didn’t work, but it was [more an implementation issue]. I guess that’s an alternative road in history…

So there was the paper, which was cool, and there was the demo, which I would say was motivated by the realities of the time. This was pre-ChatGPT and, even within a big company like Meta, it wasn’t a company priority to work on LLMs at all. So in our mind, our objective was - we were kind of deluded - being a team of 7-8 people, we were like…

Nathan Lambert [00:13:08]: This is how you have to operate if you want to be at the cutting edge. That's how great teams operate.

Ross Taylor [00:13:13]: So there were two objectives you could have had. The first is: you think that second-mover advantage is good. So you could wait for OpenAI to do something and then come in after and do it in an open way. And this is the path that actually worked for LLaMA. LLaMA was not state-of-the-art in any sense.

Nathan Lambert [00:13:27]: I've been doing this. I mean six months ago, maybe OpenAI and Google wouldn’t need to hire me because they know everything. But now I’m doing more interesting analysis where I'd be hired at a different role - but in the open. Now I'm like the person people look at. But I’m trying to tell people that “You don't understand! I'm six months behind everyone!”.

Ross Taylor [00:13:49]: Right, but to be clear, that’s a really important role - because everyone should have a stake in the future. And that's what the open ecosystem gives people.

But our objective was this: we didn't want to be second; we wanted to be first. And we were kind of deluded because we were 8 people - compared to maybe OpenAI with 200 people where their whole bread and butter was language models. But that’s why we were thinking “how do we move as fast as possible?”. And in our mind, a demo might be premature, but it would also be a way to get lots of prompts and information quickly - to understand how people would be using the model. And essentially the calculus we took was, we knew the community might not be  ready for something like this - especially with the Meta branding - but we thought this was a way to get lots of information really fast and catch up given our position. Now in retrospect, history says that…

Nathan Lambert [00:14:33]: You kind of did that. I think Meta probably got the injection of language model reality from that. It's kind of like the Gemini backlash. I think the Gemini backlash - while it's obviously stupid execution - was potentially a good forcing function for Google's structure of their Gemini org - to really move everything into the way it is now. That made them be structured more like a serious language modeling org and less like Google, I think, which people don't want to hear...

Ross Taylor [00:15:07]: For us it was just a risk we decided to take. We probably took a lot more risk than we should have done. But we just thought “obviously this is going to be huge”, “LLMs are going to power the next internet”, etc, so let's take a risk. And you know, if we ran the universe several times over - it would have succeeded in some of those runs. But [in our universe], the criticism, which was obviously overblown, reached a critical point where things didn’t work out.

And then there's the story about the demo coming down, which - I’m not sure I’m able to talk about - but I think that is one of the things where, if people knew the true reasons, they'd be like “what the f**k!?”. But yeah, that's what happened…

Nathan Lambert [00:15:44]: Yeah, this is why any company that makes a demo now has block lists, where there's certain words that if they're in the prompt of the generation, you get a really, really stupid response. Even if it's like an open model, you just put like a little filter that's like, “you can't say the most obviously bad words”.

Ross Taylor [00:16:01]: But we actually did that and that created backlash as well. Because if you have false positives, you actually exclude some words which aren't actually offensive [in certain contexts], right? And then you also offend people… so it's not a win-win situation.

But if I have to look back at it now, I think with any new technology, it's never going to be absolutely better than what came before it. With LLMs, the relative comparison is with search. If you’re going towards search and information retrieval, you're prioritizing factuality as opposed to creativity, right? And the fundamental tradeoff with LLMs is saying, “I can trade off some amount of like factuality or ‘closeness’ to the corpus for some amount of synthesis and creativity”.

I don’t think that if we had a better model, it would have helped things at all. You could say maybe if [Galactica] had RLHF, would that have helped? I'm not too sure given that the project came out of [a big company like] Meta. Meta has a really good reputation now - people appreciate the open work they're doing - but at the time, things like the 2016 election were still in people’s minds. So I think the LLM revolution was never going to start at a big tech company, in my opinion. It was always going to happen at a company that had less reputational baggage. But I think it's pretty cool now that people see things differently. Because FAIR always had a really strong commitment to open science. It’s good that they're finally getting the credit for that.

Nathan Lambert [00:17:38]: Yeah. I have two technical questions on Galactica that I find really interesting. One is from Luca Soldaini at AI2. He said that you mentioned that the Galactica log probabilities (when producing citations) were proportional to how far in the citation graph the current paper was to the cited paper. Do you have any more interesting comments on how the latent space of Galactica actually worked? Because that is cracking the most important question of a language model for science - building a better latent representation of how the scientific information is organized.

Ross Taylor [00:18:12]: Yeah. So there were a couple of aspects to that. The first thing is we had this really nice graph that showed, as we scaled the model, the distribution of citations became closer and closer to actual citations - which is what you'd expect. But this was important for us, as our main worry was - because we were thinking about deploying to Overleaf - we didn't want to prioritize the most cited documents and create a “rich get richer” dynamic.

Nathan Lambert [00:18:38]: Google Scholar already does that. Were you re-indexing all the papers rather than building off like the Scholar graph or something?

Ross Taylor [00:18:45]: I think we were building off existing ones, using things like CrossRef…but there were lots of gaps that we had to fill. The other weird thing was that we saw some strange biases in the model. So if the model didn’t know what to cite, it would sometimes cite a general review paper, which is really weird emergent behavior. It was like the model was saying “I don't know a specific example, so I'll just give you a general overview”.

Nathan Lambert [00:19:11]: It's probably in the data.

Ross Taylor [00:19:12]: I think the thing that surprised me the most was multimodality. So we trained the model on SMILES formulae and protein sequences [alongside natural language]. And the thing that really surprised me was, we had tasks which we didn't explicitly optimize for - like converting a SMILES formula to a IUPAC name for a chemical. And if you actually looked at the attention as the model was predicting the next token, it would say something like “amino” and you could see in the chemical graph, it was explicitly attending to the relevant part of the sequence.

I found that amazing because we didn't train for it explicitly. That's the beauty of self-supervised learning. But I also found it highly ironic because some of the criticism of Galactica was “it’s ungrounded”. I was like “how grounded is this? The natural language tokens are literally attending to the underlying chemical structure!”. So that was kind of cool.

And then the other cool thing was: if you prompted with a protein sequence and asked “what is the function of this protein?”, the model was really good at answering those questions in natural language. That was awesome for me.

Nathan Lambert [00:20:33]: There's another prompting thing that I had known of [for Galactica], which was asking the model to do open-ended generation tasks. The models are still out there -  people can spin them up and do demos on their own - but if you asked it something that people think of for ChatGPT - e.g. write me a poem about a sad goldfish - it wouldn't work unless you put it in a header format. It was markdown, I think? If you prompted it in that format, it would actually do a great job.

Ross Taylor [00:20:57]: Yes, so in the Galactica demo, a lot of people were being malicious with this type of prompting for markdown articles. But I did enjoy some of the creative ones. Someone was like: write me a theorem on finding a girlfriend, and it was some of the most hilarious model output I’ve ever seen. And people also generated some amazing sci-fi…but then I think some people took it too far. But whatever. I guess it was a traumatizing experience for me at the time. But with the benefit of hindsight, I was also fun in some sense, I guess.

Nathan Lambert [00:21:30]: Yeah. It makes you understand the bigger context of the work much faster than you would otherwise.

Ross Taylor [00:21:37]: It was actually crazy at the time. So many people were using it. Even then we could see that - while it wasn’t a product - we could see that most systems were going to be designed in a similar way.

I think the interesting thing was how the winning form factor in the end was like a chat interface - you know, with ChatGPT being the winning UX. I think that was actually a big part of the story [why they succeeded]. There's a debate on whether RLHF is actually a capability advance or whether it’s just alignment…but a big part of the story [for ChatGPT’s success], in my view, was the kind of UX of how you interface with a language model, rather than the actual capabilities. But I think it's obviously not monocausal at the same time. There were several factors at play.

Nathan Lambert [00:22:25]: Yeah. So the last thing on this is that you mentioned in our e-mails about language models, creativity and making discoveries. What do you mean by that? Is that the agent-like projects you worked on at Meta?

Agents are largely something that I don't have too much comment on. I'm taking the approach of wait and see what we actually get, because there are a lot of practical approaches that I think will be reasonable. People use language models for basic formatting, for code, etc. But it's easy that if they have a little bit more feedback for things like writing a paper - e.g. find me a citation for blank and justify your answer - that step is something that I think will come. I don't know how expensive it will be to run, but is that what you mean when you think about making discoveries? Is it more autonomous? Is it a grander vision? Anything like that?

Ross Taylor [00:23:18]: I think it's more like this: the killer use case right now is information synthesis. For example, I use Claude a lot more than Google now because it combines information in a better way and sometimes generalizes well to things it hasn’t seen before.

But a really cool thing would be: can a language model answer a question which is more out of distribution? That we don't see in the training data?

So an experiment I've never done because I didn't have to compute would be this. Imagine if you could train a language model on all documents up to 1905, which is the year when Einstein had his miraculous year of four seminal papers. With that model, which is trained up to 1905, could you prompt the model to come up with a good explanation of the photoelectric effect, special relativity, this kind of stuff? And what would it take to rediscover these things?

Because presumably, with all these major discoveries, it’s never out of the blue. You’re standing on the shoulders of giants, but there’s still a lot of thought and inspiration you have to do to get to those great ideas. So that's the setup. But the creativity problem is, by its very nature, hard to benchmark.

Maybe this is a digression, but my problem with the field right now is: we’re in a situation where we've almost solved a benchmark like MATH, which is a very hard benchmark, in my opinion, at least Level 5 MATH, but I don't think we've really cracked something like reasoning. So I think it's like a whole different question about how you even evaluate these frontier tasks. But yeah, hopefully that gives a flavor of the kind of questions here…

Nathan Lambert [00:24:58]: Yeah, we can go into the reasoning conversation. I think reasoning in RLHF will take up however much time we want to keep talking. I guess we can start with the basics. What do you think people that are using language models think reasoning means? And what is the way that you would interpret what you're trying to do in improving the reasoning capability of a language model?

Ross Taylor [00:25:21]: So there's a lot of controversy on this on Twitter/X. And I think people are talking past each other because sometimes people mean different things by reasoning. At a very granular level, is legal reasoning fundamentally the same thing as mathematical reasoning? Common sense reasoning? I guess my very basic definition is that reasoning is the process of drawing conclusions based on a body of observations, or in the case of deductive reasoning, basic premises.

Nathan Lambert [00:25:50]: So math is like a subset of what you think about.

Ross Taylor [00:25:53]: Yeah. And then I guess the bigger circle is the broader topic of outcome directed behavior. I have an idea of an outcome I want to achieve, but what's the best path to get there?

And then in the LLM space, I think this problem broadly equates to the technical problem of how you use compute to get from your question to your answer. In the old days, you would just  prompt the language model directly. You would just put in a GSM8k question, put in “Answer:” and then parse A, B, C, D. So you're relying on the forward pass.

Nathan Lambert [00:26:27]: Yeah, like the FLAN data is really weird. That's a popular one that people used to train on this stuff.

Ross Taylor [00:26:33]: Yeah. And then came chain-of-thought, scratchpads, with Galactica…all these ideas of using the context window to do intermediate computation. And the more recent, although to be honest, it's actually quite an old idea, is: you have chain-of-thought, but how do you better learn the internal reasoning tokens that get you to your answer? So things like, you know, Quiet-STaR and variants of this idea.

Nathan Lambert [00:27:01]: Claude now shows you when it’s thinking, and in the Claude system prompt, it has information on how many tokens to take to think about a question. We're all thinking about trying this stuff and it's all so hard.

Ross Taylor [00:27:11]: I think it's a question of how do you learn those tokens? For us, the original thing we did was just supervised learning. So we trained on some examples and let the model generalize to know that it should do the thinking in between some tokens. There are more sophisticated ways you could achieve this nowadays.

Another point is this: there’s an analogy that’s often used about language models, that they are “thinking out loud”. I actually don’t like this analogy at all. I think “thinking out loud” makes you think there’s something wrong about this kind of thinking in token space. But it’s not clear to me that the alternative - or these old adaptive computation ideas - are any better, actually.

Nathan Lambert [00:27:58]: What do you mean by adaptive computation? Because I mostly think of “thinking out loud” as being like chain-of-thought or generating its own explanation before it gets to an answer. What would adaptive computation be?

Ross Taylor [00:28:09]: So there's a paper by Alex Graves, who wrote all these amazing papers ~10 years ago, which had a lot of foresight. He did stuff like the Neural Turing Machine paper. Adaptive computation is the idea of, instead of having fixed compute between your input and your output, you can extend the forward pass to do things better, like arithmetic, where you have to maintain/manipulate state.

When chain-of-thought came out, there was an impression that it was a bit of a hack, because you're thinking in token space whereas you should be finding a way to make the forward pass dynamic. Universal Transformer is another variant of this [adaptive computation] idea. But I think there needs to be more empirics on which approach is actually better to maintain and manipulate state. I used to be more in favor of thinking, OK, chain of thought is more of a hack, but now I actually think it's probably…

Nathan Lambert [00:29:02]: What do you mean by state, like the state of the problem in that sense?

Ross Taylor [00:29:08]: So imagine that you're doing a GSM8k question, where John originally had 100 apples, then Jane gives him five apples. He has 105. And then he gives 20 away to like Susan or something and he's left with [85 apples].

So if you’re prompting the language model directly for the answer, you're expecting the language model in that forward pass to maintain and manipulate the state in a latent space, whereas the way chain-of-thought does it is in token space.

So you essentially output the intermediate steps. One of the problems with reasoning is that we have no idea how humans mechanistically reason…but if you think about how you'd solve a GSM8k problem in your head, then to me this seems a lot closer to something like chain-of-thought than adaptive computation.

Nathan Lambert [00:29:57]: Especially when you look at the architecture and attention mechanisms. A Transformer is really good at copying. So if you keep feeding in the recent information, it copies that in some way. So I think chain-of-thought and all of these things, I mean, they're only growing in popularity in my mind, along with Quiet-STaR and these kind of methods. I’ve heard the rumors about self-explanations and all these special things. The LLaMA-3 paper has all these special tokens. I don't know what all of them do, but I can see the direction. The state is stored in context and in special formatic tokens if it needs to be.

Ross Taylor [00:30:37]: So the other big picture thing is this. With the internet, you’re only seeing the output context.

So take StackExchange. If it’s a good answer, the author probably hasn’t just responded by generating words left-to-right. Maybe they’ve looked something up, maybe they’ve done a back-of-the-envelope calculation, either explicitly or in their head, right? And the internet is missing those “internal tokens”, essentially.

Now this isn’t always a problem because the models can learn how to construct them. And the effort now is to make artificial latents / internal thought, through RL or otherwise. But I think this is actually a much bigger question, which is more than just reasoning. In the end, as models become more capable, we’ll be talking more about how we can make them human-like in the way they can answer questions and solve tasks. For example, in some situations we might like the models to have [human-like] empathy, which is also “missing” in some sense.

So my prediction is that this becomes a bigger deal in the next few years: caring more deeply about the computation these models perform to reach a conclusion. And that will be the essence of alignment, in my mind. But that's a big topic!

Nathan Lambert [00:31:50]: OK, I have a long list of specific questions on this. My first question is about process reward models.

I think the canonical paper is “let's verify step by step”. My whole gripe is that it’s hard to create the data. That’s why they don’t exist in the open. But I’m guessing you can just label data with GPT and ask for feedback on each step, and just use that as an “LLM-as-a-judge” to get reasonable step-by-step labels on process rewards. But there’s so little work on this, so I don’t know if it is worth exploring. There is some research from Meta - I think Alex Havrilla did a couple of internship projects which related to this, and he’s good - but there’s such a lack of signal.

Is this something that people should work on more, or is it too complicated? Are there simpler things to do?

Ross Taylor [00:32:38]: Our big direction was integrating outcomes into reasoning - because next token prediction isn’t the objective we actually want to optimize. So the two ways to integrate outcomes are through something like PPO or inference-time search. And in both cases, you want a good reward model or value model.

Instead of (human-annotated) “process based reward”, we were exploring ideas along the lines of Monte Carlo policy evaluation (MCPE), where the key problem is how to learn a value model. It’s maybe a separate topic, but it’s underappreciated that something like MCTS - which in the public imagination is this inference-time search technique - actually has its real magic in giving you a value network for free.

This is why it was introduced in Go, because humans couldn’t come up with good heuristics for evaluation. So if you have something like MATH where you know the answer, then the question is how do you assign step by step feedback? It doesn't have to be MCTS, but something where you backprop the outcome to these individual steps is a way to get this dense feedback.

That's a way to get “synthetic process reward”. I should stress that PRM and MCPE are actually different things. Alex Havrilla was doing something along these lines also - but anyway, hopefully this gives a sense of the approach we took.

Nathan Lambert [00:34:21]: When Q* came out, that's something that I thought it might be doing. Instead of chain-of-thought, there's this idea of tree-of-thought. You could swap in the reasoning steps. And then if you could get labels on all these reasoning steps, you’re doing search over a reasoning space - which I would expect to work, but I think it needs the right datasets. I think a large part of the open alignment community right now is underappreciating datasets, where there's a lot of focus on methods, but we don't even have the datasets to use the methods… Like, why are you coming up with seven DPO variants if you don’t have the right datasets? I understand academic incentives, but if you are not an academic, you don't need to be doing that…

Ross Taylor [00:35:00]: It's an interesting question, because I guess the first chapter of LLMs had a lot of reliance on human annotations. In a way, that's a barrier to entry for the open community, because big firms can afford to pay millions for it but open source developers can’t. But more recently, you've had the rise of things like constitutional AI [and RLAIF approaches], which I believe are comparable to human-annotated datasets anyway. So is that a good thing for the open community?

Nathan Lambert [00:35:31]: I think it is, but human preference data might be a leg that is hard to remove. One of my latter questions was: can we actually do LLM-as-a-judge for human preference data fully? I think is the critical step that we don't have an answer for. Everything else in the modern RLHF stack is becoming more reproducible in the open.

And that relates to a question I have on synthetic versus human SFT. I think Thomas [Scialom] said on the Latent Space podcast that we just use generations from the model because they're better for humans on a lot of SFT tasks. Apple had a quote in their foundation model paper saying the same thing.

So I’m thinking, shouldn’t we be redoing all of our generations for our SFT dataset with the latest GPT-4 or LLaMA-405B? Why are we using GPT-4 from March 2023? That model was not as good on reasoning. So we have headroom there on synthetic data. We have prompts that we could reuse, but we don't have the right preference datasets - datasets like UltraFeedback are not big enough. And I think they're not in the same style that a lot of labs are doing this preference tuning - where it's on-policy generation.

We tried to work with Scale at Hugging Face to do this, where we had our own SFT models. We were getting data from Scale. We were labeling it every week and we were trying to retrain the models and we weren't getting a signal. This was last July/August. So we just didn't really know what we were doing. But I suspect that what people in the open should be trying to do is generating a lot, labeling it…That was a light bulb moment for me recently. This is what we have to do, but no one has done it.

Ross Taylor [00:37:21]: Yeah, I think it's definitely underappreciated how you can get better answers than a human by sampling the models [enough times]. You mentioned that Thom made this point early on in the [LLaMA] project, but you'd be surprised how this extends to reasoning as well. Even with the Galactica model - which is now an ancient model, a bronze age model - the pass@100 on GSM8k was 98%. And it's absolutely crazy to me that even now people are using GSM8k as a benchmark. In my mind, that benchmark was solved several years ago.

It’s a subtle point because the zero shot performance was ~48% but the pass@100 was 98%. The insight there is that the model already has knowledge about how to answer correctly, it's simply not reliable. This tells you that you need to invest in reward models, process based reward, outcome based reward, everything we talked about earlier…

But the same applies to the general RLHF pipeline. If you asked me to write a poem in the style of Bertrand Russell but also mix in Snoop Dogg’s style, then I couldn't do that. But the model has knowledge of how to do that, right? So why wouldn't you sample the model?

I think now with LLaMA-3, and the 405B model being out, it’s going to be good for the community that they can use it for generating data synthetically. And I'd imagine the quality will be good enough if it's done the right way.

Nathan Lambert [00:39:30]: Yeah, I think it should be doable. But there's a fundamental question of what do we think the human preference data is doing? [Compared to] model labeled preference data, is the noise that the humans provide of a different distribution that makes the human preference data better? I don't have a lot of signal on this, but I would love to know because I would guess that Meta would love to eliminate the $10 million plus estimated human preference data spend if they could. Meta is a reasonable company…

Ross Taylor [00:40:23]: Yeah, I don't know. But here’s something that surprised me. I was originally skeptical - at least on the reasoning side for LLMs - about LLMs marking their own homework. I thought they would eventually have that capability, but I wasn’t sure…

Nathan Lambert [00:40:40]: how fast.

Ross Taylor [00:40:41]: But the interesting thing we saw was as follows. We had experiments where we’d have a LLaMA-2 model that we’d sample generations from to train ORM models, and then we’d train different reward models on this data with different base models.

What we saw is that, the better the (underlying) base model, the better the reward model was for evaluating. And there were very clear patterns we saw: as the base model scaled, so did the quality of the reward model.

So that tells you that the knowledge is not in the ORM samples that you've fine-tuned the base model on. The knowledge on how to judge is within the model itself. And the pattern was so clear in the scaling. I concluded that eventually these self-verification approaches would work. It was just a question of when they would start to work for different types of problem.

Nathan Lambert [00:41:31]: Yeah. Model capabilities are also getting more dense which helps as well. Like with smaller model, there's all these experiments with better data, showing that you get a better model with X% reduction, which is kind of off-topic…

To double-down on what you said, I think this is one of the things I also debate: what makes a good model for downstream fine-tuning? I think in the LLaMA-3 report, they train the reward models directly on the base and not on the SFT model. The Apple report mentioned that they don't just use their evaluation suite for SFT models, but they evaluate with a reward model to see what is ready for RL.

I think, especially in the open, if you want the people to adopt your base model, there's a big gain in making it easy to fine-tune. For example, LLaMA has been pretty good; LLaMA-2 especially was really good for fine-tuning. There's also been base models that don't really work for fine-tuning, partially due to bugs and partially due to the state of the optimization. Is this something that you have any insight into?

Ross Taylor [00:42:43]: Yeah, I don't think I have enough insight into it to say, but I think it's definitely something that's been undervalued. I think the view of a lot of open model providers is: you get the model out, get good Open LLM Leaderboard results, and it's mission accomplished. But the real evaluation is in two days time when you get anon accounts on X saying “I'm fine-tuning this LLaMA model, it's not working”. And when you see a pattern with this kind of behavior, you have to conclude something is wrong…

Nathan Lambert [00:43:11]: It's always a chat template thing. A lot of it is a chat template thing, but those problems do get ironed out eventually. There's this whole idea of annealing and staging pre-training. I can't tell if it is boosting current capabilities at the cost of later capabilities. I think in a few years, this will all shuffle out and it's just how we do evaluation in stages. So you're always going to optimize for the right metric.

Ross Taylor [00:43:50]: There's two points to that.

The first is about annealing. It works for the kind of benchmarks people focus on the most, but then there's a question of whether you are actually just collapsing the task distribution of the model to things you're measuring - and not the true task distribution used by the community.

And I think there's a second point - which is maybe too much of a digression - but there's an interesting debate to be had about data quality being a bit of a misnomer. In a sense that when we say “data quality” we're actually saying “this data mix works well on these benchmarks”. But if you take a “No Free Lunch (NFL)” kind of approach to this, you must be hurting task performance somewhere else, right?

Nathan Lambert [00:44:34]: Yeah, I think I’m on the record of being an AlpacaEval hater. I say this all the time, because I think AlpacaEval is sacrificing actual usefulness for their own metric. If you get a 1-2% bump on alpaca eval, maybe that’s great. But you could be getting a 10-20% bump while sacrificing actual chat abilities.

We released some models trained with PPO and our PPO models are not very good at instruction following because they don't follow modifications like be concise or some stylistic things. They're also so yappy. They just say so much…but they do well on metrics and PPO especially helped AlpacaEval. So we had to figure out how to kind of use that signal without overcooking it.

Ross Taylor [00:45:16]: Yeah, it's like a whole discussion about evals, I guess…

Nathan Lambert [00:45:21]: We could come back to evals in a second. The last question that I have is: there's multiple trends like LLaMA-3 downplayed the importance of instruction fine-tuning relative to RLHF. I think there's other quotes in [Thom’s] LatentSpace podcast talking about it. Nematron also had this report where they use SFT and then multiple stages of RLHF.

I think DPO versus PPO is overblown and that'll kind of be a wash eventually. Everyone knows DPO's advantages of being simpler. But my question is this: are there certain capabilities that only come for RLHF, and people trying to do them with SFT are just wasting their time?

I always thought safety was in this bucket where it kind of makes sense - it’s hard to train a model to refuse just with SFT. But with something like reasoning, are there certain sequencings where SFT primes you and then RLHF really helps reasoning or code? Because it seems like OpenAI is really leaning on PPO to help with reasoning and code?

Ross Taylor [00:46:45]: Yeah, I think there's two ways to answer this question. First, maybe the history of this debate on the LLaMA side, and then something on the reasoning side.

So the history is quite interesting. I would say, you know, when was it? 2023? My dates have been wrong since the pandemic…But this just was after ChatGPT. There was actually a debate internally in Meta about using RL, and a lot of senior people were very skeptical. I would say the view was…

Nathan Lambert [00:47:13]: Not just at Meta. You can see when different companies embraced RLHF, if you really start to look at their models…

Ross Taylor [00:47:22]: The view was that RL was a dead end. And that even DeepMind was moving away from RL at the time, so you should just do SFT.

But, you know, at least for the folks in the Galactica team that came to lead post-training for LLaMA, we were quite scarred by hallucinations! We were definitely of the view that we needed to have the right objectives, and that we needed to make sure language models could “know what they don’t know”. So we were quite high on RL from the beginning. And eventually, I think the LLaMA-2 paper showed that a lot of the advances in helpfulness/harmlessness were via the RL stage. So I think that approach was fairly vindicated.

On the reasoning side, I would just say it’s quite simple. It comes back to the next token prediction objective not being the actual objective you want to optimize. The objective you want to optimize for reasoning is: do you get the right answer or not? Especially since reasoning is a high precision task. If you get one token wrong, unless you have a backtracking capability, you’re never going to recover…

Nathan Lambert [00:48:32]: That's a throwback, the backtracking token. Sorry, that was a random paper! That is interesting…

Ross Taylor [00:48:38]: Yeah, all these weird methods… But I think on your question, there is a point at which these techniques kind of overlap, right? So if you're, you know, doing SFT with rejection sampling: you’re doing something close to PPO anyway. And the same for reasoning: if you sample the model and pick the trajectories that your verifier says are correct, and then do SFT on that, it is a form of RL.

The final point I’d make is this: I would say the community overreacts to certain methods being used by popular models. They think: this company uses DPO because they must have found it's fundamentally better. But actually, it's usually due to either practicality or…

Nathan Lambert [00:49:22]: Yeah, that's what I think.

Ross Taylor [00:49:24]: You have a 405B model, and if you want to do PPO, you need to have a policy model, a reward model, value model etc in memory, and it’s not like…

Nathan Lambert [00:49:33]: Especially with DPO. I think with the 405B, I'm guessing what you did was cache the reference model. You could cache the log probabilities from the reference model. So you don't need to keep them in memory when you're doing the loss of the primary model. For DPO, you don't even need an extra copy of the model in memory, which therefore means you can use the same exact stack that you use for training. So you don't have to comment on this. But I think that's probably partially why LLaMA-3 just used DPO...

Ross Taylor [00:50:07]: Yeah, I think people don't appreciate how compute works either. People assume the big companies have so much compute - tens of thousands of GPUs - so compute isn't a constraint. But all these things are subject to Say's Law, right? If you have more compute, you're going to train a bigger model. And then you're going to hit the constraints again. It’s like the old thing of trying to solve traffic by building another lane. But if you create another lane, people will use that lane of traffic.

So practicality is still a factor [behind choosing methods]. Also things like which researcher is in charge, what’s their favorite method, and also politics as well.

So I think the community has made a mistake of overreacting to these choices. There was a mixture-of-experts phase too, right? I don’t think there’s anything inherently better with either method (dense or MoE), they just have different trade-offs, and it depends on what you are trying to achieve. If you’re serving lots of people with inference, then maybe a MoE approach is better. If you’re optimizing for something simple that’s easy to train and gets good results, maybe you favor a dense approach - although that’s debatable whether it’s easier to train. But I don’t think these things are clear cut.

So I would encourage people to not just copy things because they're in a paper from a big lab. I would encourage people to try things out themselves to know what works, and figure out what the problem is that you’re really trying to solve.

Nathan Lambert [00:51:20]: I think people don't have enough long term direction in their decisions. People are not trying to make decisions about what will be right in 10 years, they are trying to get a model out as soon as possible. So there are very few people with the incentives of trying to understand in the asymptote, which method is better… I might have that incentive, because I'm a nerd, and I have an audience that is okay with me writing four paragraphs around esoteric nerdy topics, but for all these companies, that is not a real incentive.

Ross Taylor [00:51:53]: The other point I’d make - maybe it is a separate thing - is this. I made this mistake throughout my career of focusing too much on novelty and complexity.

So in my first job in sports betting, we were making models for horse racing, football, that kind of stuff. And I always had the perception that other funds had really advanced, cutting-edge, complex models - but that wasn’t the case at all.

I think there is this tendency within deep learning to assume that - especially for the secret labs - that their good performance is due to some secret, amazing method. But more often than not, good performance is due to lots of small things from different people combined into one model. Really, lots of simple things done well and solid execution. And frankly, for big firms a lot of brute force too, right? Because big companies are naturally slow. But once they find a way to mobilize resources, they’re very intimidating and hard to beat. If you’re in a big company, and you’re aware of this, which approach are you going to take: are you going to prioritize novelty or are you going to do brute force if you have 10,000s of GPUs?

So I would encourage people not to be too intimidated by this perception that the big labs are smarter. I don’t think they are.

Nathan Lambert [00:53:03]: They're earlier but they're not necessarily smarter.

Ross Taylor [00:53:09]: Yeah. So obviously the constraints are different because of less compute in the open, but still: you’ve got to use first-principle thinking and be empirical as well, and just follow that path.

Nathan Lambert [00:53:21]: Yeah. So following up on this, there's a lot of discussion around what the processes are for making a successful foundation model lab. I think Armen has been talking about a few things on Twitter with great visualizations around de-risking pre-training based on FLOPs efficiency. Do you have any comments on what makes a successful post-training team and project?

I've talked to John Schulman a couple of times - he's been the king and started all of this - and OpenAI is still looked at as being the leader in the space. I think they've always been top on Chatbot Arena, and have cracked what most people like in the style. They started early. Are there different considerations for the post-training side of things rather than the pre-training side that we might hear more about?

Ross Taylor [00:54:13]: Yeah, there's probably better people than me to answer. So in our team, originally like Robert (Stojnic), my co-founder, he was kind of managing the post-training team. And then I'd say Thom Scialom was doing a lot of the work. And then more recently Rui Hou - he kind of flies under the radar a bit - but he’s been doing a lot of the work. They are all better placed to answer than me, since I was focusing on reasoning and agents.

But I think the key thing is this: post-training is just a lot of iteration. Frankly, lots of hard work - e.g. making sure at each round of RLHF you’re not regressing in certain ways, filling holes, etc. I guess it’s hard to put a finger on a single thing, but…

Nathan Lambert [00:54:58]: There's simple things like I'm trying to get people to talk about more. I’m trying to establish a good vibe test about internal culture. How do you vibe test for a good post-training culture (or for reasoning)? I remember somebody at Anthropic told me there’s still a lot of cases where you just put your finger up to the wind and you're like “model good”. And I'm sure that is still happening. And that's just a simple cultural thing of telling the team that you can’t always trust all of your numbers.

Ross Taylor [00:55:26]: I think it is maybe a more fundamental question. I wasn’t there at the early days of FAIR - I came in 2019, but FAIR was always a very bottom up organization. Which is a great thing: that's why things like PyTorch emerged. But the real insight as to why OpenAI was ahead historically, at least until recently, was that they had more of a top-down culture and focused bets. They saw the potential of LLMs early on and it was a top-down prerogative of the company to focus on that. And in essence, it was more of an engineering problem than it was a research problem in a lot of ways.

Relatedly, I think a lot of people were surprised that the LLaMA-3 paper wasn't as “novel” as they were expecting. But that just reflects the fact that a lot of it is just engineering and engineering is really hard - a lot of hard work. Not always a lot of new methods, but it is a lot of hard work.

Nathan Lambert [00:56:22]: Yeah, we're starting our next fine tuning model and everyone's asking me: “what should we work on?”. I'm trying to tell them “we just have to filter data and generate more completions”. We’ll have a lot of prompts, we have to filter them, generate completions from good models, and then we’ll have to generate more completions and keep doing this process…And in 10 weeks, we'll probably have a very good open model. We’ll just have to be boring for 10 weeks! And we have like 10 people involved.

So it's a bit of a bigger project, which I think is the right way to do it. We have just started getting improvements on IFVL by copying Nemotron. We use some open math datasets and the math scores are getting closer to LLaMA. It is really the simplest things ever. It's like browsing Hugging Face and being like, “NVIDIA released some JSON format data, some instruction format data, like we add it in and the numbers go up”.

Ross Taylor [00:57:16]: Yeah, I think I said earlier, but it raises an interesting question where this kind of approach - of grinding until the open LLM leaderboard numbers get to 100% - I think we’re going to get to a situation where all the benchmarks are solved, but where we haven't really, in my mind, at least solved intelligence.

What does it mean that we'll get close to 100% on MATH, you know, without any inference time search? I think sooner or later, while it looks like we’re on an exponential with LLMs, we’ll realize we’re actually on an S curve. Eventually we're going to get back to this mode where we have to do new things. And I think that's great, because that's what motivates me.

But yeah, I think there's waves, and we’re in this heavy exploitation mode right now with LLMs - away from the glory days of architecture exploration. But my hope is that we'll get back to the stage where, after exhausting all the [current] benchmarks, we say: OK, now we need to do something completely different. But who knows?

Nathan Lambert [00:58:26]: I see it similarly. I think we still have a year or two, at least in the open. If the closed models start saturating and they start doing things differently, that's fine. But eventually it'll all get there. And in that phase, I mostly keep working just to make sure that the ecosystem doesn't fold in on itself. So that's probably the one-sentence summary of what I'm doing these days: add transparency so that regulatory capture doesn't nuke everything. And that's fine, but I think it's still going to be longer than people expect. I don't think we have true signs of saturation at the top. We'll see what GPT-5 does - if GPT-5 never comes out - and then we’ll really know.

But it seems like it's going to come. I think there's enough signs that it'll come eventually. I think I don't know the answer to this - and it's not really our expertise - but I'm interested in the potential architecture of GPT-5 and if it's GPT-4o like and they're using more multimodal data to try to keep the data engine going relative to just going bigger. I don't know the answer, but that's kind of the future questions I’m thinking about.

Ross Taylor [00:59:34]: In my mind, like three years ago, the thing on the horizon I saw was agents. That’s where a lot of people are working right now: long form tasks where an agent doesn't have to answer a question immediately, and [can instead] go away for a while doing some research and answer later. I think that will take up a lot of time in the next five years.

It's both a compute problem of bigger models - more scale will do better - but also a data problem of how do you generate these trajectories? How do you get reliability? So it’s more successful and less error-prone at each step. And I think in principle it's solvable, but I just think it would take some time.

Nathan Lambert [01:00:18]: Yeah, it seems that engineering is required. It doesn’t seem like something that's just going to emerge. It's building a whole system and scaffolding around agents. Just unglorious work.

Ross Taylor [01:00:32]: Yeah.

Nathan Lambert [01:00:34]: OK, anything else you want to add? Do you want to get people excited about your start-up or is it too early? Maybe too early, yeah?

Ross Taylor [01:00:43]: Yeah, what else should I say? It has been nice to step back for a bit and look a bit ahead into the future. For me, my best days creatively were my teenage years when I got back home from school and spent the rest of the day programming. It’s quite nice to feel like that again: to be in that zone again where I can shut the world out and do some work.

But maybe just to give a hint of the areas I'm interested in, I think it comes back to this problem of how alignment is going to be a process of making AI more human-like. For example, how do you control for things like deception - which Anthropic has done a lot of really good work on.

Essentially… the latents of AI are [potentially] misaligned with human latents, and the question is: what do the human latents look like anyway? And how do we model these things?

That is very abstract and high level, but that is the fundamental question I want to work on. But yeah, I think I can talk about it later in the year!

Nathan Lambert [01:01:49]: Yeah, sounds good. Thanks for coming on. This was great. I think people are going to get a ton out of this. I think just a very sensible conversation on fine-tuning, reasoning and some of the things that got us here. And that's what I was hoping to get out of it, so thanks again!

Ross Taylor [01:02:06]: Yeah, great to talk, Nathan. Have a good one!

Apple, Meta, and Nvidia all agree -- synthetic data, iterative training, human preference labels, and lots of filtering.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/frontier-model-post-training

00:00 Llama 3.1 post-training and the new normal for RLHF
01:18 A new standard pipeline
01:45 Human preference data
02:59 Scaling RLHF
05:03 Synthetic data
06:10 The new normal
06:51 Data quality is king
07:18 Apple confirms the new normal

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_018.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_020.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_031.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_033.png
Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_035.png

This week, I had the pleasure of chatting with Sebastian Raschka. Sebastian is doing a ton of work on the open language model ecosystem and AI research broadly. He’s been writing the great Ahead of AI newsletter (that has the biggest audience overlap with Interconnects, at 26%, so a lot of you know him) and multiple educational books, all on top of being a full time machine learning engineer at Lightning.ai, where he maintains LitGPT, which he described as being like Karpathy’s NanoGPT, with slightly more abstractions.

This conversation mostly surrounds keeping up with AI research, the state of the open LLM ecosystem post Llama 3.1, and many narrow topics in between. I learned that Sebastian used to be an Arxiv moderator, which gives some simple color on how Arxiv and sifting through thousands of papers works. We cover a lot of ground here, so I hope you enjoy it.

Listen on Apple Podcasts, Spotify, and where ever you get your podcasts. For other interviews, go here.

YouTube

Chapters

* [00:00:00] Introduction & Sebastian’s background

* [00:04:28] The state of deep learning and language models in 2018

* [00:08:02] Sebastian's work at Lightning AI and LitGPT

* [00:12:23] Distillation and its potential in language model training

* [00:14:14] Implementing language models and common pitfalls

* [00:18:45] Modern architectures: Mixture of experts models, early v. late fusion multimodal

* [00:24:23] Sebastian's book on building language models from scratch

* [00:27:13] Comparing ChatGPT, Claude, and Google's Gemini for various tasks

* [00:38:21] Vibing and checking new language models during implementation

* [00:40:42] Selecting papers to read and moderating Arxiv

* [00:45:36] Motivation for working on AI education

* [00:52:46] Llama 3 fine-tuning

* [00:57:26] The potential impact of AI on jobs in writing and education

* [01:00:57] The future directions of AI

Transcript

Built with smol-podcaster and with love of Latent Space.

Nathan Lambert [00:00:00]: Hey, Sebastian, welcome to this kind of interconnects, normally researcher interviews. You were a professor, so that definitely counts. You do a lot of different things these days. Let's get talking into language models. Welcome. Yeah.

Sebastian Raschka [00:01:35]: Thanks so much for the invitation, Nathan. I'm a big fan actually of the interconnects newsletter, so I'm hoping we can have some fun chat about research, LLMs, and what's hot these days, basically. Yeah.

Nathan Lambert [00:01:48]: I have a little section on the end, which is keeping up with AI research, writing about AI and process, because you do so many things, but I kind of want to jump into how you got to AI, because you have an interesting career path. So you were a professor at Wisconsin Madison for years. I saw in statistics, which ... I also went all the way back to find your PhD thesis, which was uncovering hidden patterns of molecular recognition. So this was a while ago, and is this kind of ... Can you explain your background and how you got into AI? I'm guessing it's through computational statistics or something like this.

Sebastian Raschka [00:02:24]: Yeah. Close. So yeah, you did some research there. Interesting. So yeah, it's been a long time since my PhD thesis. This is maybe seven years now. And back then, it started even earlier when I got into AI, that was like, I would say 2012-ish. I was in grad school and I was taking a statistical pattern classification class. And in that class, yeah, the star of the show was basically naive Bayes classifiers, or in general, Bayesian methods for pattern recognition. And from there, I kind of really got into machine learning. So there was, I would say, more statistical-based, but it was all about classifying things. And then I think it was also right about the time where Cozera was launched, and I saw Andrew Ng's Cozera class. That was, I think, the first class in 2011-12 back then. And yeah, that's basically how I started from statistical pattern classification into machine learning. And I applied that for computational biology problems like molecule and drug discovery, like pharmaceutical drug discovery. And yeah, from there, I joined at some point after my graduation, the University of Wisconsin in Madison, where I was in the statistics department, but I did mostly deep learning research, essentially. I was the only one basically doing Python, deep learning, machine learning stuff. So yeah.

Nathan Lambert [00:03:48]: What year was this, and what did it look like at the time?

Sebastian Raschka [00:03:52]: That was around 2018, I think August 2018, when I joined the department. And yeah, I mean, so it's the statistics department, but my work was technically all machine learning and deep learning. I mean, a lot of students were really excited about learning machine learning. I think it was just around the time where it got really popular. And yeah, I was teaching machine learning and deep learning classes as well. They were always like, you know, full and crowded, like a lot of students were excited about that. Also, in general, like the time learning about Python, machine learning, data science, all these topics.

Nathan Lambert [00:04:28]: It's, I mean, it's very interesting because I was a student, I was a grad student at this time or that time in like 2018. That's what deep RL was really taking off. And it probably feels like that probably felt kind of like the language model thing was like as a student at the time, where it's just like, there's so many people in all these classes. And now language models have more of a real world application, but I think as a student, it probably feels so, so similar. Yeah.

Sebastian Raschka [00:04:50]: So also back then, if I may say that it's like large language models already existed. I think the GPT paper, was it 2018? Something like that?

Nathan Lambert [00:04:59]: Yeah, 2018 or 2019. Yeah. For GPT-2, I think.

Sebastian Raschka [00:05:04]: Remember covering, like I had a whole hour or two hours on large language models back then, but it was all focused on BERT models and basically also using them for more like classification tasks. Now, I would say maybe a lot of business problems still evolve around classification, but everything else is basically generative, generating text, generating images and stuff. So it has changed a lot.

Nathan Lambert [00:05:28]: Yeah, for sure. It's like a sequence of like, is it like the transform, is it like Elmo, BERT and the transformers are probably the things that you're talking about all the time? Just very interesting. I think Yitay had this, did you read Yitay's recent blog posts on language model architectures and kind of walked through why encoder decoder is no longer in vogue? Did you see this?

Sebastian Raschka [00:05:51]: Yeah, I think I haven't seen the article, but I remember having discussions with people about that recently. I mean, I think there was actually, it's interesting. So I think T5, if you would train it and fine tune it, it would still be a really good model for sequence to sequence tasks, like language translation and stuff like that.

Nathan Lambert [00:06:10]: Yeah. Cohere for AI did this with AYA. They used T5 for their first AYA version, which most people were like, oh, they've Cohere branded it so well, but no one realized they're using T5.

Sebastian Raschka [00:06:21]: See, I even didn't know about that. And so also on that note, I would say there was something else I wanted to say. So then there's also still the classification thing and using LLMs for classification. And it was also usually either a bird like encoder, or you could also use an encoder decoder, but mostly an encoder. But I've seen also recent papers using just decoder models for that. Just basically removing the, I saw two papers on that actually, like removing the causal mask. So basically reverting it back to an encoder using LLMA and then removing the mask. So in that sense.

Nathan Lambert [00:06:59]: And it works well as a classifier. You can just kind of use it. That's awesome.

Sebastian Raschka [00:07:04]: I mean, you could even do that without removing the causal mask. So you could just tune the last token basically, but yeah, if you remove it, yeah. They found that you could use probably the first token even, because if you have the last token, you don't, you have to have padding always because you have to pad it to the longest sequence. Otherwise the last token would be a different one in each training example. And so in this way you could use an earlier token basically, and then keep it fixed.

Nathan Lambert [00:07:30]: Yeah. Yeah. Now with your work at Lightning AI, do you do a lot of these things like hacking around with language models? Because I think it's kind of an underexplored space where just like people remove layers and plug things together. I think there was like, when merging was just getting going, there was like Franken Llama 2, where somebody made like a Llama 2 30 B by just chopping layers and stuff together. There's so much unexplored signal there that I just, do you have your, have you ever looked at these things or you don't do that much?

Sebastian Raschka [00:08:02]: I must say I'm not a big fan of merging. Maybe I'm just not good at it. I rather prefer fine tuning, start changing things or training and fine tuning things. So yeah, I do a lot of this type of hacking. Sometimes voluntarily, sometimes involuntarily, because I make a mistake or something or like, because at Lightning I developed this library, LitGPT, which is an open source library, pre-training, fine tuning and serving and deploying LLMs. But it's basically a from scratch implementation. You can think of it as a NanoGPT from Andrej Karpathy, but for all types of LLMs, like Llama, Gemma, PHY, all of them. But the focus is also like NanoGPT is on readable code or like keeping it relatively simple. Of course it gets a bit more complex there when you add multi-GPU training, tensor parallel, fully sharded data parallelism and stuff like that. So if you add all these settings, it gets a bit more complicated, but the focus is still on having a code base that you can easily work with. And in that context, it's very easy to remove layers and change things. I mean, yeah, so that is usually, I build it like for colleagues at Lightning, but also like open source community, but then also for myself to tweak things, to change things and stuff like that. So yeah, I should also say, it's not just me, it's Carlos and Adrian who started this library. Currently I'm like the main person maintaining it, but a lot of people contribute to it. So it's actually a nice playground.

Nathan Lambert [00:09:41]: There's kind of two follows odds for this. One is like, what part of the language model training stack, if somebody is going to start with libgpt or HuggingFace or whatever, like they're trying to fine tune a model, you can do an example. And then what is the thing that they should do to go like one level deeper to learn how these things work? Because you're saying with libgpt, you can do all these different architectures. I don't know if I would recommend architectures, but it's a good way to learn how like the attention implementation and how different layers are shaped and things like this. Is there different areas you'd recommend people to look at?

Sebastian Raschka [00:10:14]: Yeah, I would actually, okay. So it's like a shameless plug, but in my book, I have a book where I do this step by step, the implementation. And this is for only one model, like a simple model, a GPT-2 model. Because it's like the, I would say the one that started all of this, right? Like the main architecture and everything else is kind of like a derivative almost of it. So I would think in a good way that it is making tweaks and improving things, but basically starting with one architecture, like you said, not looking at different ones at first, and then just understanding what is, I would say the best way is what is the input data here? How does it look like? What does go into the LLM and really how does it pass through the layers? And then from there, okay, we understand how a model learns to generate one word at a time and then going from there to instruction, fine tuning, and then even like alignment with a DPO, for example. So doing like all these different lifecycle things from implementing one architecture, pre-training, fine tuning, aligning, and then from there, I think it's a useful or interesting exercise to see how different architectures make slightly different choices, like replacing the Gelu activation with a Silu activation or pre- and post-layer norm and like these like nuances, changing the number of heads or number of layers. And yeah.

Nathan Lambert [00:11:38]: Yeah. I mean, in industry, everyone kind of is converging to similar things or like people converge to a similar recipe and then they stick with it for infinity. So like each of the orgs have these recipes that it's too risky to change and like AI2 are like still converging at a recipe. So we're like learning things that the Llama team does and it's like RMS norm and they think it's very important or like these different things. And I wonder how like the open community is going to converge on pre-training things. So like what scale of models do you recommend people train for your book? Are they training like the hundred million scale GPT-2? Is it smaller? Because I think in Colab, you can fine tune maybe with Laura, a 7b model, I think. Is that true?

Sebastian Raschka [00:12:23]: Yeah. So this is true. But I think for Laura, if you want to fine tune 7b model, you would need, I think, bits and bytes of quantization, the normal float for like some quantization. But yeah. So for the, or maybe going one step back for the book, it's really the smallest model, like the hundred, what is it, hundred something million. But I also have settings. If you like, if let's say your machine permits, use the larger version. So there are four larger versions, like 300, 700, and 1.5 billion. But it's really up to the reader. I have all the examples with the smallest one so that it even runs on a MacBook Air. So on this podcast, I'm here on my small MacBook Air and all the models train in a few minutes fine. Of course, I'm not doing the whole pre-training for that. You would need a GPU for a week or maybe I would say maybe even longer than that now. I mean, it depends on the GPU, of course, but H100, maybe a week. But also the other reason is yeah, in practice, you would probably use pre-trained weights and then you can find, so you can do continued pre-training and then fine tune. So the focus is basically understanding how the pre-training works, then loading pre-trained weights. But then also the fine tuning is like the full, the full thing, like doing it to fine tune a classifier, but also instruction fine tuning essentially. And that doesn't take too long. I would recommend using a GPU, but it would technically run on a CPU. And get back to the question you had with a 7 billion model for that one A100, I would say yeah, one A100 would probably work for a 7 billion model. But you can also, if you have Litt-GPT or if you use Litt-GPT as a setting, you can set the number of devices and shard it over multiple GPUs. Yeah.

Nathan Lambert [00:14:14]: I mean, all of this stuff is getting so much easier. I think, I don't know, when did you start writing this book and all of these chapters? Because I've seen the GitHub, I haven't looked at when it started.

Sebastian Raschka [00:14:23]: Actually longer than you might think. It took a long time. It's almost, at this point, one and a half years approximately.

Nathan Lambert [00:14:30]: Because at that time, like a 1 billion parameter model, like what was the state of the art 1 billion parameter model a year and a half ago? Some random model. But today, like people are trading 1 billion parameter models for 15 trillion tokens. So the fine tuning that you can do there is getting extremely good. And I'm going to guess that people are going to start training even smaller models with these distillation losses. So have you looked at distillation at all? I think it's full on coming in the next six months. We can shift it to like the LLAMA3 and the state of the open ecosystem section, because it kind of goes in. It's like LLAMA3 was not distilled. It's a specific loss function. I hate it that there's synthetic data came around and people call, I was on this paper, the Zephyr paper, the title is Direct Distillation of Language Models. But now the technical definition of distillation, which is like knowledge distillation from a teacher is becoming popular. So the whole synthetic data and alignment and everything is like screwed in a doubly defined word.

Sebastian Raschka [00:15:30]: So basically what you're saying is that people who just use synthetic data refer to it as distillation because it's from a larger model. Yeah. Yeah. Yeah. Confusing. I think Gemma too did that actually recently. So that was an example where they did that. And I do think, you know, I think it's also coming. So I have for my book, that's like the core chapters I have, but I have a whole long list of bonus material that I want to cover and distillation, knowledge distillation is one of them. So this will be something over the next few years, but you know, doing tutorials on those and yeah.

Nathan Lambert [00:16:04]: Because I think people can actually use it as a thing. So how distillation works, I've thought about implementing it, but as it works is that if you have a fine tuning corpus, you get all the predictions from your big model. So all the log probabilities from your big model and you store them in memory. And then as you're training the model you're training, which is smaller, you essentially weight them by those predictions because you store them from memory. So you don't need to store the big model in memory when you're training. So I think people should be able to like, or someone will upload a data set file of like a giant log probs of Lama 405B and that people will just try to fine tune from it. I'm surprised that Lama 3 didn't use it, but I think it's just because they're focused on scale and data more than any fancy things.

Sebastian Raschka [00:16:49]: Yeah. And I think the, I can, I think I probably know why, but also, yeah. One thing is I should, one should also add is why I think it's also becoming more popular is like Lama 3.1, they just allowed doing that. I think before it was according to the license, technically not allowed to use Lama 3 models to improve other models, but now, now we can. So I think, like you said, it's probably going to be a hot topic, but I do think they didn't do that because the 405B Lama model just finished, I think. So I think, I mean, if you think back, they shared the Lama 3 model, it's like, I don't know, half a year ago or something, many months ago. So I think it's really more like, yeah, it hasn't finished training, but maybe for Lama 4, we will see more distillation using the 3.1 model for that.

Nathan Lambert [00:17:38]: Yeah, it's more architecture things. So for while we're talking about distillation, almost like Cloud Flash or Google Gemini Flash is confirmed as distillation. And it is very likely that Cloud Haiku and GPT-40 mini are distilled in the technical sense of the word, which is like, I think it's obvious that that works on pre-training. And I think there will be a breakthrough fine tuning model, kind of like the likes of Zephyr, Starlang, I'm forgetting more names, but ones that really reach the narrative from fine tuning on distilled data. I think that'll come in the next six months. So honestly, I'm telling the people I work with, we should try to do this before something new, because it's so obvious now.

Sebastian Raschka [00:18:22]: One thing I've seen also a trend, I wouldn't say backwards, but a thing that doesn't seem to be that popular anymore is a mixture of expert models. What do you think about that? Is that like something like that was like a fad and now people don't, you know, they explore other things like distillation. I mean, you could do both, but it feels like a mixture of experts is not as hot anymore

Nathan Lambert [00:18:45]: somehow. I don't know.

Sebastian Raschka [00:18:45]: What do you think?

Nathan Lambert [00:18:47]: There's two things. Small mixture of expert models are definitely coming out. Essentially, you get a fixed improvement in flop efficiency at pre-training. So essentially, if you're going to pre-train like an X billion parameter model with mixture of experts, it'll go like 40 percent faster or some pretty appreciable number. There's a lot of rumors and discussion that scaling up mixture of experts models is really hard from a stability point of view. So a lot of these open people, you could get it started and we're playing with these AI too. So we want to play in the mixture of experts space as well. And doing a small model works, but there's a lot of headaches. I think like some of the friends at Databricks Mosaic ML have been the clearest about this. It's just like you do not, like you at AI too, do not have the engineering throughput to deal with the headaches that comes from mixture of experts. So I think there's still clear signal from industry and people and like, I mean, Deep Seek's releasing MOEs. I think Quen has a small MOE and these are pretty good models. But I think it's a really heavy engineering lift to get to mixture of experts to work. I like GPT-4 scales. I expect Meta to figure it out. I think it's just on their list and they figured out dense first. The thing I'm more interested in for GPT-4, I don't care if it's mixture of experts. I think they have the compute to do either way. But for Llama-4, God, all the numbers throw me off so bad. But I think that OpenAI and Google might be slightly ahead by having the early fusion model. So essentially with these multimodal models, there's the concept of early versus late fusion. The first visual models that people were playing with the GPT-4 were this late fusion. And now like GPT-4.0 is early fusion. And it seems like Gemini is probably early fusion, which means they take in direct audio, video, text directly at the input, the training data changes. And I don't know how much of a heavy lift it is to get that to work. I think that might be the bigger change. And that might be harder for Meta to catch up on than anything else. But no one's really talking about it.

Sebastian Raschka [00:20:58]: But also here, I think that is something I feel like others have. I mean, I remember even like last year, there were a lot of papers with a late fusion thing, like I think Llama adapter papers and stuff like that, like retrofitting the models. But yeah, I haven't seen that much focus on that from Meta. But I mean, they had a section on that in the paper, but it felt almost like an afterthought. I don't know. It's like where, yeah, I think maybe there's a different team at Meta that works on

Nathan Lambert [00:21:26]: that. There is a Chameleon team that was doing this, and I think a lot of them have left. My question, essentially, that I want to debate and I don't know the answer to is like, because essentially it takes so much different data pipelines. So you have to have a much clearer balance between video images and audio and text when you're training early fusion than with late fusion, because you just add a bunch of images at the end. And like if that data curation step is going to be a big bottleneck for kind of shifting and if Google and OpenAI have an advantage by just scraping YouTube, like Google obviously can't scrape YouTube and I'm not saying that they are, but like if it becomes a way that you can get more data and like GPT 5.0 is the first model that OpenAI releases, then I'll be like, OK, the GPT 4.0 thing was just a pivot. And I actually think this could happen. I don't put this at like a one percent probability. I could see this as being what the labs are betting on. It just takes so long to spin up this entire new pipeline of training.

Sebastian Raschka [00:22:25]: But one question here is going back to a point you mentioned earlier regarding the knowledge distillation where you can just precompute all these things, you could technically do that also just once for the whole data set. Let's say you have a very good image encoder, audio encoder. You would never have to redo this if you do it well. Right. I mean, it would be something you do it, take care of it once and then you pass it just as tokens to the to the other team, basically.

Nathan Lambert [00:22:49]: Yeah, probably. I don't know. I'm not like I don't have as much insight into really advanced pre-training practices as I would like. I'm mostly of a similar boat of like fine tuning models and playing with things because I'm trying to play like, have you played with Llama 3405b at all? For context, the recording is like, what is this, like a week after, like six days after. Like I haven't gotten it set up, but I'm really curious. Like I don't have clear expectations on how the open source community, like the open language model ecosystem kind of evolves from here with these new Llama models, the new Mistral models. It feels like a total, from like a technical and a policy perspective for me, it feels like a pivot. I think the educational side of things, it's actually more of the same. Like we knew we knew this was coming, but it just it feels like it could be qualitatively different going forward. Do you see anything? Have you tried anything?

Sebastian Raschka [00:23:45]: Yeah, I did actually try the Llama 3.1 models. I, when they came out last week, we added them to Litchipiti. I took care of the eight and 70 billion models. And my colleague Adrian, he also added support for the 405 billion models. So just briefly trying it, it looks really good. So the thing is with a 405 billion model, it's a bit tricky. So I think the problem here is, of course, it's free. Everyone can use it, but in a sense it's still expensive to run it because you need, so we were running it with bits and bytes of quantization, like a normal float four on eight H100s. And this is expensive, right? I mean, eight H100s, it's probably more than a hundred bucks an hour.

Nathan Lambert [00:24:26]: I was trying to do the same and I messed up the BLM installation. I was like, okay, I spent an hour on this. Yeah.

Sebastian Raschka [00:24:32]: So you can try Litchipiti maybe. So it works with.

Nathan Lambert [00:24:36]: Yeah. And there's a related question. One of the things I'm trying to ask people who are hands on, just like, how do you, what do you do to vibe check a new model as you go through so much AI research material and language model material? It's like, everyone has their procedures and how do you go about that?

Sebastian Raschka [00:24:51]: So for me, it's like, I, I mean, I use these more like for making sure they generate the correct answers and stuff like that, or something that is reasonable. So honestly, really simple questions for me just to see, so this is more like, I'm not necessarily benchmarking these models. I'm more like making sure the implementation is correct. And for that, I use simple questions like what do llamas eat? What is one plus two? You know, like just making sure, because it's actually easy. Something I just fixed this morning. It's easy to mess up things like KB caching, where you cache, you don't clear the cache and then there's something from the previous answer and the answer looks kind of correct, but it's kind of weird. And, you know, like simple questions can sometimes reveal that. So basically what I do is I ask it multiple, multiple questions the same time. So, sorry, repeatedly, like the same question repeatedly and see if the outputs still make sense and stuff and then mixing them up, but like in a loop basically, but I'm not so much like, that's a great way to make sure the implementation works.

Nathan Lambert [00:25:53]: Cause I think in transformers, they had a missing end token. There's so many little things like this when implementing stuff. Like the, the end tokens is such a ban or like the chat templating can always break things. Cause it also can happen that you mess up pre-training and then you need to have something in the chat template that people might not know. I think in one of the early Olmo models, we like missed a new line in, in one of our documents when we were annealing it. So in order to fine tune it, you had to like have an extra new line before the chat template and like most people will just miss that. Yeah. This is very, very interesting point.

Sebastian Raschka [00:26:28]: It's like, you don't even notice it usually when you use something like, I don't know, chat GPT, because it's applied behind the scenes. But if you implement these things yourself, you have to be really diligent and careful to do it very consistently. Like one little, like you said, new line throws it totally off. It's, it's, yeah, it's interesting. It's like, you have to be, I noticed that I was actually working on some DPO stuff this weekend and my template for fine tuning and DPO alignment, the one that I'm working on alignment, the prompt template was a bit different and I got like garbage results. And then, oh, I, I stripped some line here, the new line character, basically something similar, like you said. So it's, it's very sensitive to that.

Nathan Lambert [00:27:04]: Yeah.

Sebastian Raschka [00:27:04]: Yeah.

Nathan Lambert [00:27:05]: This, this makes sense. Um, related, do you use Clod, chat GPT, any of these regularly in your workflow? Are you team Clod?

Sebastian Raschka [00:27:13]: Uh, so yeah, so it depends. I have both and I flip back and forth between them. I don't know. I'm probably not really good at prompting, but sometimes I get better results with one over the other. Um, I think. I wouldn't say one is better than the other. They're just different. I would say I'm using.

Nathan Lambert [00:27:31]: That's kind of what I think. It's important. Like, it's good. Like, what do you think of both of them? I think it's good for people to know this because it's, it takes some practice to understand and using both. Both people don't use both. Yeah.

Sebastian Raschka [00:27:43]: I would say when I use also GPT-4, I must say I use the, uh, it's called legacy now, but the original GPT-4, I don't like the mini and old versions. And, uh, for Claude, I use the opposite of the, not the new one, but the one, the previous larger one, the slower one. And, um, I think for me it's like coding wise, it's kind of weird, but most of the time I like GPT-4 better for code stuff. But then I think also, uh, I think, you know, what, what's better with GPT-4 was it's, it's a bit more up to date, um, with knowledge, I think. But Claude has, I think better, you know, when you say improve my writing or something like that, it has more, it has less, you know, like these, like I delve into something, these weird words and stuff like it, it's a less, it's more natural a bit, I would say, but

Nathan Lambert [00:28:34]: also not always.

Sebastian Raschka [00:28:34]: I agree.

Nathan Lambert [00:28:36]: It's like, it has a bit more flair and a bit more unpredictability. So I like use a Claude on my phone, but I've found, I've tried to use Claude for like information transformation tasks, like LaTeX or taking, taking data out of a table. And sometimes it just like refuses. Like I do research on like AI safety, like safety and bias. So if I put anything into Claude that I'm trying to transform that data, it just says no. Cause it's like, I can't comment on like a mean story. Well as OpenAI will just do it. And it's like the processing that OpenAI does is very good. So I actually like canceled my GPT subscription when I started Claude, but I kind of regret it now. I'm like, oh, now I need both, which is, which is a little annoying. Yeah.

Sebastian Raschka [00:29:16]: It's like, yeah. So one thing is what is interesting though, is we, we're talking about GPT-4 and Claude here, but we haven't even mentioned Google Gemini.

Nathan Lambert [00:29:24]: I don't know.

Sebastian Raschka [00:29:24]: I personally, I tried the early versions. I don't want to say the newer versions are not good. I just haven't tried because I didn't need to, but do you have experiences with Gemini

Nathan Lambert [00:29:34]: or? I was using Gemini in search preview. So if you have the Google app, I can, I'm recording this in, in video. Like you have the Google app, like at the top, you could click on Gemini, which I was doing for a while just to play with it. But like, I don't use it on the web. I, they do have a nice interface that looks exactly the same, but somehow I got grandfathered into like AI studio, which I use for, if I upload, record a podcast, I upload the podcast and I'm like write chapters or something. And it actually works, which is pretty cool to be able to upload like an hour long podcast. But for whatever reason, the Google interface, other than the Google app, hasn't stuck for me. And I think that's the biggest, biggest limitation. And I use it more in a googly way. So I'd not, I'm not as perceptive to style. I see. I see.

Sebastian Raschka [00:30:20]: So also I'm curious. I just yesterday saw Apple's on device AI is a bit delayed, I think. And for that, I think it's an interesting one. We will see how this will work because this will be, I think also smaller models. And there's a, for me, it's like, I never really care about speed for these things. It's like, I just want the best possible models. So this is also why I was a bit disappointed when GPT-4 O came out and GPT-4 Mini came

Nathan Lambert [00:30:46]: out.

Sebastian Raschka [00:30:46]: It's like, ah, I don't really care about if it's faster or not. I just want it better. You know, I want to have better quality. I don't know. It's maybe it's just me.

Nathan Lambert [00:30:53]: I think for building applications, speed is really good. So I have a few friends that run startups that are heavily built on language models and they have a similar stack to perplexity, which is like the user passes in a query that have a primary language model request and they have a series of feedback requests or small requests on top of that. So when you're concatenating multiple requests, like speed is extremely important. And when you're like selling a product, speed is extremely important. But if you're like tinkering and trying to learn, it is much slower. It's true. Yeah. Yeah.

Sebastian Raschka [00:31:19]: It's like the real world, like, sorry, not real world, but the individual user, um, yeah, using it as a tool in everyday life versus really building an application based on an API that makes sense.

Nathan Lambert [00:31:32]: Yeah.

Sebastian Raschka [00:31:32]: So there are two different use cases.

Nathan Lambert [00:31:34]: Yeah. Yeah. I think we're kind of talking about style. I have a section on RLHF here. I just wanted to like, what do you think you do spend a lot so much on AI education is like, what do you think is most confusing to people about this kind of whole post-training thing, which is instruction tuning, reinforcement learning from human feedback, other safety modules, like adding a filter and stuff like this. I'm really on the bandwagon of trying to convince people that RLHF is deeply tried with style, which is like this, how this discussion of cloud versus, um, open AI and Google and all these things. And I don't really know how to portray that in like an educational technical point of view. So like, I'll do an analysis of the paper and I'll do like DPO and like scores and all these things. But at the same time, for most people reading my articles, the most important thing is probably to know that open AI is really smart about their style. And that's why they're so high on chatbot arena. But like, I've written about it a couple of times. I have another article in the drafts, which is essentially like why GPT 4.0 mini like broke chatbot arena. Because everyone's so upset that it scored so highly, but it's not that surprising if you look at historical events.

Sebastian Raschka [00:32:39]: So it's basically exploitation of the benchmark almost you're saying or like the benchmark

Nathan Lambert [00:32:45]: is focused on style and it really penalizes refusals. So like I get refusals when I use cloud. So it's definitely going to like be downweighted. And like open AI is really good at this. This is what they've been doing for a long time. But I don't really know how to educate this. Like, have you thought about, like, there was a question on Twitter of why didn't you include RLHF in your latest? It was kind of a joke, but I took it out.

Sebastian Raschka [00:33:09]: Well, if yeah, I can maybe answer that. It's it's in the works. No, so there are multiple reasons. And so one is it's so there are page limits per chapter. And originally it was meant to be in chapter seven. It got way too long. It's actually even without it. Chapter seven is the longest chapter already. And what is the other one is fine tuning.

Nathan Lambert [00:33:29]: Oh, sorry.

Sebastian Raschka [00:33:30]: Instruction fine tuning. Yeah, I called it not instruction fine tuning. I called it fine tuning to follow instructions, which were originally, which was originally meant to have both, but then it got too long. And the other thing is, you know, like one book chapter takes about two months and a lot of people who really want to book before the new semester starts. So it's like, you know, it's, there could be another chapter on it, but it would be

Nathan Lambert [00:33:54]: another two months.

Sebastian Raschka [00:33:54]: And that, I mean, it's not really an excuse, but the other one is I was not happy with the results. And this is a very mathy topic. And I was like, okay, I have this book, which is very clear and makes hopefully a lot of sense. And then I have this really super complicated chapter at the end. I don't know if that's very satisfying to read or death.

Nathan Lambert [00:34:15]: Yeah.

Sebastian Raschka [00:34:15]: Where it's like, so you read this book, everything makes sense. And then it comes to this huge...

Nathan Lambert [00:34:19]: Why is RLHF so much mathier? Like, I know a couple, there's a couple of core equations. Like the core equation is like the RL optimization step, which is expected expectation, maximization of reward subject to penalty. And like, where does most of the, like compared to pre-training, which is like one equation, like that is also one equation, but there's a lot of downstream stuff, I'm guessing. Yeah.

Sebastian Raschka [00:34:41]: I think it's the explaining a bit about reinforcement learning. I mean, you don't really have to explain reinforcement learning in a classic sense, maybe, but yeah, there's still like KL divergence and penalties and reward margins. And there are lots of things happening at the same time. And the code is also very long if you especially want to track the rewards and stuff. So for my instruction fine tuning chapter, I'm using exactly the same training function I implemented in the pre-training chapter.

Nathan Lambert [00:35:14]: And it's really nice.

Sebastian Raschka [00:35:14]: It's like, well, you can actually reuse everything. It's, it fits together.

Nathan Lambert [00:35:18]: Yeah. Like what we're doing on OMO, we can baseline our instruction fine tuning in our fine tuning code base, which also has some RL things and in our pre-training code base. So it's nice to have both, but that is definitely why it's simpler. And the RL is only getting worse in my mind, I think. Like we've seen that LLAMA has used rejection sampling for two iterations and there's no public implementation of rejection sampling that at least public enough to know that people have actually trained models with it, which is the idea of ranking completions to a reward model and then running instruction tuning again on the top completions.

Sebastian Raschka [00:35:54]: I think also in the recent LLAMA 3.1 paper, they used rejection sampling with DPO, for example. Like they didn't use the RLHF with reward model, but then they used the reward model for the rejection sample. And yeah, so I must say, I have the code for the DPO. I wanted to do TPO because it's also more resource efficient. You don't have to train that reward model for, let's say the book, but I was not really happy with the quality of the output yet. So I must say it's like, okay, this is not, it's not helping the instruction fine tune model. And it's like, I think a general thing where I, I mean, you might correct me if I'm wrong here, because you are the expert in RLHF, but for me, it's like, it's like a optional thing where unless you need a specific style or need to deploy something in like a safe manner, it's maybe not giving you the best results. If you need a private model that just runs on your own computer and gives you correct answers, I don't think DPO or RLHF will make the answers more correct. They will just change how they look like.

Nathan Lambert [00:37:01]: And yeah, I mostly agree, especially on what we have in public implementations. The public implementations are really good at improving on like alpaca eval. But if I'm training a model that I actually want to use, don't worry about alpaca eval. I think I'm like the most annoying person internally running these experiments because I just get so annoyed when only alpaca eval goes up and be like, that has made the model worse. Like we've, I've been building internal demo tools, which is just like making Gradio better and showing how to use VLLM for serving. But it's like a lot of the models we put out for research are like really, really annoying to talk to. You put no yapping or just be concise in the prompt and it doesn't do anything. So like a lot of the open datasets, and this is something that Nibetron and Lama3 have shifted to is this new evaluation, which is like IF eval, which stands for instruction following eval, which I think is a great one. So it's like write a response with less than 300 words or something. And it has these verifiable claims. And this is something that the Nibetron report showed that like doing fine tuning really unlocked a lot more performance in the DPO stage. So I'm hoping that we start to get more evals than just alpaca eval that are helped by this RLHF and that'll help the whole ecosystem come forward because it is in a kind of young, rough state right now. Yeah.

Sebastian Raschka [00:38:21]: And also one last thing about this topic is for me, like you said, the last sentence is kind of also one of the reasons is where I was like, okay, if I include something on DPO as the last chapter, I don't know if it's still going to be used next year or if there's so many variants, ORPO and QTO. And I mean, right now, I mean, Lama3.1 used DPO, which is like a big endorsement. But to be honest, I'm not sure if this exact variant is here to stay.

Nathan Lambert [00:38:47]: And so I think DPO is here to stay. DPO will be a canonical example, much like PPO. But I think the things that people are using will go away. Like PPO has stood the test of time of multiple eras of RL. So I don't think that people use it in its exact form, but people are always looking at it. And same with DPO, just because DPO is so simple. Like the exercise, this is like one of the best getting started with RLHF exercise is taking like the hugging face trainer and modifying it to use the DPO loss because you could use all the other infrastructure for like most of the infrastructure for batching and stuff like this. And then add that loss function, which is a few lines of code. And like, that's a good, that's like the entry point to doing RLHF implementations. Like when I interview people, I'm like, make sure that they have looked at this DPO loss function before. And if they haven't, I'm like, I don't know if you're in the weeds enough. I feel like you should look at this.

Sebastian Raschka [00:39:37]: Speaker 3 And if you need, if you are listening to this and you are about to get interviewed by Nathan, I will hopefully have by next weekend a tutorial on DPO, on implementing it from scratch. I was, this weekend I used actually Lama 3.1 to make a synthetic data set for that and got much better results. So it looks good enough to probably upload it next week. So nice.

Nathan Lambert [00:39:58]: Okay. Let's shift gears into like AI research and AI education, which is I think the thing that you have some of the most insight into. So you're a head of AI newsletter. You, I wasn't originally reading it when I subscribed, but now I almost always skim through to kind of see what papers you uncover. I'm pretty interested in like how you select papers, like how much you actually prioritize reading papers and why, and just like any advice for people, because it's hard to sit down and do this. And I, speaking for myself, sometimes writing is like how I force myself to read some papers. I don't know if you're in the same boat, but like, what is your worldview around reading AI papers these days and skepticism or excitement, everything?

Sebastian Raschka [00:40:42]: Yeah, that's a big topic. So I must say, so I, I look at more paper than I actually literally read. I mean, I look at the abstracts and the titles and then that's like a huge funnel as a section

Nathan Lambert [00:40:54]: processor.

Sebastian Raschka [00:40:54]: I must say for like, I was an archive moderator for the machine learning archive a few years back and that got me into the habit. So how it worked was basically as a, maybe it's useful because some people complain when

Nathan Lambert [00:41:06]: How did someone become an archive moderator? I didn't know that it was like a community position.

Sebastian Raschka [00:41:12]: So that was originally by Tom Dietrich. He was doing it by himself and he was looking for people to help him with that. Because as you mentioned, there is an ever increasing number of papers. And so how it works is essentially that when you submit a paper to archive, you select the categories. But a lot of people, they select not, let's say the correct, I wouldn't say not correct, but like the preferred categories because Yeah, the AI and ML.

Nathan Lambert [00:41:39]: It's like ML, AI, and then everything else. Yeah.

Sebastian Raschka [00:41:42]: And AI in archive is interesting. It's more like the classic AI. It's like, it's not LLMs. It's more like symbolic AI, that kind of stuff.

Nathan Lambert [00:41:51]: What do you think the difference between, or like as an educator, how do you define AI and machine learning? This was also one of my favorite interview questions to like see where they're at.

Sebastian Raschka [00:42:00]: Well, right now I would say I go back and forth on that. Right now I would say AI is this big umbrella thing where you have deep learning and machine learning as subfields. But if you think about it, if you consider a logistic regression classifier, it is essentially machine learning. And if machine learning is the subfield of AI, you would say, okay, then logistic regression must be AI. But is like classifying iris flowers really AI? I don't know. So today I would say

Nathan Lambert [00:42:28]: I also think about search as AI. Yeah. Like, yeah.

Sebastian Raschka [00:42:31]: Like, yeah. So there's like the good old fashioned AI. So I would say with AI, yeah, you have both, you have the machine learning and deep learning branches, but you have also, you can also implement AI with if else statements, I guess, like, you know, like, so. So that's how I would define AI. But I think nowadays when people talk about AI, they mean specifically gen AI, like generative AI models, like LLMs, stable diffusion, that type of stuff. But yeah, so the archive thing. So just briefly, basically there is in the background, it's also using machine learning or NLP to detect whether the title based on the title and the abstract, if the category is actually matching. And if there's a mismatch or in general as moderator, you go through them and, oh, this looks good.

Nathan Lambert [00:43:17]: This looks good.

Sebastian Raschka [00:43:17]: This looks good.

Nathan Lambert [00:43:18]: They started exposing this to the user. So I submitted a paper recently under ML and I was like, this looks like language. And I was like, I've been in moderate, I've gotten papers stuck in moderation. So I was like, I'm always going to hit, except if they tell me it might be in the wrong category, because archive moderation is a black box that you don't want to get stuck in. No, no, like as a user, but I understand the service it's providing. So it's good to expose that to the user. And if anyone's listening, just click it, click. Yes. It's not worth delaying your release. We get stuck in moderation and help archive out. Yeah.

Sebastian Raschka [00:43:50]: And so just the last thing on that is by default, everything gets accepted. However, sometimes it's something gets flagged. If there's duplicate content, if it doesn't look like a paper, sometimes people submit like one page blog posts or something. So there is this thing where sometimes there are also false positives and then it gets stuck. But long story short, that got me into the habit of reading the titles. And that's what I still do. Also for my head of AI newsletter, I just look through the titles and select. How have titles changed?

Nathan Lambert [00:44:21]: Like titles have changed a lot though, as I feel like they used to try to be. Accurate. Mostly descriptive. Yeah. Descriptive, right? And now they are a mix of, it's more of a storytelling than descriptive. I think it's the right way to tell it.

Sebastian Raschka [00:44:36]: At least we don't have the, it's all you need anymore. I feel like this went away finally, but yeah, you're right. It's more.

Nathan Lambert [00:44:43]: It ended with Ryland Schaefer's test set. Training on test is all you need. Yes. Did that make it on archive? It did.

Sebastian Raschka [00:44:51]: I think I also had it featured in my newsletter one time. I think. Or not featured, but at least mentioned. And so how I select papers is also often selfish. I read or select papers for the newsletter that I find interesting. And because I think this is also for education. When people ask me about how I would suggest doing things, I think the most important thing is to talk and work on things you are interested in. I think it would be really hard to do a good job if it's a topic that is not interesting to you. For example, I know, I don't know. R, sorry, or Rust is interesting, a very important topic, but I'm not into it. So I don't try to, let's say, make videos or content.

Nathan Lambert [00:45:35]: Yeah.

Sebastian Raschka [00:45:36]: So it's like, I think if there's something you're excited about, I think it comes almost naturally that you want to talk about it. So in that sense. So the newsletter, I almost, it's weird, but I almost write it for myself. It's like, I find it interesting.

Nathan Lambert [00:45:49]: How much do you spend reading versus writing when you're reading these papers and writing a blog post? I'm guessing a lot of it is just the natural process of synthesis is what you put into the newsletter. It's not like you're doing it from my read. It's not like you're doing a ton of scaffolding and editing after the fact, which seems similar to what I do.

Sebastian Raschka [00:46:09]: Yeah, you're right. I don't do, I don't spend too much time on it in the sense that I wish I could, but I have a full-time job. It's literally just the weekend project where I aim for one newsletter per month. Of course, I would like to do more, but there was also a book to write on weekends or sometimes I'm doing videos. It's like keeping it fun, you know, like where it's like, okay, this is not a chore. This is something that is supposed to be fun. Like in that sense, I read a paper and then I take notes and then I collect them and spend maybe half an hour, an hour to polish them a bit up or make some figures. And that's it per paper, I would say. And so I also don't write the whole newsletter on one day or one weekend. It's really spread over the month. I read a paper. Oh, this is an interesting one for other people. Let's write this up basically. And then this way I collect material over the month and then.

Nathan Lambert [00:47:00]: Yeah. What motivates you to work on this stuff? Is it purely like education? Because I, in some ways relate to that. I've been in that mode before.

Sebastian Raschka [00:47:09]: Yep. So if you have noticed, I don't have any sponsorships or something.

Nathan Lambert [00:47:14]: Never done that. Respect.

Sebastian Raschka [00:47:16]: I will never say never, but it's not something I do. It's really just a hobby. And I do like discussions that come around it. There's a certain satisfaction that if you put it out, it helps others and people tell you positive things about it. It's kind of very gratifying. I don't know. There's like a reward in a sense. And what's also cool is there are a lot of people. It's like being part of the community and exchanging information because there are also a lot of people who sometimes know something I don't know. And this is really, I think, really cool. You write about something and then someone, Hey, have you seen this? This seems like it's taking it to yet another level. Or this is the same idea. It's even better or something. And this is super cool where you get this effect where you learn by doing this, actually, because there's always someone who knows a bit more than you do in a specific area. So, yeah.

Nathan Lambert [00:48:07]: Yeah. I feel like it's increasingly important these days and increasingly impactful because so much of research has become closed off and for business reasons. So there's fewer people that do more of the work. I don't like it. I always feel like people don't realize how few people are informed and share on any given topic like AI research. If you take away three people, I've yet to find people that just tweet the same random RLHF crap that I tweet. It's like, I don't do it because I just say random things, but there's not that many people that represent each of these corners. Ahead of AI, I think Jack Clark's important AI. I should have him on the pod. I think I've talked to him a few times. He's great to talk to. And his is the same thing. It's like these few people that are disseminating AI information, which is crucial for policy at future angles. Have you ever gotten criticism that your work is accelerating AI and that you are a safety risk? I've gotten some critical emails that are like, you shouldn't talk about this.

Sebastian Raschka [00:49:07]: Yeah, I've more gotten emails about the fact that I talk about LLMs is not good because LLMs violate copyrights. I mean, not that I do it, but that other people's LLMs do it.

Nathan Lambert [00:49:21]: And I'm happy that I haven't had this audience very much, but it seems this is like one of the challenges of having a tech audience is like you cultivate it in kind of one of two, like there's multiple ways to go. And one of them is like this all data is for language models is theft thing. And I just don't know how to deal with it because like I disagree, but the normally people that aren't receptive to it, which is really hard. It needs to be played out. Yeah.

Sebastian Raschka [00:49:47]: My book also just to make extra sure all the data I use there is so the pre-training data is public domain data, like a book from Project Gutenberg. And for instruction fine tuning, I did my, I created my own data set basically. So just to avoid any issues, you know, like. Did you do, you wrote it by hand?

Nathan Lambert [00:50:06]: Yep.

Sebastian Raschka [00:50:06]: So I took, no, actually I used, I used part of an LLM and some by hand.

Nathan Lambert [00:50:12]: Yeah.

Sebastian Raschka [00:50:12]: So it's a great exercise.

Nathan Lambert [00:50:14]: Yeah. Yeah.

Sebastian Raschka [00:50:15]: And for the synthetic one, I use LLAMA 3.1 now too. I mean, yeah, you can tell me also about that a bit. I mean, that's maybe interesting for the audience, how to generate a preference data set, because there are multiple ways, I mean, naturally it's crowdsourced, right? So you ask people, you have the model generate two answers or have flavors of the model generate answers and then, oh, which one do you prefer? But it's not really scalable. And so you could technically do the same thing with an LLM. You could basically have the LLM generate a more polite version because I think LLMs are very good at, even the small LLMs, the open source 7b models are good at rephrasing things or evaluating things. They're not necessarily good to generate the answer in the first place if they don't have a reference, but given a reference, I think it's super useful to use open source LLMs in that sense.

Nathan Lambert [00:51:07]: I'm surprised that this hasn't caught on sooner, but I think it's starting to catch on. I think in the meta report, they essentially have edits. So then they rank, they make their preference pairs as edited better than chosen, better than rejected. And that's like, you can create multiple players by binarizing. There's a few research projects that have done this where they have like, constitutional AI is popular, but that's not really reproduced. One of my collaborators slash friends at Synth AI Labs, Louis Castricado, he did a paper on like the pink elephant problem, which is like using provisions to get the model to not just say whatever is in the question if you ask it not to. We did a follow-up work that's out literally today, which is like on self-directed synthetic dialogues where you have the language model generate a plan, and then it follows the plan. And then you can also do revisions on it. So I think Nemetron did this with Prompt. So it's really getting going, but it's something that took longer than I expected. There's the kind of question, this is like too big of a topic to go into, but it's like, how do you use GPT-4 feedback? Do you use like, are your completions from two different models or the same model with different generation settings? How do you use humans? I think that the labs are using humans for preference data because it eliminates some of the problems in language modeling. And then that's one of the biggest impactful research questions in alignment. It's like, we can't afford the $1 to $10 million dataset. How do we do this? And that's what, we're starting a project to do that AI too right now. And it's a big open, like, I don't know where it'll go. I don't know how much, like how far can we reproduce the LLAMA-3 alignment methods. Yeah.

Sebastian Raschka [00:52:46]: So I would say the LLAMA-3.1 paper or the LLAMA-3 paper, it was like a 93 page paper

Nathan Lambert [00:52:52]: and it was great.

Sebastian Raschka [00:52:52]: I love it. It's like a lot of detail, but on the alignment part, I feel like I wish there was more information

Nathan Lambert [00:52:58]: about it.

Sebastian Raschka [00:52:58]: Even like LLAMA-2 had more information where they showed what is the improvement actually over the different stages when they added to supervised fine tuning.

Nathan Lambert [00:53:05]: So I'm talking to Ross Taylor tomorrow, and I'm going to ask him the specific thing. On latent space, like Thomas S., one of the leads, said that most of their gains come from RLHF rather than SFT. So I think the open source community is over-indexed on instruction fine tuning because it is accessible and we have the data. And this is like one of my, like, try to guide the community by doing things is like, go do RLHF. Don't worry about instruction tuning data sets. Don't worry about that. We'll just leave that the same and go find more preference data and keep playing with this. And don't worry about the DPO methods. Just literally go make preference data and keep trying to train things. Like don't implement a new loss function.

Sebastian Raschka [00:53:48]: Practical question to an expert like you. How good is actually a preference data set if you download it, if both the chosen and the rejected answers, if you download a preference data set, they're not generated by your model, right? And if you have a model and you use the responses that the model has never basically seen before, does this actually work or would it be advisable?

Nathan Lambert [00:54:11]: So the most, the two most popular preference data sets in the open right now are UltraFeedback and Nectar or variants of them. Both of those are collected from large suites of other models. And part of my, there haven't been data sets or papers that have trained really good models using on-policy preference data from the model you're training. And I think that's a question that we need to answer. It's like, how do we get UltraFeedback level results with on-policy data? Because all the labs are using on-policy data. I wrote about this in like Barry to one article. I have a theory that UltraFeedback and Nectar, these general data sets work so well because within them, there is something close enough to your distribution and you don't have to get it quite right. But it's just like a gentler, more uniform learning signal for the models doing preference tuning. But we don't know. That's something that I want to answer.

Sebastian Raschka [00:55:02]: Yeah, this is an interesting one. I would also like to know the answer because that is one thing where I got a bit stuck when I was writing this DPO chapter with smaller models. I think bigger models also, they hide these weaknesses a bit because they have been trained on so much data that like you said, it's kind of in distribution already. But if you train a small model, it would be out of distribution, right? If you use someone else's preference data set. I noticed even something simple when you train a model on one simple instruction data set, let's say something like alpaca. And then let's say you have just to have something visual. You want the model to generate Yoda speech, like where every sentence is reversed. But the model has never seen sentences like that unless it was maybe in the training data. But in that sense, it doesn't work well at all because you ask the model during preference tuning to write sentence structures. It has never grammatically written before. And so in that sense, I think what I found is it's much better if you, I don't know, you say be more polite or like you have a more polite answer because you use the same grammar or so. So things like that basically. And yeah.

Nathan Lambert [00:56:08]: Yeah, I think that's a smart approach. It also might be why learning rates are getting so low. Where like all the learning rates for DPO and things have been going down in the fine tuning space. And it might just because distributionally, like we're far off from the model. There's the other theory that the model is like really, really done training. So they get it to a really good optimum. You don't want to move it from them. But it might just be that like our data sets are in the wrong space. Yeah.

Sebastian Raschka [00:56:32]: So you try to be gentler with a lower learning rate.

Nathan Lambert [00:56:36]: Yeah. All of this stuff changes fast, but not fast enough. Like this ultra feedback data set they were talking about came out last October. So we're like almost 10 months in and it's still the state of the art data set. And it's only like 50,000 examples. So there's so much opportunity for someone to like at this level, like go build data sets if anyone is watching. Because it's like, I think we're so far off where we could be just because people don't know how to make good preference data sets.

Sebastian Raschka [00:57:02]: Well, now we have LLAMA 3.1, 70 and 405 billion that allows us to do that, right?

Nathan Lambert [00:57:08]: We'll see. Yeah. I was wondering, this is a change of topic, but how do you think like, do you think AI will change our jobs in writing? How do you see AI coming for this kind of educational space? Like how much of what you do as an educator could be taken in N years by AI?

Sebastian Raschka [00:57:26]: Well, I think it's like, of course it will automate away some things because nowadays you would ask a model something instead of searching for it and reading it on a website. But I do think the creation process, you still need a human to put it together well. Because I don't know, I think LLMs are not nowhere near like generating a whole article that is actually, I would say even good where it can generate the right things, but you still have to put it together. It can generate good blocks of text or something like that, but you need to, as an edit, like you become maybe more like the editor then in that sense. But I'll try this.

Nathan Lambert [00:58:09]: Also like, do you write, do you have AI write any parts of your articles? I'm so scared for like moral reasons to have any AI writing in it. I'm like, it's just a slippery slope. It feels like I could get addicted. Yeah.

Sebastian Raschka [00:58:21]: So sometimes I don't have it write anything from scratch, but I sometimes do do that. And especially, I don't know, I have a, I mean, I'm a non-native language speaker and sometimes I have a harder time than other days to make the sound right. It's like, okay, this is what I want to say, but it doesn't sound right. And then I, can you revert this with a focus on XYZ or something? So like, it's basically like a, you know, like a thesaurus where you find similar words, you find similar sentences, like just rewording it, like these types of things. But one also, now that you mentioned it, one weakness it has, or LMs can't do really, is they can't generate figures. You know, maybe that's coming.

Nathan Lambert [00:59:01]: I don't know.

Sebastian Raschka [00:59:01]: You can do that probably with ticks, like the latex thing where at one point, but right now nowhere near, can you generate any useful figure? And I think learning is very visual too. I think if it's just text, it would be really hard to learn anything.

Nathan Lambert [00:59:17]: Yeah.

Sebastian Raschka [00:59:17]: So you can, of course, but I do think, you know, there's a saying, image is worth a thousand words, right? So yeah, in that sense, you still need someone, you know, like the mastermind behind an article, even if it's just an editor, I don't think LMs can replace everything at least. And we'll see. I mean, I don't know how much better, I mean, we just don't know how much better, let's say GPT-5 as a placeholder here will be then GPT-4, you know? So maybe if it's saturating, who knows, right? So maybe it will be five more years till we, yeah, get in a more scarier territory in terms of replacements, you know? So we'll see.

Nathan Lambert [00:59:55]: Yeah. I mostly avoid the agent word, but it does seem like there's enough culture and cultural investment in the Bay Area and tech executives to do something. Like they're going to get to something that is triable, which I think is mostly like automatic Google searching, more code execution, which is going to be interesting, but I have such wide expectations of what it actually means. That's probably the next big shift. I think this LLAMA 3.1 is probably right now leading the year in terms of AI news. This recent DeepMind thing on the math might be a better example of what's really hot news. I need to go read more about it. There's some long write-ups on how the qualitative between the AI math and the human math and the different directions they're going. So that's kind of what I want to read about it. But it'll shake things up. We're multiple years into this fast phase. It's not exactly new at this point. Yeah.

Sebastian Raschka [01:00:57]: Last thing on that is I do think, though, LLMs make good assistance in the literal sense where one thing where I use it for my newsletter for is at the end, I have a list of all the papers I have found interesting, like 30, 50 papers usually. And usually per hand, I edit the author names, like the last names of the first three authors. And now I use an LLM to go to the website and get the names of the authors, basically. And so this is where it saves a lot of time. You could do that without LLMs. You could write some code to do that, but it would probably take me half a day to write because I'm not good at this web scraping code to do that type of thing. And I think in that sense, it is actually a useful assistant for certain things like

Nathan Lambert [01:01:44]: delegating actions. I think it'll keep creeping up. I don't expect their usage for those things to go down because they already are so useful. And the little coding things, the hacking data together, the automatic searching, people aren't going to want to stop using that. I don't know if it supports the whole valuation we have, but it's fun to be in a space where we get to try new things. As a computer nerd, it's really fun to have a new type of software that we can try all sorts of things in our workflow. And I think that's underrated. So I don't know. Thanks for coming on. Any last things you want to discuss?

Sebastian Raschka [01:02:19]: Yeah, I just wanted to say thank you for the invitation and I hope you keep creating these awesome newsletters. I think this is much needed because there's so much hype, like you said previously, it's

Nathan Lambert [01:02:32]: creeping up on us.

Sebastian Raschka [01:02:32]: There's a lot of over, let's say, evaluation and praise. And I think something that is kind of like cutting through this is it's much needed like this honest, straightforward, no b******t content. So yeah, I hope you keep creating that. It was fun to chat. And yeah, to everyone out there, I think also what keeps us motivated, I think, is the awesome community that people give feedback and discuss things and bring things up. And yeah, I think without people also giving us feedback, we wouldn't be probably doing this because it's kind of a lot of fun to be in that space, I must say. Yeah, it's fast moving, but there's always something interesting every day.

Nathan Lambert [01:03:14]: Yeah. Yeah, this is really interesting. We covered a lot of kind of low level of just what it's like trying to use language models on the day-to-day basis in July of 2024. So thanks for coming on. And I'm sure we'll talk soon. All right.

Sebastian Raschka [01:03:27]: Yep, it was nice meeting you and see you then. Bye.

And how to understand Llama three point one's results.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/gpt-4o-mini-changed-chatbotarena

0:00 GPT-4o-mini changed ChatBotArena
3:23 Llama 3 in the arena
5:13 Partial solutions and next steps

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_013.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_015.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_019.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_021.png
Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_025.png
Fig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_039.png
Fig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_043.png

Defining the future of the AI economy and regulation. Is Meta's AI play equivalent to the Unix stack for open-source software?
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/llama-405b-open-frontier-model

00:00 Llama 3.1 405b, Meta's AI strategy, and the new open frontier model ecosystem
01:37 Meta's open frontier model
03:51 Zuckerberg's vision for open-source AI (vs. reality)
08:35 Does the Llama 3.1 license support open-source AI?
12:55 Different futures for regulating frontier models

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_008.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_010.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_015.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_018.png
Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_050.png

SB 1047, AI regulation, and unlikely allies for open models
The rallying of the open-source community against CA SB 1047 can represent a turning point for AI regulation.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/sb-1047-and-open-weights

00:00 Introduction
01:53 SB 1047 and targeting regulation
07:57 Unlikely allies of "open"
12:05 What would I regulate today?

I Switched to Claude 3.5
Speculations on the role of RLHF and why I love the model for people who pay attention.
This is AI generated audio with Python and 11Labs.
Source code: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/switched-to-claude-from-chatgpt

00:00 I Switched to Claude 3.5
03:57 Product priorities
05:15 RLHF's peak?

Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_016.png
Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_018.png
Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_020.png
Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_022.png