Ep 43: CEO/Co-Founder of Contextual AI Douwe Kiela Reaction to o1, What’s Next in Reasoning and Innovations in Post-Training

Douwe’s contributions to AI are truly a part of its bedrock foundations. He wrote the first paper on retrieval-augmented generation (RAG) and has raised over $100 million to help enterprises build contextual language models that fit their use cases. Before Contextual he was the head of research at Hugging Face, worked on the Facebook AI research team (i.e. Llama) and remains a professor at Stanford. Douwe was incredibly open about his take on AI’s recent history and where he thinks it’s going.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  Dawal Kiela wrote the first paper on retrieval augmented generation and is the CEO and co-founder of Contextual AI, a company that's raised nearly $100 million to help enterprises build customized contextual language models for their use cases. Dawal was previously the head of research at Hugging Face, he spent five years at Facebook AI Research, and he's an adjunct professor at Stanford. He was the perfect guest on Unsupervised Learning to talk about a bunch of things, including innovations in post-training, his view on OpenAI's latest O1 model and reasoning in general, barriers to enterprise deployments, and the research areas he's paying attention to. I think folks are really going to enjoy this episode. Thank you so much for coming on.

• 0:33Speaker 01

  Yeah, thanks for having me.

• 0:34Jacob Effron?

  So I think we're talking one day after the O1 release, and so I'd be remiss not to start there and just ask you, what do you think of yesterday?

• 0:43Speaker 01

  Yeah, I think it's very exciting. I think a lot of us in the field have been thinking about test time compute, basically, right? And I think one of the things that's very interesting about the release is that it's going in this much more important direction. I think of thinking around systems rather than models, right? So essentially what it's doing is it's compressing a lot of the chain of thought, ideas that have been around for a while, into the model using ROHF, which has been around for a while, and turning that model into a more complicated system, essentially. So I really think it's very encouraging to see that this works so well for reasoning specifically.

• 1:25Speaker 01

  My own team has been working on very similar ideas, but then much more on the retrieval side of things, right? So how can we compress a lot of this systems thinking into models? Or at least make sure that we address the entire system and not just kind of having this one thing that predicts the next word and that's it.

• 1:41Jacob Effron?

  Right. And now that we have a model out there like this, I mean, should we expect most future models to follow this approach?

• 1:47Speaker 01

  Not necessarily. So I think it depends a little bit on the latency constraints of the deployment, right? So when you have to do a lot of thinking during test time, your latency goes up. So it's not always the best thing to do. I think there are all kinds of different trade-offs that you have to make when you don't think about AI deployments.

• 2:04Jacob Effron?

  I thought it was interesting in the benchmarks yesterday, obviously, it's much more powerful on certain things like math and law, but other things... Actually, the old models do better and they're certainly much faster.

• 2:13Speaker 01

  Yeah.

• 2:14Jacob Effron?

  I think it'll actually be, it'll be so interesting to see what UIs are built around the fact that like you might have to wait. I mean, this one, the wait time isn't crazy, but as you could imagine, over time, there'll be models that you wait hours to get something back from.

• 2:24Speaker 01

  And yeah.

• 2:25Jacob Effron?

  How do you kind of build?

• 2:25Speaker 01

  Especially if you have like very long context and there's a lot of like complicated reasoning that you have to do. Then your thinking time goes up, but the same is true for humans, right?

• 2:34Jacob Effron?

  Right, that is, that is very true. When like a new model like this comes out, like, you know, you're a guru in the space. Like, how do you figure out if it's, you know, what it can, can't do? Like, obviously, I'm sure you look at the benchmarks, but.

• 2:44Speaker 01

  Yeah.

• 2:44Jacob Effron?

  What do you do to play around with it?

• 2:46Speaker 01

  Yeah, so it's mostly looking at the benchmarks and so I have a very good team. So I ask them what they think and they give me their, their opinion. Yeah, it's always good to follow along closely with these kinds of releases. For us, I think it's very encouraging that we've been thinking about very similar things, but then in the retrieval space. So I think it's exciting.

• 3:06Jacob Effron?

  Yeah.

• 3:06Speaker 01

  I, I think it's possible that some of the ideas that are in O1 were already in other places, models, without them telling everyone explicitly. I think you can guess what I mean there. So, so this idea, it's not super new, but, but as usual with OpenAI, you have to give them credit for, for how they executed on it. I think they did a great job there.

• 3:26Jacob Effron?

  Awesome. Well, I think that's probably a good segue to maybe talk about contextual a bit. And I think, you know, there's a lot of interesting angles I'd love to dive into today, but maybe the first is, you know, you started this with kind of maybe a different thesis to model building than, than some of the other foundation model players. And I'm curious, like, you talk a little bit about how that came about and, you know, over the course of building the company, you know, the milestones that have kind of given you conviction in that approach.

• 3:48Speaker 01

  Yeah, yeah. So, so why we started the company was because we saw this frustration in enterprises where everybody was super excited about Gen AI. Like this is going to change the world, but at the same time, it just wasn't ready yet for primetime, especially not in enterprises, right? And, yeah, we knew that RAG would be one of the solutions to a lot of these problems. We also knew.

• 4:12Jacob Effron?

  You knew that because you came up with it.

• 4:13Speaker 01

  Because we came up with it, but we also knew we could do much better than RAG.

• 4:17Jacob Effron?

  Yeah.

• 4:17Speaker 01

  Right? So RAG was a. Old idea and so the original vision was actually much more ambitious than than what ended up in the paper. And so what we're building with contextual is that much more ambitious vision. And and so I think how how we're different from an OpenAI or an Anthropic is is really that we think about things from kind of two core principles, I think, that are different. One is systems over models. So I just mentioned how encouraging it is that OpenAI is also thinking about systems, but

• 4:50Speaker 01

  they're still compressing all of that in a model and we think a model is maybe ten, twenty percent of this much bigger system that has to solve the problem. Right? And what you want to buy, if you're an enterprise, is the system, not the model, and then you have to build all the system around it and it doesn't work and it's you know, very complicated to build amazing AI systems. And we are very good at that. And the other thing I think that we do differently is that we do not do AGI. I think AGI is a great idea. I think fundamentally it's a consumer product. Because you have no idea what the consumer really wants, right? So you need a generalist intelligence to satisfy their needs.

• 5:29Speaker 01

  But in an enterprise, you often know exactly what you want from the user and you often do not want it to be a generalist, right? You want it to be specialized for this one particular use case. You don't want it. So this came up recently in a discussion with a bank. So in the European Union, if you use your AI system for doing a performance review that is very, very bad, you will get very heavily sanctioned. So now you have this generalist AI deployment and you need to like constrain everything back again, right? So I think the the right way to think about enterprise AI is really through specialization, not generalization.

• 6:05Jacob Effron?

  Yeah, and you guys obviously do a really compelling job kind of tying the whole retrieval process, you know, in the model directly to the the outputs of large language model as well, versus separating them out.

• 6:14Speaker 01

  Yeah, so when you have that that systems approach and with this specialization component, you can really just say like we end to end specialize all the parts together. It's a very integrated system. And then we focus specifically on high value, knowledge intensive use cases where that really pays off. Like being extra integrated and extra specialized.

• 6:33Jacob Effron?

  Anytime we talk to some of these end customers and there's like a company selling one very specific part of the infrastructure stack or one, you know, and they're like, yeah, you know, put us together with like these other eight solutions and you got something. They're like, I feel like they just, those companies get dragged into like, actually, can you just please do everything for us? Or please just like, help us solve this problem. And so I imagine that really resonates with the, with the folks you talk to.

• 6:53Speaker 01

  Yeah, so for us, we're kind of taking the opposite approach. And I think that goes against the startup playbook in a way, right?

• 7:00Jacob Effron?

  Yeah, do like one very specific thing.

• 7:01Speaker 01

  Yeah, it's like do this one thing and do that really well. And so you get this layered cake, which is what you would use to build Frankenstein's rag essentially. And we're saying we're just slicing vertically through that cake. And so that gives us a very nice compounding effect where, because we control the retrieval and the reranking and the generation and the post-training and alignment and fine tuning and all of that. That just makes us very good at solving the problem. And that's what you care about in the end.

• 7:28Jacob Effron?

  Yeah, and I guess I'm curious on the enterprise side, it feels like maybe from my seat, you know, the, the, the things that have killer product market fit seem to be anything in co-generation space. You know, customer support, obviously, and then a lot of this like internal search, you know, chat your docs type stuff. But you guys see this at a far deeper level. What else are you seeing in terms of patterns of things that are actually getting deployed?

• 7:48Speaker 01

  Yeah, the pattern I see in terms of things getting deployed is that there are still just lots of demos happening. And yeah, we are focused exclusively on production deployments, which is very different, right? So I think we still see a lot of very compelling demos built in this Frankenstein sort of layered approach.

• 8:10Speaker 01

  That then when they get to real user testing, just fail completely. That's not even just about like the machine learning side of things. It's also like the deployment, like everything around it, right? So like risk and compliance and security and.

• 8:25Jacob Effron?

  What are like the fall on your face things you see like, you know, amazing demo and then like, okay, that did not work.

• 8:31Speaker 01

  Yeah, so demos are often built on like, I don't know, like 20 PDFs or something, right? There's a lot of climbing, like hill climbing directly on the test set happening with those dots, right? So, yeah, directly hacking the test set. And then when you, when you expand that to the 10,000 PDFs it has to work on, everything breaks down completely.

• 8:51Jacob Effron?

  You were the author of the first paper on RAG. So like, so now it's become, everyone knows what it is, it's become, you know, the standard way of doing things. But like, how did that work come about? Like, maybe talk about that a little bit.

• 9:00Speaker 01

  Yeah, yeah. So we always joked that we should have come up with a better name than RAG if we had known.

• 9:06Jacob Effron?

  It is kind of catchy.

• 9:07Speaker 01

  It's kind of catchy, but still, it, you know, it doesn't, doesn't really do it justice.

• 9:10Jacob Effron?

  What would you rename it if you had a choice?

• 9:12Speaker 01

  Yeah, something with the word contextual in it.

• 9:15Jacob Effron?

  Yes, yeah, with the company.

• 9:17Speaker 01

  Exactly, yeah. So then I would have to name the company something else, I guess. I don't know. But, yeah, how that came about was I have always been thinking about grounding in my career. So my PhD was on grounding language in perceptual information, right? So the very early multimodal systems at the time. So like trying to understand the meaning of the word cat much better by knowing what cats look like. So integrating pictures of cats into your machine learning systems, your NLP systems.

• 9:51Jacob Effron?

  Pictures of cats always crucial for most.

• 9:52Speaker 01

  Exactly, that's where you start, right? And then, then when we were at FAIR, I was working with Ethan Perez, who was my PhD student, and we were thinking about this grounding thing, like can't we just ground in Wikipedia? So we built this amazing prototype, sort of like the first version of RAG that actually worked. And that was mostly possible because of the amazing work by the FACE team. So Facebook AI Image Similarity Search, I think is what it stands for. So that's the archetype of the vector database, right? The very first vector database. So anyway, it was very easy to just try out that idea. And then what we had to figure out was how do you back propagate into the retrieval

• 10:35Speaker 01

  mechanism so that you can actually train the system? So we didn't even consider. Consider that in the future, most people would not backprop into the retriever because that's too difficult. They would just use everything off the shelf, right? So we had this nice prototype working, but then unfortunately, for I think visa reasons or so, yeah. University reasons, he had to go back to university. He wasn't allowed to work with me at FAIR anymore. So then there were these other people in London. So Patrick Lewis and Sebastian Riedel, working on very similar problems in open domain question answering, which was like the killer use case for RAG. So then we teamed up and out of that came really amazing collaboration. So I miss those days.

• 11:21Speaker 01

  That's really like the golden era of FAIR for me. After that, it became a bit less good.

• 11:27Jacob Effron?

  I guess at the time of publication, was it kind of clear to you? I mean, it's just become such a standard paradigm way to do things. Was that apparent to you at the time or was it like, oh, we'll see what people end up doing with this?

• 11:37Speaker 01

  It's never really apparent. I think anybody who would say yes there is probably lying, right? Like when papers come out.

• 11:44Jacob Effron?

  I mean, the transformer paper, I guess, is the canonical example.

• 11:46Speaker 01

  No, well, Mike, so I think that history has been completely rewritten on that paper, honestly. So when the transformer paper came out at FAIR, we were very underwhelmed. We thought it was a little bit silly because we had RNNs with attention. And it's like, oh, like if you cut the recurrence, then you can paralyze it better. It's like, no shit, Sherlock, right? And it didn't even actually work that well on the machine translation benchmarks that they were evaluating on. It was tiny differences and it was all very finicky and stuff. So we thought like, why not just do kind of sliding window attention, aka convolutions, right? So convolutional neural networks. So I think some of the folks at FAIR were working on basically the same idea,

• 12:29Speaker 01

  but then using the convolutions paradigm. So that came out a couple of months later, maybe around the same time. Didn't get the credit and then Google proceeded to rewrite history where they invented like the AI revolution. This is something I think that actually just happens with historians in general, sort of like the great. Man storytelling, right? Or so great person storytelling, I should say. But so in in scientific history, we we do the a version of this where it's like the great paper. But that's not how it works. Like there was so much stuff going on at around the same time. That it was very obvious that this was something you could do. It was mostly actually about the engineering contribution. Like, how do I get this to work so well?

• 13:14Speaker 01

  The reason transformers work so well is not because they're amazing, it's because they they are optimal for GPUs.

• 13:20Jacob Effron?

  Yes.

• 13:21Speaker 01

  So, yeah, GPUs are the reasons for transformers, not the other way around. So maybe Nvidia deserves all the credit. Or or I would say the real credit should probably go to the people who invented the attention mechanism. Because that was really a radical new way to distribute information across sequences.

• 13:39Jacob Effron?

  Well, one, the history is still being written, so we'll see what what what specific points we point back to. But, you know, I think for most, a lot of discoveries, you know, a lot of times there's simultaneous work happening in a bunch of different places. Yeah, I guess that's been true since you mentioned calculus, right?

• 13:52Speaker 01

  Yeah, yeah, exactly.

• 13:53Jacob Effron?

  I guess like on the on the rag side. You, I mean, I'm sure enterprise comes to you all the time and they're like, OK, here's the seventeen things I want to do. And I'm sure you have this mental model of like, yeah, those seven good idea, those ten like never going to work with the current models. Like, what's your mental model today of like what you can and can't do?

• 14:11Speaker 01

  Yeah, so so I think the the main question we we we often get from customers is can I put this in front of my customers? And we tell them to be very careful with that. So the higher value the use case becomes, the the more risky it becomes to directly expose that to your customers. So we're really trying to help enterprises find the optimal sort of ratio of AI to human. Very much keeping humans in the loop, solving problems that are within reach now. And then over time, you can make it more and more complicated. I I think companies are becoming much more pragmatic about that now, actually. Like, the the first conversations we had when we were just getting started, we we had companies go like, I want this like investor application.

• 14:56Speaker 01

  And it's just going to make investment decisions and.

• 14:59Jacob Effron?

  That sounds great. If we're ready for that, let me know. I'll start looking for other jobs.

• 15:04Speaker 01

  So yeah, it's not it's not that easy, right? So the more concrete you can make it, the better. So if you give investors great tools so that they can make better investment decisions, that's what you want to do.

• 15:17Jacob Effron?

  Yeah.

• 15:18Speaker 01

  So I think we're quite a ways out from really just replacing people in those kinds of roles.

• 15:25Jacob Effron?

  Yeah, I know you've obviously, you know, a big focus of your work in textual has also been around alignment and a lot of the work you're doing on reinforcement learning. And I think what would be helpful for our listeners, if you could just explain some of that work, but also like it feels like that space is moving so fast and there's been all these improvements. And obviously there is, you know, the reinforcement learning from AI feedback, DPO. I mean, there's like a soup of acronyms, I feel like, in the space. And so it might be helpful to just kind of contextualize what's been happening.

• 15:49Speaker 01

  I like that word.

• 15:50Jacob Effron?

  contextual for contextualize, you know, what's been happening over the last year and then, you know, why you guys have chosen to focus here and what you've been doing.

• 15:56Speaker 01

  Yeah, I think alignment is a super interesting problem area. That is really kind of getting at the core of like, how do we make systems maximally useful for the end users? And so as you know, right, if we take a step back on RLHF, what's really the secret sauce that made ChatGPT work? There were some instruction tunings or SFT, right? So just next token prediction. But then you can't really capture human preferences all that well. So what you ideally want to do is capture it at the full sequence level. So not at the next word level, but the full sequence level. So then the right way to do that is through RLHF or some similar method.

• 16:37Speaker 01

  So RLHF is great, but it has two big problems. One is you need to train a reward model and so that you can propagate those rewards back to your sequence, right? And that model needs to be pretty good itself, right? So it's very expensive to train that model up. And then so you're kind of throwing that model away. You're never using it in your actual generation.

• 16:59Jacob Effron?

  Right.

• 17:00Speaker 01

  The other problem is that you need to have preference data. So that means that, let's say you have a real deployment of a system and you get like a thumbs down like, oh, actually, this was the wrong answer. Now you need to go to a data annotation company, or you need to have somebody internally annotate what the thumbs up would have looked like. So that's obviously very slow, very expensive. The more specialized your use cases get, the more expensive. Exactly, yeah. So ideally, you want to break both of those dependencies. So if you can do this without a reward model, that's much more efficient, right? So that's what DPO was about. And then what we looked at, contextual, together with Cowan, brilliant Stanford student, was

• 17:44Speaker 01

  can we break the dependency between the preference pair? So can we directly optimize on the feedback without needing to do data annotation? So we did that through something called KTO, an order acronym, Kahneman, Versky Optimization, after the behavioral economists, utility theory, and so prospect theory. So really brilliant work, I think, from Cowan. And so we recently went a step further with this, with Carl, who interned with us, where we looked at the kind of under-specification problems in alignment. So one of the big problems with these preference data sets, or even if you have individual, like thumbs up, thumbs down, is understanding what the real preference signal is.

• 18:29Speaker 01

  Yeah. Right? So you just have kind of a random ranking of like a bunch of answers, and you say like, this one is better than this one, but you don't really capture why, right? Or like what the underlying sort of causal structure is. So one way to make that much tighter is to contrast the revisions. So rather than having two options, you can say like, this is an option, there's something wrong with it, let's fix it. And now say like, this one is better than this one, but it's only a very small difference, and like the difference is to fix. Yeah. Right? So that means that it's much less under-specified. So that's called CLARE. And then when we were looking at CLARE, we found that there's a very interesting relationship between the data and the model that isn't really captured in the standard ROHF, or even in the DPO/KTO paradigm.

• 19:14Speaker 01

  You ideally want to factor in sort of how good is the model that you are training your preference data on. If the model is Better than the preference data, which is possible these days, right? It's kind of a crazy idea, but so it's possible that that's the case. So then you don't want to learn from that preference data that like this this good one is what you should be doing. You just want to learn like this one is better than this one. There's still information in there, but you you want to get the right information out, right? So just the ranking, not like this one is the right answer. So this kind of anchoring on on what the model quality is, we call that anchored preference optimization. So APO.

• 19:55Speaker 01

  And so I'm very excited about this because you can just see, and we see this ourselves in in customer data sets. You can just see this work much better and you have much more control over what the the quality of the data set means for your model quality after you've trained.

• 20:10Jacob Effron?

  I mean, it sounds like you're doing a lot of reinforcement learning with your customers for whatever their use cases are like. How does that actually work in practice? Are they having people like, you know, rank these things? Are you are they staffing folks to label data? Are you just using external parties? Like, how does it actually work with the customer?

• 20:22Speaker 01

  Yeah, so we have our own data annotation team, the our data engine. And our customer deployments, when we deploy with them, they give us a lot of feedback. So we have a thumbs up, thumbs down mechanism, and we can learn from that very directly. Using our algorithms, which you couldn't really do with a standard RLHF approach. So, yeah, I'm excited about that. But honestly, a lot of this alignment work just happens for the core model.

• 20:47Jacob Effron?

  Yeah.

• 20:48Speaker 01

  Right, so just during post-training. I think that that is also one of the things that like. So alignment is a big part of this post-training, but post-training really is where a lot of the magic happens in AI, right? It's like, okay, we have this pre-trained model, but now we need to like make it actually be good at what we need it to be good at.

• 21:05Jacob Effron?

  Totally.

• 21:05Speaker 01

  And so, yeah.

• 21:07Jacob Effron?

  And you're obviously aligning it around a set of business use cases versus like, you know, the ability to both do Shakespeare sonnets and give you a really, you know, dry customer support answer the next.

• 21:15Speaker 01

  Yeah, exactly. Yeah, we we always say like, if you want to use AI, right? Like, let's say you work in finance or something, you don't need that system to know about quantum mechanics or Shakespeare or Japanese haikus, right? You need to be really good at that one particular thing, and you get that through. Alignment, right? It's just much more specialization, customization. And through that approach, we find that you can make it past this production bar much more convincingly, like integrated system, custom alignment, specialization. That, yeah, you can see real ROI from these deployments.

• 21:48Jacob Effron?

  Yeah, super interesting. I was thinking of other research you've been doing. You obviously, I think, recently released this really powerful, small MOE, open source model. And I'm wondering, like, you know, I think it obviously begs the question. It feels like a model of that size, you can kind of run on device. And it feels like there's this overall trend of, you know, potentially a lot of things moving on device. So wondering, one, what inspired the work? And two, you know, is that a trend you expect to see a lot over the next few years?

• 22:09Speaker 01

  What inspired the work is Nicholas.

• 22:14Jacob Effron?

  Usually it's a set of people.

• 22:15Speaker 01

  The first author, so it's all about people, right? And Nicholas is incredible. And so this is a collaboration with the Allen Institute, right? And I would say honestly that they have done most of the work here. So I'm not taking any credit for what happened. But there wasn't a high quality, fully open source mixture of experts model out there for people to play with. And yeah, that was just a big opportunity, I think, because you can see this trend where everything is moving much more towards smaller models that can potentially be deployed on edge devices. Nicholas had already done some great work with us on Grit, Generative Representational Instruction Tuning.

• 22:59Jacob Effron?

  You guys have done an impressive number of acronyms.

• 23:01Speaker 01

  Yeah, yeah. Every new thing we do needs a new acronym.

• 23:06Speaker 01

  So I recently hired the new head of marketing. He's been complaining about it, actually. But so with Grit, what we showed for the first time is that you can use the same model for the retriever and the generator. So the same weights, which means that you can cache a lot of the compute. Because when you think about a RAG system, you have to encode the query basically twice, right? Like once for the retrieval step and once for the generation step. So if you can reuse that compute and just kind of figure out how to combine the parts there, then you can be much more efficient. So what I'm excited by is like, can we take? Omoi and then combine it with great to give you like a super powerful rag system that you can put on your phone.

• 23:48Jacob Effron?

  Yeah.

• 23:48Speaker 01

  So this is like happening very, very soon, I think.

• 23:52Jacob Effron?

  I mean, one thing, it seems like you do all these, a lot of these collaborations. And obviously in another hat you wear, you're like an editor professor over at Stanford. And so, you know, and obviously came up in the academic world, like, how do you think about the role academia plays? And, you know, obviously, a lot of the research coming out of these labs and companies like you, you know, I think traditionally might have been done in an academic setting. And so how do you think about what makes sense, you know, for folks at Stanford to focus on versus companies like yourself?

• 24:16Speaker 01

  Yeah.

• 24:16Jacob Effron?

  Where's that all go?

• 24:17Speaker 01

  Yeah, so I think academia is super important for the progress of AI and continues to be extremely important. I think obviously the research that happens in academia has changed, right? So five years ago, at Stanford, you could still be like seriously looking at pre-training at scale.

• 24:38Jacob Effron?

  Yeah.

• 24:39Speaker 01

  At the time.

• 24:39Jacob Effron?

  Scale is very different.

• 24:40Speaker 01

  Yeah. And so now the scale is very different. It's just not really possible to do that anymore at the same scale to get like a full model out, right? But you can still do lots of very interesting research in pre-training, just with smaller models. But I think a real blessing to my point earlier is that post-training is so important that you can take a pre-trained model that somebody else trained.

• 25:02Jacob Effron?

  Your former employer.

• 25:03Speaker 01

  Like, I was gonna say that Zuck has generously donated to the world. And so from that starting point, you can do amazing research, right? In post-training methods and like better alignment methods, for example. So I think, yeah, academia will continue to play a super important role in all of this.

• 25:23Jacob Effron?

  And then I guess, you know, on the enterprise side, I mean, it seems like obviously one of the strengths of what you're doing is this end-to-end solution. So there's no like gaps of things, but I'm sure even, you know, I'm sure you have a wishlist of things. You're like, God, if someone else built X thing, that would just be so helpful for us to plug into our end-to-end solution. Like, what's top of your wishlist these days?

• 25:39Speaker 01

  Yeah, it's a very long wishlist, but.

• 25:44Speaker 01

  So initially, I think we were hoping that we could use off-the-shelf extraction systems.

• 25:49Jacob Effron?

  Yeah.

• 25:49Speaker 01

  It's just really hard to contextualize language models properly. Right, sort of hence the name of the company in a way. But so language models are like, most language models are just really good as long as you give them the right context. So how do you give them the right context? You need to do the right retrieval and sort of re-ranking if you have a more sophisticated retrieval pipeline. But you can only really do the retrieval properly if you did the extraction correctly. It turns out that extracting some information from a PDF is very, very hard. And there just isn't anything good out there in the open. So we had to build our own models for doing that. So that's kind of on the beginning of the pipeline. And so, yeah, there's just a lot of like, kind of boring stuff that needs to happen

• 26:33Speaker 01

  to build that into a generalizable system. And also, on the other end, I think evaluation, especially for enterprises, understanding sort of the risk of a deployment and the real accuracy of the system, hugely important. But there still isn't really a right way to evaluate systems that enterprises can rely on. So I would love there to be kind of a standard way to do that. We might come out with some ideas of our own in the near future. But yeah, that's really like a very underexplored area.

• 27:06Jacob Effron?

  Yeah. Is that like a solvable problem? Or does it end up just being really custom and bespoke to like whatever the company's trying to do? Like, and some people come on this podcast and I think said, convinced them, like, you know, part of building a good product is specifying upfront what you want it to do. And then you basically created your evaluation set through that.

• 27:21Speaker 01

  Yeah.

• 27:21Jacob Effron?

  And it sounds much more bespoke, but obviously, there's plenty of folks trying to prioritize this.

• 27:24Speaker 01

  Yeah. So one of the real problems is that a lot of people don't understand what they want.

• 27:29Jacob Effron?

  Yeah.

• 27:30Speaker 01

  Right. So what we do when we have a customer engagement, like, we really spend a lot of time just figuring out like, okay, what do you really want? What does success look like? Then we hill climb together to achieve that success in the sort of prototype setting. And then we productionize it. And I just think that's the right way to do it. And yeah, I'm worried that a lot of people are not taking evaluation seriously enough. They just have like a spreadsheet somewhere.

• 27:56Jacob Effron?

  Right.

• 27:56Speaker 01

  With like 50 examples and then the super high variance. It's just, it's not principled enough.

• 28:02Jacob Effron?

  What parts of it then will end up being common, though, across different customers? I mean, obviously, everyone probably has some sort of bespoke nature of what they're doing. But like, you know, it sounds like you think at least there's some sort of common set of tools or framework you may be able to build across. Like, what might that look like?

• 28:16Speaker 01

  Yeah, so it's the same system, right? You need to extract information from your data at very large scale, tens or hundreds of thousands of documents and not failing. Then you need to put that in some retrieval mechanism. It's not just one retriever anymore, I think, in modern deployments. So we use what we call a mixture of retrievers approach. And then on top of that, you contextualize the language model with that information. Then you do a lot of things on top of the language model. So I think the components there are relatively clear.

• 28:50Jacob Effron?

  Yeah.

• 28:52Speaker 01

  And they have commonalities across everyone. But there's still just a lot to gain from making the components themselves more specialized and then sort of tuning the knobs between the components to make everything more optimal.

• 29:06Jacob Effron?

  Totally. I mean, today evaluations are done through like a spreadsheet. Like, you know, in a few years, we hopefully have figured this out in a much better way. Like, how do you think people will do evaluations?

• 29:14Speaker 01

  Yeah, so one idea that we're working on that I'm very excited about is thinking about what developers really need, right? So I think one interesting observation is that what AI developer means has changed over time. It used to be like somebody who could do machine learning and knew a lot about PyTorch and things like that. These days, AI developers often are people who are very good at calling APIs, right? And so I think on the evaluation side, we should have something very similar, right? So instead of having a lot of kind of machine learning, traditional knowledge around sort of test set creation and statistical testing to ensure that you have a significant difference between two models.

• 30:00Speaker 01

  That's all like machine learning, data science. It's not what these API AI developers want to use, right? So we're thinking a lot about like, what does the evaluation framework look like? If you want to make this accessible to developers.

• 30:16Jacob Effron?

  Yeah, makes sense. I guess, you know, switching gears, you've obviously been on the forefront of a lot of this research over the years. I'm wondering like something you might have changed your mind on in the last year. Like a strongly held AI research belief that you now think differently on.

• 30:29Speaker 01

  Yeah, many, right? I think like one part of my job is that I have to change my mind very quickly when things happen and there is a lot happening in AI. I think, yeah, there are lots of them. I think synthetic data working so well was a bit unexpected for me. I didn't really see that coming. I think agents, even though it's still kind of ill-defined what it even means to have an agent, but agentic workflows with tool use, I think is actually much more possible these days than I expected and probably than I would have guessed a year ago. In general, like this kind of like test time compute.

• 31:13Speaker 01

  So I think when like chain of thought and things like that came out, I thought, oh, that's a cute gimmick and like you can like tip the language model and it will do better. And I think a lot of that work was very flawed from an evaluation perspective. But maybe I was a little bit too sort of academic elitist around those contributions where I was like, oh yeah, that's kind of silly. Like that's not a real like cool machine learning contribution, but it turns out it works really, really well, right? So I think it's really great that the ecosystem is looking at this from so many different ways and combining them in very exciting ways. And so going back to the where we started, if you roll out a lot of chain of thought

• 31:54Speaker 01

  things and then you train on those with ROHF, maybe fixing them with either a language model or humans, then you get a more powerful language model. It will be slightly less efficient during test time when it rolls out the actual sequence, but it will give you a better answer. It's like, it sounds sort of obvious, right? But, uh, yeah, I, I think, um, like, the parts themselves didn't feel that special. It's more like when you combine these ideas that, that we suddenly see these, these nice.

• 32:21Jacob Effron?

  Also, very challenging data problem. Unfortunately, the internet is not littered with a ton of step-by-step, uh, the way of doing things.

• 32:27Speaker 01

  Yeah, exactly. Yeah. But, uh, so, yeah, it turns, turns out that, uh, uh, language models are pretty good at thinking through things step-by-step.

• 32:34Jacob Effron?

  Yeah. Yeah. I mean, a lot of people, you, you mentioned synthetic data. I feel like a lot of people talk about, you know, in, in the, in the road to AGI that like, you know, uh, obviously we're, we're using a lot of the tokens that are out there and this question of like, you know, where do we get the extra data from? Um, are you, is that something you worry about? Like, do you think?

• 32:49Speaker 01

  No.

• 32:49Jacob Effron?

  Okay.

• 32:49Speaker 01

  I think that's misguided. I, I really don't understand where, where that comes from. I, I, I think somebody said this once and a lot of journalists kind of picked up on it and it became a thing like, oh, we're running out of tokens. Uh, we produce huge amounts of data as a society every day. Uh, and you can train on all of that data. Uh, and I, I don't think we, we are at the, the kind of rate of compute where, where we, uh, sort of train faster than we produce data. It's just that not all the data we produce is very high quality as a society. Uh, so there, there's an interesting trade-off between, uh, data quality and data quantity. Uh, so if you have lower quality, you need commensurately more quantity.

• 33:33Speaker 01

  I, I think that's kind of when, when you ask researchers like me, like, okay, what don't we have enough of? I would say high-quality data, but we have plenty of data to train on, then we can still learn from it. We just need much more of it. So that's one, one observation. Thing to utter is that that's maybe true for text, but if you go multimodal, like, we haven't even scratched the surface, right?

• 33:53Jacob Effron?

  It feels like video data is massive.

• 33:55Speaker 01

  So we, we were talking about like the meaning of cat and like looking at cat pictures, like, how many cat videos can you train on to really understand the concept of cat in your huge multimodal transformer, right? That we're only just getting started with this stuff. And like, I think that's one of the shortcomings still. Uh, and that's like my PhD was on this, right? One of the big shortcomings of systems is that they don't understand our world. They don't look at our world in the same way that we. So they kind of observe it by proxy through our linguistic behaviors, i.e. the text we produce on the internet. But that's kind of a poor proxy. Right? So if they can actually watch lots of cat videos, they understand much better how humans think about cats.

• 34:35Speaker 01

  Um, so, so I, I think like multimodality is really a, a, a great way to add even more data. And then the third point is that, um, there was this paper where they, they showed that synthetic data doesn't really work. Or like you can't train a language model on its own data. But so that paper is full of flaws. It's like, it's not actually a realistic scenario. This is again like sort of PR flag plant. Um, and, and it, it killed a lot of very interesting directions because people were like, oh, this is not interesting. I'm going to work on something else. Or like probably like companies didn't get funded in synthetic data because this silly.

• 35:09Jacob Effron?

  All the data leading companies were very happy to retweet that.

• 35:11Speaker 01

  Yeah, exactly. So, so, um, but synthetic data, if you do it right, is super powerful. And then if you combine that with smarter algorithms like KTO and APO, uh, then you, you can get super far without needing to do any data annotation. And without needing to do like super compute heavy things.

• 35:30Jacob Effron?

  Yeah. Um, what about like reasoning? Obviously, uh, I think we're seeing early, you know, early signs of it. But like one, I'm just curious, like what, at what milestone, you know, for you, will you say, okay, this model has, you know, reason? What are you looking out for in terms of this like wave of, of, you know?

• 35:45Speaker 01

  I think models already have reasoning capabilities. I like, yeah, and they have had for a long time. It's not like that just came out, right? That like, uh, maybe OpenAI wants to make it seem like that just became real and before that it didn't exist. But these models have been reasoning for, for quite a while. Um, so I, I think that's very interesting. Uh, obviously they can get better at reasoning still. I think there are easy ways to do that. A lot of it is just data driven. Uh, so this is different kind of data, right? So whereas like high quality, like if you hire a bunch of mathematics PhD students and you make them run through very complicated reasoning, then you can probably learn quite a lot from that. I think if, if you invest a lot in that data, and I think all of the, the big like AGI companies are doing that, then we still can.

• 36:30Speaker 01

  Improve quite a lot.

• 36:31Jacob Effron?

  Are there milestones that you're, I mean, obviously, I guess they tongue in cheek had the number of R's in strawberry. It's like one of the ones they they did, but like, are there are there kind of milestones you're looking out for? Like, oh, it would be really, you know, amazing when a model can.

• 36:41Speaker 01

  Yeah, so I have a fun, fun paper I did with Tristan, who's a PhD student at Stanford on meta linguistic tests. It's related to the strawberry one, but it's about self reference in a way. So like, understanding like, there are X words in this sentence, like, for what X is that true? Basically, right? So like self referential statements like that. And so we had a lot of fun like making this data set and writing the paper. It's so it's called I am a strange data set. So I don't know if you know this book from so Douglas Hofstadter, right? Go to Escher Bach, like amazing guy. He wrote this book, I am a strange loop, where he basically argues that human consciousness

• 37:25Speaker 01

  is a result of us perceiving ourselves and our own behaviors. And then we sort of come up with a narrative for for why this thing that we're observing is behaving that way. And that strange loops or becomes consciousness as you grow up. So like amazing theory, I think. And so I'm not saying that, you know, I believe in artificial consciousness or anything like that. But but if these systems can actually nail these problems, like that would be very interesting, I think. Because that's really like a sort of, you know, like you have mathematics, then you have meta mathematics. That reasons about the math. And that's really like something that I think systems can do now, even if they are good at mathematics itself. So, yeah, if if systems like score very high on that data set, then I would be very impressed.

• 38:10Jacob Effron?

  That's super interesting. Of all the research happening now, I mean, obviously, you mentioned your chain of thought. A few years ago, people were doing all this kind of work. Like, what's what's happening right now that maybe is like a little under reported that you could see in like a few years? Hey, this this could really drive some of the advancements in these models.

• 38:23Speaker 01

  So a lot of it is just very practical work, honestly. Right? So like, I just talked a bit about like mixture of retrievers. It's like, why would you have like one dense vector database? That's really like not optimal, right? So I think a lot of. The interesting work, at least that I'm excited by, is just like real pragmatic, practical, how do we get this stuff to actually work? And then how do we figure out the right product form factor for a lot of the AI research that we do? It's kind of a weird new paradigm. It's not like in old school like SaaS or something where you understand what your customer wants and then you just build it. There's this much more complicated product pipeline where the product is basically the

• 39:04Speaker 01

  behavior of your model, which is a reflection of the data that went into it and maybe like the model that produced that data in the first place. So it's just a very, very different way of thinking about products.

• 39:16Jacob Effron?

  Yeah. And the idea that like, I mean, I always find it funny that sometimes the most powerful, impactful thing you can do is just like an extra sentence you add to a prompt or something. Or like, you know, and like, it's not traditionally what you think of as the, you know, what the tech engineer would do in a product, but it actually can end up having a massive impact.

• 39:31Speaker 01

  Exactly, yeah.

• 39:32Jacob Effron?

  Yeah. I guess you have, switching gears to your former employers, you've obviously worked at some interesting places, maybe starting with FAIR. You know, was it kind of always clear that Zuck would buy a ton of GPUs and give this to the world and like, what do you, you know, is this something that can kind of continue going forward? How do you kind of think about that?

• 39:48Speaker 01

  No, we, so we never had enough GPUs. Probably the people who work there still complain about GPU shortage.

• 39:55Jacob Effron?

  I don't know if anyone feels like they have enough GPUs.

• 39:57Speaker 01

  Yeah. Yeah, I think FAIR is a really incredible place. When I joined it, that was 2016, it was basically like a year old, I think, that they were really kind of taking it seriously. And yeah, it was very, very free and academic, maybe a little bit too free in a way. I was working on like Wittgensteinian language games of like multi-agent systems, talking to each other in emergent communication protocols and things. All of that, by the way, is going to come back. Like this year or next year, like I was working on multimodal and that became.

• 40:29Jacob Effron?

  Yeah.

• 40:30Speaker 01

  So I think the thing that comes next after this is sort of multi-agent systems, because it's kind of related to this like systems approach, right? At some point, when you have a group like that, it becomes very hard to maintain that kind of freedom and then there became a bit more like politics and things like that. So it's just very hard, I think, in the long run to preserve this kind of dynamic environment. And I think they did a great job preserving it as long as possible. And they did some really amazing things, right? So, okay, like RAG, it's like, that's cool and cute and everything, but like PyTorch, like that really changed the world, right? And I think.

• 41:08Speaker 01

  Probably fair doesn't deserve enough credit for for what happened there. And I think Zuck doesn't get enough credit for being willing to take that kind of bet, right? Like, why give away PyTorch? Like, why give away React? Why give away Llama? No. I think it's really visionary leadership that he's doing there.

• 41:26Jacob Effron?

  Why give away Llama? Like, what do you what do you think's behind that?

• 41:29Speaker 01

  Well, I I think, so he was recently on this other podcast, Acquired.

• 41:36Jacob Effron?

  We'll let you mention it. Acquired, we'll we'll we'll make an exception for you. You can mention that.

• 41:40Speaker 01

  Yeah, so I mean, they filled out.

• 41:41Jacob Effron?

  Did you go to the live recording?

• 41:42Speaker 01

  Yeah, they filled out the whole Chase Center.

• 41:44Jacob Effron?

  Yeah.

• 41:44Speaker 01

  So that's something for you to aspire to.

• 41:47Jacob Effron?

  Yeah, I don't I don't think we can do the Chase Center yet. You know, we'd have to do some really small, local high school gym or something.

• 41:53Speaker 01

  That would be fun, maybe.

• 41:54Jacob Effron?

  Yeah, exactly.

• 41:56Speaker 01

  Yeah, so I think he had some very interesting observations around like platform, right? So if the new platform is language models or generative AI,

• 42:06Speaker 01

  it is very important for him to not have to serve that up on somebody else's platform.

• 42:13Speaker 01

  So I think that's a very pragmatic reason. I also think it does them like great benefit. Like, I think like the whole public perception of of Facebook. When I worked there, everybody hated Facebook and I had to like field a lot of like questions around like this is like this evil empire and stuff. And all of a sudden, like, everybody thinks Zuck is cool, right? He has become.

• 42:36Jacob Effron?

  The wardrobe changed.

• 42:37Speaker 01

  Yeah, he has become a lot cooler after they given that. But so I I think people are just seeing the contributions that happen here. And and I think it's helping them a lot with hiring. It's really putting them as like one of the top players in the field. Yeah, they're so they're one of the the cool kids again.

• 42:53Jacob Effron?

  Totally.

• 42:54Speaker 01

  Finally.

• 42:55Jacob Effron?

  You mentioned while you were there, you do this work on multi-agent systems. And I feel like that that feels like it's coming, right? I mean, I feel like we're seeing all these kind of architectures of, you know, you have agents that are particularly good at things. And then it's how do you coordinate among those agents and have them work together? And, you know, you as a person have your like fleet of 30 agents that you're sending off and doing different things like. How close are we to that? Like, how do you think about, you know, that feels like another space where demos are going to be really cool and then, you know, actually productionalizing that will be really hard.

• 43:20Speaker 01

  Yeah. Yeah. So in a way, I think we already have that, right? So there are agents and you can connect agents to each other. And I think when you think about it from a kind of machine learning perspective, like synthetic data is also about agents, right? It's like there's this one agent training this other agent. So I think a lot of this is already just happening. But yeah, I think for me, the broader observation is that we should really not be thinking about this as like monolithic models. We should be thinking about it as systems and those systems can comprise many models and those models together need to solve the problem. So that is a multi-agent setup.

• 43:58Jacob Effron?

  Totally.

• 43:59Speaker 01

  So yeah, we're sort of already there. I think we're just not calling it that yet. So somebody's PR department at somebody, some company is going to at some point make a big splash around that.

• 44:10Jacob Effron?

  Yeah. But I'm sure there'll be all sorts of issues in actually figuring out how you get folks to collaborate and route the correct way.

• 44:16Speaker 01

  It's a very different optimization problem, right? So I thought like these emergent communication protocols that emerge when you get these agents to work together, like that they might display very interesting like linguistic behaviors that humans also happen to have and that that's actually a reflection of like the organization of the agent structure. So we have some very cool papers where we look at the topology of the agents. So like which agent is connected to which agent? Like, do they all talk to each other? Or are they like only connected to the next one sequentially, right? So very different dialects kind of emerge.

• 44:53Jacob Effron?

  A lot of like, I think everyone's always debating org structure within companies and what the.

• 44:56Speaker 01

  Yeah, like culture. Exactly. Yeah. So like that is getting at culture. And I think like in the even longer run. Like, humans will be agents in those kinds of like multi-agent systems as well.

• 45:10Jacob Effron?

  And then the other agents will decide, we definitely don't need to call Jacob for this.

• 45:14Speaker 01

  Yeah.

• 45:14Jacob Effron?

  Well, maybe if this other thing will come.

• 45:16Speaker 01

  Yeah. Hopefully, you will still be the CEO controlling all of these agents for you.

• 45:21Jacob Effron?

  Yeah.

• 45:24Jacob Effron?

  You know, I guess, obviously, another place you worked was Hugging Face. And curious, like, obviously, Hugging Face does a lot of really cool research in a bunch of different places. You know, what do you think the kind of future role of Hugging Face is in this broader AI ecosystem?

• 45:35Speaker 01

  Yeah, there's just an amazing company, right? It's like, joining Hugging Face for me was a very different experience from Facebook, because everybody hated Facebook, and it was very different to work at a company.

• 45:46Jacob Effron?

  Emoji logo definitely helps you.

• 45:47Speaker 01

  Everybody loves the company. So, yeah, I think Hugging Face really has become kind of the central place for a lot of publishing of things, right? And I think it's maybe a missed opportunity for some of the other kind of more traditional, like, so GitHub, where you publish your code. Like, why didn't they do the model publishing earlier, right? Or, we were actually looking at this with PyTorch as well. Like, there was a model hub and everything, and it didn't really take off, because it wasn't, like, properly resourced, I think. But out of that, they built a very special place for themselves, I think. But, so, yeah, in the long run, I think, so, it's a place where you publish models.

• 46:31Speaker 01

  So, they do exist by the grace of Zuck.

• 46:35Jacob Effron?

  Yeah.

• 46:35Speaker 01

  In a way, and others, right? Like, so, they're basically just kind of making money off of other people's open-source efforts.

• 46:45Jacob Effron?

  Right. Where do you think they go from here? I mean, is that always kind of what they'll be? Or are there other interesting directions that you can build? You know, obviously, it's a place to potentially deploy these models, or?

• 46:53Speaker 01

  Yeah, so I think they're doing lots of cool experiments around that, basically, like, trying to figure out what, where to take that, yeah.

• 46:59Jacob Effron?

  We have a lot of founders on the show and as listeners, too. And so, I think one thing that's interesting, you obviously have, you know, impeccable research credentials. You're building a really interesting company, you've got some great enterprise partners. I imagine people just want to throw money at you all the time. And I thought this is true of a lot of AI startups. And so, I'm curious how you thought about, over the series of, you know, the course of running Contextual, like, how much money you should raise? You know, how do you kind of, obviously, there's other people that raise insane amounts, like, you guys have raised a good amount of money. Like, how have you kind of thought about that throughout the journey?

• 47:28Speaker 01

  Yeah, so we're trying to be very pragmatic around these things. So I see a lot of companies probably raise way more money than they need at valuations that they will have a very hard time living up to. So I think it's important to balance that out, especially because it is likely that the AI hype cycle is going to die off a little bit in the next year or two. And so we're trying to plan for coming out of that as one of the big winners. But I think a lot of companies are going to really struggle when that happens. So yeah, in terms of the money we've raised, I think we've made some like very clear decisions in that we are not going to train our own big base models because we know that Zuck is so great.

• 48:14Speaker 01

  So that means that we can be much more capital efficient or that we can allocate our capital to other things that are necessary, like having amazing people and also having these people work very closely with customers to really solve the problems. So I think there's a lot of confusion in the market around the maturity of AI products, where they think like, why can't I just have like a turnkey thing and I just like throw my data at it and everything works. It's like, we are not there yet. So that means that for me as a CEO, we need to think much more about how we can deliver amazing experiences to our customers with like something that in the long run will become a turnkey thing that everybody can just use out of the box, but in the short

• 48:57Speaker 01

  term needs like, you know, white glove, a bit of a more dedicated support. So what I always like saying is, we sell like the Tesla Roadster right now. It's like high-end sports car. It's kind of hard to drive. You need to like tune in properly, right? And so we give you some some mechanics and engineers with the car to make sure that it really drives the way you want it to drive. And while while we're selling these Roadsters, we're also working on the assembly line for the Model S.

• 49:24Jacob Effron?

  Totally.

• 49:25Speaker 01

  That is going to come at some point.

• 49:27Jacob Effron?

  Do you think we're on this like exponential of model improvement where the Model S is coming like literally in six months? Or like, are we on some sort of S curve where like, okay, you know, that we're gonna, there will be some sort of plateau for a bit and then, you know, obviously, I'm sure you're a bull long-term on this, but like, you know, some people are like, look, in a year, you know, these models, it's just gonna be night and day. You'll look, you'll look back at GPT-4 quality models and laugh. And other people are like, well, I don't know, actually, it seems like there needs to be some sort of algorithmic breakthrough.

• 49:50Speaker 01

  Yeah, so I think we're really only just getting started with a lot of this stuff. But I do want to point out that the things coming off the assembly line, the cars are systems, not models, right?

• 50:00Jacob Effron?

  Yeah.

• 50:01Speaker 01

  And so I think like a lot of people still talk about scaling laws and things. But at some point, you don't really have to scale anymore because there's so much other stuff you can do to make it better, right? So you just scale up to a certain point and then you do a lot of like post-training around that pre-trained point and then you do things like, you know, chain of thought and like distilling that back into the model to like taking the systems thinking and putting it back into the model. So there's just a lot of, lot to be gained from doing that too. So we're kind of scaling in many directions, right? So yeah, I think there's, there's a lot of exciting stuff still ahead of us.

• 50:39Jacob Effron?

  Yeah, well, we always like to end our interviews with a quick fire round where we get your take on, on a few common questions. And so maybe to start, we'd love one thing that's overhyped and one thing that's underhyped in the AI world today.

• 50:50Speaker 01

  Yeah, so overhyped and underhyped, I think is agents, actually. So they're on, in a way, they're really overhyped. It doesn't really work yet, but they're underhyped because like it is showing signs of life, I think. So yeah, that's probably the one that springs to mind at first, yeah.

• 51:09Jacob Effron?

  Nice. And then what's been the biggest surprise in building contextual? I guess, you know, be an example of one thing you thought would work and it didn't, or one thing you were like, I don't know, it was a random experiment and it actually worked really well.

• 51:18Speaker 01

  Yeah, yeah, so, so I think like when, when you are not in a startup, you think that a lot of the kind of infrastructure things are, are easy, right? Like, let's just have a cluster that works. Like that, we're, that's, we kind of assumed, I think, that that was a given. It turns out that having like a high-end research cluster that actually works is incredibly hard. And even if you look at like the latest Lama paper, they have some stats about how often they had hardware failures. It's, it's amazing how often they had to like swap out GPUs or even entire nodes, right? So the, the infrastructure that we're using to build a lot of this technology is a lot more, more fickle than you would think.

• 52:00Jacob Effron?

  I know this isn't a one or the other, but obviously people love to debate closed source versus open source models. And so I'm kind of curious, you know, which one you think will be the, the, you know, the dominant usage for most business use cases long-term? Is it state-of-the-art closed models or actually much more specific, you know, open source models?

• 52:16Speaker 01

  I think a mix, right? But I guess I'm a little biased here. But so I used to always talk about this as sort of a triangle where at the top of the triangle you have these high-end closed-source models that are kind of frontier, right? They're also more expensive. Then at the bottom of the triangle, you have these open-source models. Everybody can access them and like do things with them. The most interesting part is sort of the middle of the triangle where you get the right trade-offs between sort of like capital efficiency for the company, for like usefulness for the customer, like serving it up at the right price point, right? So I think if you start from open source and then you have an amazing post-training capability, then you can just do a lot on top of that to end up in the sweet spot, sort

• 53:01Speaker 01

  of in the middle of that triangle where you have the right trade-offs.

• 53:03Jacob Effron?

  Which AI startup outside of your space are you most excited about?

• 53:07Speaker 01

  So I've been really impressed by where AI-based entertainment has been going. So sort of like Suno and things like that, right? And like all of the video generation companies and there's just, that is going way quicker than I than I would have thought.

• 53:25Jacob Effron?

  We had Mikey on the podcast and I feel like that's that's the one, I'm always trying to like, you know, show my family like this cool demo, that cool thing. And Suno's the one that's like, that's stuck.

• 53:34Speaker 01

  Yeah.

• 53:35Jacob Effron?

  It's a really compelling. And the video stuff is also really cool. It obviously just feels very early.

• 53:39Speaker 01

  Yeah. So it feels very early, but so it's still going quicker than I would have thought. And so you can just extrapolate where that's going to go, right? Like in 10 years, we are not going to watch one movie anymore, right? There's like, everybody will watch their own personalized movie.

• 53:56Jacob Effron?

  Right.

• 53:57Speaker 01

  Or you can watch like infinite episodes of whatever your favorite show is, or you can make up a new ending to Game of Thrones that is actually good, right?

• 54:06Jacob Effron?

  If all of humanity focuses on that, hopefully someone comes up with a better one.

• 54:10Speaker 01

  Well, so everybody will get their own personalized ending.

• 54:13Jacob Effron?

  Right, right. Yeah, I guess it would be different for.

• 54:15Speaker 01

  That is optimal for them, right?

• 54:16Jacob Effron?

  Yeah.

• 54:17Speaker 01

  So there's just a lot of like personalization.

• 54:18Jacob Effron?

  I'm not sure anyone will stick with the existing one, but there's got to be someone out there that was optimal.

• 54:21Speaker 01

  Well, maybe, yeah, the directors, I guess, of the show, right? They thought it was great.

• 54:26Jacob Effron?

  What company besides Contextual do you think would be like most interesting to run AI at? If you could just plop you in anywhere?

• 54:31Speaker 01

  Yeah, like there are obvious answers like sort of Google or Meta or something like that. But I think because of what I'm currently working on and what I'm most interested in, I think actually a lot of the cool stuff in AI is about how to turn it into real usable products for people, right? So if you think of like, I don't know, JP Morgan or something like that, like think of all the amazing problems that their head of AI gets to solve using all of this amazing technology.

• 55:00Jacob Effron?

  Well, I think this is some episode in the 40s and this is the first time someone's answered JP Morgan. That's a good, I don't know what the next 60 years will end that either. I love it.

• 55:08Speaker 01

  Yeah, no, but so like all these like more traditional enterprises where suddenly you you encounter this like new technology that's going to change everything. And if only you can figure out how to do that before everybody else, you have this massive opportunity. I think that's very exciting.

• 55:23Jacob Effron?

  If you, you know, weren't running Contextual and it was, you were building an AI application, like what application would you build?

• 55:30Speaker 01

  Probably also something in the entertainment space. I think that that could be very interesting. Yeah. Yeah, it's kind of hard to choose, honestly, like so for us, I mean, I think like in most alternative futures, I would still end up working on Contextual because I think like work is the most obvious place where this technology is going to change everything, right? So and like our company mission is to change the way the world works like through AI. And yeah, if we can really make that better for everybody, I think that is a very noble mission and will have huge impact. So that's the thing I would focus on in most alternative futures.

• 56:09Jacob Effron?

  Well, the nice thing about this personalized entertainment is you'll be able to make a movie of yourself having done, you know, 100 different companies.

• 56:14Speaker 01

  Yeah, exactly. So you still have your alternative futures that way.

• 56:19Jacob Effron?

  Well, this has been a fascinating conversation. I'm sure folks are going to want to pull on lots of threads of things we talked about. I'll leave the last word to you. Where's the best place for people to go and learn more about your research? Contextual, anywhere you want to point them.

• 56:29Speaker 01

  Yeah, contextual.ai. Follow me on Twitter. My name is hard to spell, but you can find it. There's only one person in the world with my name.

• 56:39Jacob Effron?

  That was impressive.

• 56:42Speaker 01

  And yeah, so yeah, parting thoughts. I think we're just really at this amazing point in history where we get to see the world changing in front of our eyes, changing how the world works. So yeah, exciting times ahead for everyone, I think.

• 56:59Jacob Effron?

  Yeah, well, thanks so much for coming on. This was a fascinating conversation.

• 57:02Speaker 01

  Thank you.