Ep 34: Eric Ries and Jeremy Howard (Answer.ai) on the Biggest Mistakes AI Founders are Making and Building the Bell Labs of AI

In this week’s episode of Unsupervised Learning, we delve into the forefront of AI innovation with Eric Ries and Jeremy Howard. Eric Ries, renowned for pioneering the Lean Startup movement, has consistently influenced modern entrepreneurial strategies with his emphasis on agile, sustainable growth through innovation. Jeremy Howard is known for his contributions to deep learning and data science, co-founding the fast.ai educational initiative that democratizes access to cutting-edge AI learning.

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

• 0:00Jacob Effron?

  Eric Reese and Jeremy Howard are building the Bell Labs of AI in Answer AI, a company at the forefront of building smaller, cheaper, more affordable models, as well as applications in legal and education. We had a really interesting conversation on supervised learning that hit on what applying the lean startup process to AI looks like, when AI companies should think about moats, where the opportunities are in legal and education, AI safety, and more. I think folks are really going to enjoy this one and stay tuned for Jordan and my debrief at the end.

• 0:27Jacob Effron?

  Well, Eric and Jeremy, incredibly excited to have you both on the podcast. We've been looking forward to this one for a while.

• 0:31Speaker 01

  Thanks for having us.

• 0:32Jacob Effron?

  Eric, I feel like it's one of the great joys of this job to get to talk to the author of a book that I feel like has defined a lot of how many of us in tech think about startups and building, kind of innovating within larger enterprises. One thing I'm taking by is I feel like a lot of the lessons that you kind of put forth in the lean startup, it feels like we're seeing some different things sometimes in the AI world. We're seeing some big rounds, a lot being spent on models and compute long before anything touches market. Curious just your high level take on that and whether there should be different lessons here or whether the same thing applies in this industry.

• 1:07Speaker 01

  Well, thanks. Thanks for the kind words. And yes, it is a bit of a Back to the Future feeling. And it's true in the startups, you're seeing it across the entire industry. You're seeing it in corporates who are trying to use AI. You see it in people trying to build products out of foundation models. And I think partly it's an artifact of the fact that AI makes just such unbelievably good demos. It's just they're so magical. And so it's really easy to convince yourself that you're the exception and you don't need to test it with customers because the thing is self-evidently cool. So you're just looking at it, look what it can do. And yet I think the same principles, at least at a deep level, the same principles are always true. That is, you can't know in advance what customers are going to want.

• 1:53Speaker 01

  And people will hear that, they're like, well, but customers don't know what they want either. And it's like, that's true. We're not saying that you're going to, it's like saying, someone is like, I want to do chemistry. You say, oh, I can't do chemistry because I can't ask the electrons what they want. It's like, no, we can experiment. We can go discover what people want through their revealed actions. And I think the other thing that's making it especially difficult is People have taken the SaaS stack and just copy pasted it to AI. I mean, they basically just crossed out the labels with crayon and wrote in new labels, assuming that everything's going to be the same.

• 2:25Jacob Effron?

  Right, that's like dot com to dot AI, right?

• 2:27Speaker 01

  I mean, it couldn't be more transparent. Look, a lot of people raising a lot of money doing that. So far be it for me ever to criticize something that helps entrepreneurs raise money. That's always good. But I think one of the problems with that is you have a lot of companies that have pushed the product market fit question to a different layer of the supposed stack. You think about all the AI companies that just make an API. They're seeing, some of them, tremendous usage of that API. And they're like, just assuming that their customers are going to be able to find product market fit with their customers. And if you follow that value chain, in many cases, it's not even two layers deep. It could be two, three, four layers deep between the model and the end product. And the issue there is that if anything is different about this stack, if it's not assembled

• 3:08Speaker 01

  the same way, there's going to be a lot of carnage in applications that are set up that way. And I've always been an advocate for wherever you are in the supposed stack, to really understand who the end, end, end customer is and what they want, to make sure that you're part of creating something that can achieve product market fit. So I think that's not any different. But a lot of the tactics are going to look different. I mean, you know, software development and model development are similar, but not the same. The economics of AI are completely different than what we saw in software. And frankly, I'm drawing a lot more on the lessons I learned doing physical manufacturing and, you know, deep sea oil drilling and nuclear combined cycle power plants and nuclear plants, like things that had real infrastructure and operating cost and also market risk.

• 3:52Speaker 01

  You know, it's actually, you're amplifying the risk, not reducing it by bringing in all this computational cost. So I think it's really, it's a really interesting time to be applying those principles. And I think it is important to be open minded about what has to change and what is the same. But I think the core, core, core idea that, you know, as Peter Drucker said, that a business is an entity that exists to create a customer. That's no different now than it was then. And the agents, you know, the AI agents are not the customer. Human beings, you know, we say, you say, metrics are people too. That's all they are. It's just human beings doing something. So if you're disconnected from that, it's very easy to go astray.

• 4:27Jacob Effron?

  Yeah, no, I mean, it makes it makes a lot of sense. And I feel like on the flip side, there's been, you know, and it's going to be at Arvin from perplexity on the podcast. And I think one thing he talked about is a lot of people feel like they're paralyzed into action because of this idea of a priori, like, do we have notes? And, you know, is there defensibility here? And a lot of people are maybe even looking at things where they're like, actually, this could provide value to an end customer. But like, is this defensible? And, you know, I think different opinions on like, should you jump in from the start if you're not sure on that, you know? Arvin's like, look, you have to build something that customers want. And then you have to earn the right to, to, you know, actually think about like, do you have a mode or not? Curious like your, your take on that? Because I think a lot of people are, there seem to be a lot of things customers want. And then a lot of people intellectually thinking about if they have a business model around it.

• 5:12Speaker 01

  It reminds me of the old parable of picking up dimes in front of a steamroller. So you have these like mega, mega platforms with tremendous power. And they're so big, just by definition, how expensive they are, how many things? Think about all the different things that OpenAI could be building. So although in theory, anything you do, they could just nuke you from orbit by deciding to do it too. In practice, you know, they, they're, they're pretty fast for a giant behemoth. But there's only so many things they can do at once. They have to be focused. And so the vast majority of things, they're not going to get around to doing, you know, right now. And so the, the old lesson is that you can, you know, if you're fast enough, you can jump into the street, grab the dime, jump out of the street. And if you're really smart, you can do that over and over again, picking up these use cases one after the other.

• 5:53Speaker 01

  And the only downside to that strategy is that if you trip and fall even one time, you know, then you get totally flattened. And that's like, you know, in tech, we used to talk about that all the time, back when we had bona fide platform wars, you know, Mac versus PC, Intel versus, you know, you know, now we're into, you know, x86 versus ARM. But, you know, if you go back to Motorola versus Intel and just these been, we've been through a lot of these over the years, different compiler platforms. You know, obviously, social networks was quite violent. Android iOS wasn't actually that much of a hardcore platform war because Google never really had their heart in it. But anyway, we've been through this before, but I feel like it's been a little bit, a little while. And so people have kind of forgotten that a huge part of building startups, you know, in the tech industry, used to be having some sense of industry analysis

• 6:38Speaker 01

  and understanding that the strategic options and stuff like that. But all that being said, It's so much easier to talk yourself out of doing anything than it is to, you know, like, put yourself in a position where you're able to pivot. I'm kind of with Arvind on this one, where, you know, yes, it's not that, it's not that analysis is not useful, of course, we want people to do analysis. And, and look, there are plenty of startups where you're like, man, even if it works out, there's no economics here, even if it works out, there's no impact. And I mean, I even meet startups where I'm like, even if that works out, it's pure exploitation. And so you might make money from it, but you're not going to sleep well at night. So there's plenty of things you can rule out. But if you have something that you passionately believe in, if you're not sure how to analyze it, or, you know, even if someone who sounds really smart tells you that you're doomed,

• 7:22Speaker 01

  I don't feel like that's a good reason not to do it. But I do think the humility does have very concrete company building lessons you have to take. In particular, you have to build in such a way that you can do rapid iteration and you are able to pivot. You have to keep your, to be alert to the possibility that your assumptions could be wrong. And that's, you know, people, I think, sometimes think, okay, because I've given up on analysis, therefore I don't have to iterate. It's actually the opposite. Now, you, you need it even more because you're not completely sure of what's going to happen. And with the level of uncertainty that we're facing with AI, I mean, anyone who thinks they know even just what are the capabilities of models going to be, you know, in a few years, or what are consumers going to want? How is it going to affect society? People are making very confident pronouncements left and right about this.

• 8:06Speaker 01

  But I don't see how anyone can attach like real, serious confidence that they're correct in those things. You're much better off preparing yourself for a range of possible futures. And that, that definitely, I think, is a strong argument for having a strong vision, but being prepared to pivot.

• 8:20Jacob Effron?

  Yeah, I mean, I think just given the pace at which things change, it feels like the rapid iteration and, and kind of ability to quickly adapt and get feedback is like, is even, has never been more important, right? Just given how, how quickly things move.

• 8:31Speaker 01

  I mean, what do people expect me to say? But yeah, obviously, I think so. But, but I mean, I really do, I really do think so. And I feel like the, we're, I mean, we're already starting to see the first wave of companies who didn't do that run into serious trouble. So we kind of, I feel like we're speedrunning the unusual industry hype cycle. Where usually it takes a couple of years before people start writing the article about how you don't need lean startup anymore. And then the thing, you know, then they write it about Quibi or something. And it's like, it's always the companies they write it about that are then later have a monumental crash. We've skipped that step because we're going right, going right to the crash. That's progress of a sort, I think.

• 9:04Speaker 02

  So we'll definitely come back to the lean startup in a second here, but I wanted to turn our attention to you, Jeremy. Obviously you started Fast AI with the intent of teaching deep learning to folks. And I wanted to get your thoughts on if you are an L M curious person today, what do you think is the best way to get caught up?

• 9:20Speaker 03

  I mean, you know, fast.ai and answer.ai are not so different. You know, the goal of both of them is the same. So answer.ai is a, is a public benefit corporation. It's stated public benefit is to maximize the public benefit of AI for as many people as possible, which was exactly what fast.ai was created to do as well, you know? I think the difference is that language in particular is such a powerful modality. And we've now started to unlock that modality with computers that it deserves a particular focus.

• 10:05Speaker 03

  You know, also it's when, when my wife Rachel and I started fast today, I was so speculative. You know, this was, you know, we were the first broad online deep learning course for on PhDs. The idea that people who don't have a PhD could learn and use this tool was generally considered pretty ludicrous and pretty naive. And the idea that you could write software that made it, you know, so that you could create applications in a few lines of code was also considered naive and ludicrous. You know, now that we've gone quite a few years down that track and it turns out that these things actually worked out pretty well, I think apart from anything

• 10:49Speaker 03

  else, it's just time to invest a bit more into this mission. So, you know, now, now we've got, you know, VC and angel money behind us, we've got a team of people behind us. And of course, we've got Eric working on this mission with us as well. And so, yeah, it feels like we're really kind of setting ourselves up for for success here in terms of this, in terms of this mission of creating widely applicable, societally beneficial applications of artificial intelligence.

• 11:27Speaker 02

  So that's a great segue into Answer AI. You know, obviously you two together have just incredibly credentialed backgrounds and could really work on anything you want to. And so why did you decide to really go after this? And how did the idea come about in the first place?

• 11:42Speaker 03

  Well, Eric and I had been talking about it for years, really. You know, whenever we chatted, we were always just chatting about AI and about like, how many opportunities it created. And I was kind of asking for his advice. You know, one of the pieces of advice I asked Eric about was like, fast.ai was hamstrung by the fact that everything that we created, all of our research, all of our courses, all of our software, required a strong coding background to take advantage of any of those things. And for something that's attempting to make AI as accessible as possible, you're restricting yourself

• 12:21Speaker 03

  to less than 1% of the world's population by doing that. And yeah, it seemed like maybe there's a huge opportunity to actually.

• 12:34Speaker 03

  Make it way more accessible, you know, by by taking advantage of of natural language and other natural modalities, vision and so forth. So, you know, one piece is kind of saying, well, how do we take advantage of this? And another, you know, that was the opportunity. And then there was this kind of threat of like, you know, also, it seems like the thing that fast.ai was created to counter, which was Rachel and my concern, that AI will lead to massive amounts of centralization of power. And as a result, decrease opportunities for for most people. That centralization seemed to be accelerating, you know, like

• 13:13Speaker 03

  in that sense, I felt like fast AI's mission was was failing. So that's really where it started, I guess. At least, Derek, that's my memory of it was like, how do we take, you know, try to capture this opportunity whilst trying to do this.

• 13:26Jacob Effron?

  Do you agree with any of that? Or do you have a completely different, different memory of where it all began?

• 13:30Speaker 01

  Yeah, what was talked about, we were, we were on Barbados doing a, I think, yeah, I mean, that's, that's definitely right. I mean, I think to call fast AI a failure is frankly ridiculous. So I don't agree with that. But but I think, and I think Answer AI is made possible by the tremendous goodwill that, you know, Jeremy and Rachel have built up over so many years, helping so many people enter the field with not just educational resources, but really a real sense of values and ethics about how this technology could and should be developed.

• 14:05Speaker 03

  One of the things we both struggled with was thinking about was like, well, this is, you know, unlike fast AI, which was entirely funded by Rachel and my personal bank account, you know, this was something we felt was worth investing more money into. And we didn't know who the hell would fund such a thing, because it's not a normal startup. Answer.ai is a for-profit R and D lab. So questions like, what exactly is the proprietary part of your technology? What exactly is your product market fit? You know, who exactly is your market? What's your five-year financial plan?

• 14:46Speaker 03

  We don't have any of those things. Like, the nature of an R and D lab is that you do the R and then you get the D.

• 14:51Jacob Effron?

  Yeah.

• 14:51Speaker 03

  And, you know.

• 14:52Jacob Effron?

  I mean, if you told me that the, uh, the author of the lean startup was starting an R and D lab, I would have thought that those two were like almost antithetical to each other in some way.

• 15:01Speaker 01

  No, I, listen, that's a very common, very common misconception. And I've actually worked in, in like old school industrial R and D labs. Like, you know, we use that as a metaphor here. For I'm talking about people that are doing actual research on heavy industry. And part of the problem, both in academia and in industry, is we have, uh, over the course of the past century, had this tremendous drift. It's natural, I think, as a result of specialization, but it's also a bit of indictment of the way our capital markets work and time horizons. We have had this hyper-specialization where the R and the D have become so separated. We actually now do really large amounts of research, completely untethered from practical applications, in ways that Thomas Edison would find bizarre. And so some people look at that and they say, okay, um, the right solution to that is to have a specialized,

• 15:46Speaker 01

  you know, research function and a specialized development function. And you run lean startup in the development function and you, you know, you tightly integrate them together and you have two. But, but Jerry and I both happen to have the idiosyncratic belief that the very best research is done only when the researcher, it's him or herself, is coupled to the application. And so it's not to say that this is the only kind of research that can or should be done. Of course, there, there are many, many different approaches, but I think as a society, we're way under-invested in this style where the iteration loop, um, you know, goes from the customer all the way through into the scientific inquiry and back out again. And I could tell you funny story after funny story of times when I've done this, doing, you know, lean startup transformations in a corporate setting, in an industrial setting,

• 16:31Speaker 01

  where the thing that the researchers are like practically trying to win a Nobel Prize solving is not of value to the customer. So it's like first we have to win a Nobel Prize so that we can then, you know, make a material science breakthrough so that, you know, I remember a very funny example of building data centers, technology for data centers. And, you know, they were really focused on the energy efficiency of the thing and, you know, how, what its operating costs would look like. They had all these like extremely difficult performance optimizations that the scientists were working to figure out how to do. And I'm not being facetious about Nobel Prize level. It was required like fundamental physics breakthroughs to achieve.

• 17:12Speaker 01

  And the first time we made them take an MVP of the thing to the data center builder, we learned the very basic facts that first of all, the person building the data center doesn't operate it. So they definitely don't care. And they, the first question they asked was a question that no one was asking in the lab, which was, well, what's the footprint? How big is the thing? You know, like, oh, to make it energy efficient, we only have to increase its size by 300%. So that should be a really good trade. But, you know, if you think about it from the point of view of a data, where data, you know, data center builder, they want to know how many racks they can squeeze in. And you're going to take how many racks away from me in order to make, save money by someone else? Why don't, I mean, it was just like such a classic thing where they're working so hard on the thing that doesn't matter to the customer. And when we bring that into alignment, once we brought that fact back into the lab, you

• 17:56Speaker 01

  know, it took one of the junior PhD, you know, level researcher like five seconds to be like, oh, well, if you're, if you're willing for a different set of trade offs, then the scientific problem is really different. Let's go have this other breakthrough. And then, and then we're off to the races. And so, yeah, that's kind of a trivial crude example. We're just doing it one time. But if you do that in a continuous loop, then tremendous breakthroughs are possible.

• 18:18Speaker 03

  So coming back to the source of this part of the discussion, you know, given that this kind of R and D, for profit R and D lab hasn't appeared in the world since the early days of electricity. We weren't sure who the hell would fund it. And we had a lot of good luck because actually.

• 18:39Jacob Effron?

  I think it was from one of your podcast appearances.

• 18:41Speaker 03

  The last time I did a podcast interview.

• 18:43Jacob Effron?

  Exactly. I feel like we're gonna have to give you a term sheet after this.

• 18:45Speaker 03

  And I, and I reached out to one of the interviewers, Alessio afterwards and said, like, could I get you some advice? Cause it seems like you guys talk to everybody. And yeah, you know, and I said to him, like, we're thinking about doing this like R and D lab thing. But I don't know, like, what kind of investors, you know, would give money to such a thing? He said, oh, we can give you a term sheet tomorrow. So, you know, there was a lot of like, I know how much luck there was now because, you know, since that time,

• 19:17Speaker 03

  we, we also decided to bring some angels in. And literally, I think every, like AI insider I reached out to, maybe you're but one, who I, who I invited, would you like to invest in this? Said, yes. So, you know, amongst that group, everybody wanted in. But then we had a few people reach out and be like, oh, Jeremy and Eric are doing something like, you know, traditional investors.

• 19:47Jacob Effron?

  Yeah.

• 19:48Speaker 03

  We, we definitely want to end like, please talk to us. And then when we talk to them, they're like, so what is your proprietary technology exactly? Okay. And like, who is your market? And, and whereabouts in the SaaS stack are you? And it's just like, wow.

• 20:01Jacob Effron?

  All right. I'll erase all those. We had all those questions teed up for this episode, but I'll just, I'll just get rid of them. So we're not mercilessly mocked.

• 20:08Speaker 01

  The, the key insight, I think that maybe some of your listeners would benefit from is,

• 20:14Speaker 01

  the idea of a stack to begin with, the idea that there's abstraction layers. And this goes back to like, you know, right? The like, the layers of the IP stack and, and the way that the internet technologies will work, you know, and, and the stack that people developed around, uh, semiconductors and how that relates to the development of software like that. Those are very useful abstraction boundaries. When you understand enough about the thing in question to know how the layers interact. And even then, the reason we prize full stack engineers so much. Is there are these times when you have to break the abstraction boundary in order to get the thing that you want because you're trying to achieve something that only by integrating across layers or doing something unusual can you can you achieve?

• 20:54Speaker 01

  When you have a nascent field like this, the reason why, you know, as you said, the development helps the research, but the reverse version is also very valuable because I think prematurely deciding what the abstraction boundaries are just leaves a lot of opportunities out of your consideration set. And we've already, we've only been at this a few months. We've already had numerous situations where someone who really deeply understands the technology will be able to say, you know, your assumption about XYZ is just not true.

• 21:22Speaker 03

  What's an example of that that's happened for you guys? I think a great example is, you know, lots of discussions about agents. So there are places that agents are well-connected with the capabilities of language models, and there are places where they aren't. Most of the places, they're not. But you only understand those places if you have a deep understanding of the underlying theory of the so-called linearized subgraph matching operations on which LLMs work. When you deeply understand those mathematical foundations, you realize that any kind of

• 22:04Speaker 03

  agent which requires a kind of novel planning, and I don't mean novel in some kind of like Nobel Prize level novel, but just like a planning sequence which isn't almost exactly in the training data, it won't be able to do it. Anything which is largely various mixes and matches of stuff that you've seen in the training data, it will be amazingly good at. So, you know, agents which kind of help you answer your emails, you know, or, you know, or label them or whatever, or add some CSS to your HTML or, you know, iteratively debug

• 22:47Speaker 03

  common Python problems are going to be great.

• 22:52Speaker 03

  You know, or as agents that attempt to You know, independently create a research breakthrough. Or, or the thing I come across all the time is I, I, because I'm a researcher, a lot of my code is creating algorithms. They're not rocket science, like often they're just four or five lines of code, you know? And I've never, ever had a language model help me at all in creating any of those algorithms, because by definition, they're not in the training data. If they were, like, I'm, I'm pretty good at algorithms, so I wouldn't be asking the language model for help that everybody already knows about. So like, this is an area where I think like, understanding,

• 23:34Speaker 03

  you know, the low level research really changes, you know, the, the type of development that we try to do.

• 23:42Jacob Effron?

  So interesting. I mean, obviously you guys started, and I think you, you kind of recently released some stuff around, you know, an efficient fine tuning of Lama three. I'm curious, like, how did that come about? You know, can you talk a bit about what you found and kind of the implications of that for, for builders?

• 23:55Speaker 03

  Yeah, definitely. So we like, this is such a great example of, of so many of our values and premises here. So i- in short,

• 24:05Speaker 03

  the Lama three fine tuning breakthrough was done by one guy. His name's Karam. He lives in Turkey. And he was like the best master's student I ever had.

• 24:20Speaker 03

  And, you know, and then like, okay,

• 24:24Speaker 03

  how, how did the management occur or whatever, which we decided to resource that project? It didn't. We don't have managers. I'm actually not a manager in our company, and I don't do any management, when nobody does.

• 24:36Speaker 03

  He just decided to. You know, he decided it would be a good idea. And he started it and he started doing it. And along the way, some of the team were like, this is really cool. Let's find out how we can help Karam. And they helped, you know, here and there.

• 24:53Speaker 03

  But then, why were we working in this space in the first place? Well, we were working in this space in the first place because we operate as an R and D lab, again, in the same way that Edison's lab did. It's what, uh, there's a great historian of these labs. His name's Eric Gilliam and Eric did an analysis, actually, of answer.ai.

• 25:13Jacob Effron?

  Yeah, we'll link to that piece, a really, really good piece.

• 25:16Speaker 03

  Yeah, it's a great piece. And he describes this as a long leash with narrow fences. So Eric and I created narrow fences, you know? So from early on, you know, we tried to figure out like, you know, what's, what's the research thesis here? What's the idea? And, you know, we thought, well, there's a real overinvestment in training foundation models from scratch. There's a real overinvestment on doing things on like gold plated, super expensive hardware. There's a real underinvestment on like the real world, which is resource constrained. You know, and so we, we brought on people that care about efficiency, that care about, you know, doing things at a reasonable cost, you know?

• 26:00Speaker 03

  And they're actually often folks who therefore come from backgrounds where they had resource constraints themselves. So they deeply empathize with these things. Yeah, we can also, we've, we've all been working as a team and trying to think like, how do we bring the price down? And it's really interesting because like when I went and visited one of the big labs in the Bay Area and had a chat with their kind of main Gen AI team. And they were like saying to me like, what are you doing? And I was trying to explain this and like, I remember one of them like looked at me and said like, so what's this for? Like, it sounds like it just brings the cost down. Like, yeah, that, that's like, oh, like when I say, let's make it more accessible. That's the definition.

• 26:43Speaker 03

  Like if we can bring the cost down by 10x, which I think we can. Then like lots more people can use it. But but to them in these big labs, the idea of like, it's just the same thing, but cheaper is like tedious.

• 26:59Speaker 03

  Whereas to us, it's like, it's what we're trying to do.

• 27:02Speaker 01

  Yeah. The other thing I'll say is that there's there's these moments in in, you know, industrial evolution, where a difference in degree becomes a difference in kind. So people are like, OK, well, so so what if you make the inference cost cheaper? That just improves margins, but it doesn't just improve margins. Where this is the first time the software industry has been dealing with like actual supply chain constraints, like in a long time, like physical installation and power access and how much H HBM memory can be manufactured in the whole world at any given time. So so first of all, like it makes things possible that wouldn't otherwise be possible. But I also think we're going to see a lot of uses. A lot of the things people really want from Gen AI are not possible now just because of the cost.

• 27:47Speaker 01

  So, you know, when you can imagine a future reality when it's so cheap to fine tune a model that you can do it continuously. So rather than actually viewing the model as pre trained and then fine tuned and then deploy, we basically, we only deploy Amnesiac models today that have no, can't remember anything about what they did. It's like this like Mento all the time. So they're like writing down on the context window, a scrap of paper to try to remember like, who is this person I'm talking to? And they easily.

• 28:14Jacob Effron?

  Wow, I love, I've never heard the Memento comparison, but that's a great, a great one.

• 28:18Speaker 01

  Yeah, well, it's very obvious to me. I better write that down on my skin before I forget.

• 28:25Speaker 01

  It's not going to be that long before that's going to like that. That's just fundamentally not going to work for so many use cases where you want the hyper personalization and and context and memory of of an actual artificial intelligence. So if it was cheap enough that you could continuously fine tune, we call it continuous pre training, of an individual agent, you know, on an individual inexpensive virtual machine, then, you know, that that opens up a whole range of use cases where you might you might do that. We actually might be willing to dedicate, you know, to have dedicated resources for each customer or for each thing that needs that level of customization. And so I think I think it's very easy to dismiss cost. You know, in the physical world, what they call manufacturability.

• 29:09Speaker 01

  So many of the applications of electricity were not, didn't require some kind of super genius to figure out what you would want to do with it. But making it practical was extremely difficult and required hundreds of thousands of individual experiments painstakingly done, you know, by a relatively small group of people. And, and a common criticism of Edison was that he was hyper obsessed with that stuff and not, you know, didn't, didn't have the prestige that you would have doing real science. But he didn't care. He had, he understood what he was trying to accomplish and, and was just willing to knock down the obstacles to get from here to there. And so we need a little bit of that, I think, in this world with, with, we have, we have too many splashy demos and not enough deployed products. And I think a lot of it had to do with the ergonomics of the technology itself, its

• 29:53Speaker 01

  deployability, the cost, the, the, just the usability of it. And then, yeah, there's a few, you know, there's UX breakthroughs that are required. There's, there's real creativity required too, but let's not neglect the, the need for real infrastructure. And, and again, that's, we haven't had that in tech in a while.

• 30:08Speaker 03

  There's two kinds of inventions, I think, right? There's the light bulb kind, which as you say, Eric, was basically like, oh, you know, electricity, we can store and distribute energy. Energy can be used to make things glow. Yeah, obviously, let's, okay, let's figure out how to create a light bulb. And the invention, Edison's invention factory were good at those. But they were also good at the other type, which is like, okay, we could use that stored energy to rotate a disk. We could put little divots in the disk. We could have those little divots use a magnet convert to movements of like a, a, a flat thing, and that could then create sound waves and we could use that, you know, we'll call that a phonograph, we can use that to play music.

• 30:51Speaker 03

  Or we could use the energy to compress gas and when gas is compressed, it makes things cooler and we could use that to create a kind of refrigerator. Like, those are clever. Like, those are a different kind of invention, which is like, a big step. Now, interestingly, The thing that that lab had that allowed them to create those things was getting all these, you know, people together, not heaps of them, but these people who live and breathe electricity all the time, you know, they're not necessarily professors of physics, although some were, you know, some were boiler makers or, you know, engineers or whatever, you know, and

• 31:33Speaker 03

  just have them making lots of stuff. So that, you know, Edison always said, you know, that that he felt that their group of people had this great gut feeling of what this technology could do. And that there were very often situations where fancy physicists would write papers about things electricity, you know, provably can't do. And they'd be like, yeah, but obviously you can, you know, obviously it can. And this is the kind of like AI factory. We want, you know, one where this is this group of people that just has

• 32:15Speaker 03

  a really good intuition for like, what kind of stuff we we can make, what kind of stuff doesn't make sense to make. And it doesn't all have to be based on theory, you know, some of it will be for like bulb kind, you know, like the stuff we've already released is definitely of the light bulb kind, you know, everybody, I think, understood that if you combine quantization and distributed computing, you could dramatically improve, you know, the accessibility of fine tuning and drop the price. And just doing it was really difficult. But, you know, I I hope we'll also be successfully able to create

• 32:54Speaker 03

  the, you know, the phonograph style inventions by by creating this group of passionate enthusiastic individuals. So one of the things we do, for example, is we give everybody in the team a $500 a month credit card to buy any AI stuff they come across and use the hell out of it. Because

• 33:19Jacob Effron?

  What have been the most common purchases so far?

• 33:23Speaker 03

  Oh, you know, it's just the normal stuff like coding helpers, you know, Copilot and, you know, everybody loves perplexity and all that kind of thing. But, you know, the idea is like,

• 33:35Speaker 03

  There's a lot of overlap in the Venn diagram of what people choose to use, but there's also a lot of differences. And it's in those differences that we get this kind of like these great conversations where people say like, oh, well, you know, if you've ever used

• 33:49Speaker 03

  Cohere's Command R plus model, you know, then you'll realize that their particular approach to retrieval has these particular benefits without having those particular drawbacks. And, you know, we can see, and it's like, you know, you've got to be, yeah, really deeply invested in in the ecosystem and the products as well as the research to, to appreciate these opportunities and constraints, I think.

• 34:12Speaker 02

  So I know we've talked a little bit about this already, but I'd love to go a little bit deeper and understand if there are other areas or ideas that you're going really deep into right now, or if there are any that you've had that lightbulb moment with that you mentioned.

• 34:25Speaker 03

  There are broad areas where this new ability to compute with language and to interact with a computer with language is such an obvious fit. So two particular examples we're excited about, and this is not at all controversial, is the law and education.

• 34:50Speaker 03

  These are both like pretty heavily language-based things.

• 34:58Speaker 03

  One of the lawyers we works with describes a law firm as just a very large language model. You feed in text and you get out text.

• 35:09Speaker 03

  Also, there are places where there's huge opportunities to to help society a lot, right? So, so let me take both of them in turn.

• 35:23Speaker 03

  The law.

• 35:26Speaker 03

  Is used a horrifying amount as a weapon by wealthy people and organizations to stop less wealthy people and organizations from doing things that they really ought to have every right to do. And every time that happens, that's a grave injustice. But it's something that as a society on the whole, we just kind of all accept now. You know, it's like, oh, yeah, you can't, you can't win that battle. It's like, well, what the hell, what was the law there in the first place for? You know, if like not to protect you in that exact situation from that exact thing happening.

• 36:06Speaker 03

  If we can, again, as a price thing, if we can bring down the cost of high quality legal advice, you know, that's, that's just damn good for the world, you know? That, that means that the law can be used by a weapon less often, you know, by, by less people. And also, you know, particularly in countries like the U S, the law is often used for gatekeeping. So, you know, you, you can't enter this regulated market unless you go through all these loopholes. And, you know, this has always been something I'm passionate about is like if somebody has an idea

• 36:50Speaker 03

  for a thing they want to bring into the world, I want to help knock down all the things that would stop them from doing that. I'm a homeschooling dad. Uh, I have spent many, many hundreds of hours, probably thousands of hours, researching the academic literature of education and listening to teacher interviews and of course, teaching my own child. You know, there are so many opportunities to just improve education so that more people can.

• 37:25Speaker 03

  Do the things in their lives that they have in their head, that they can build the things they want to build, that they can be the people they want to be. You know, currently, not nearly enough people have that opportunity. You know, teachers are amazing, but they have to work within this very constrained environment where they have a whole bunch of kids that have to go through the same path at the same time and the same way with pretty limited ability to customize. So I've seen for my own child, how delighted she is when we remove that. If we can help, you know, more people access the law and more people access education,

• 38:07Speaker 03

  both of which are really heavily language in, language out things, that would be amazing.

• 38:14Speaker 02

  Jeremy, I'd love to discuss a piece you published yesterday, actually, around the bill SB 1047, which is placing, you know, limitations and regulatory checks around training foundation models. We'd love to hear your thoughts on that and why you're concerned about it.

• 38:28Speaker 03

  You know, let me start by saying, you know, this, this particular law is a California, is a proposed Californian state law. I think, you know, it's clearly created with the best of intents, you know, which is, and it has some good features. But, you know, I want to particularly talk about one part, which I don't think is a good feature.

• 38:52Speaker 03

  Which is a part which attempts to ensure the safety of AI models, which sounds like a good idea. We don't want unsafe anything. We don't want unsafe planes or cars or toothpastes. So why would we want unsafe models?

• 39:11Speaker 03

  So as it happens, about a year ago, I spent many, many months studying this exact question. Because at that time, a group called Governance AI, along with groups like OpenAI, were involved in releasing a, a new framework for creating laws of exactly that kind. And they were kind enough to, to provide a draft copy of it to me and I agreed to kind of red team it, basically.

• 39:37Speaker 03

  And I interviewed over seventy experts across the spectrum of all the kind of fields that touch on this question of is such a policy a good idea? And it turns out counterintuitively that not only is it likely such a policy would be ineffective, but in fact, it would be likely to cause the opposite result. It would actually be likely to create a less safe situation. And there's a number of reasons for that, but in the end, the key reason is a model,

• 40:16Speaker 03

  the model itself, you can't ensure anything about it because it's the it's just the most purely kind of dual use technology like a pen or a piece of paper or the vote or education or a calculator or the internet.

• 40:38Speaker 03

  Which is to say, if I make a model, right, and I can test the hell out of it and be like, oh, it never creates bioweapon recipes or whatever, but then you can take it and fine tune it, you know, or prompt it or whatever to do whatever the hell you like. I can't stop you from doing that any more than if, you know, you could buy a calculator from me and use it to do calculations necessary to create, you know, a nuclear bomb. Or you could buy a piece of paper from me and use it to write down a death threat. You know, I can't ensure that you don't do those things with my dual use technology. So what happens in practice? So if you put in such a regulation, what has to happen in practice is that you never

• 41:23Speaker 03

  get the pen and paper, you never get the calculator, you never get the model, all you get is some kind of layer on top of it. It's just like, oh, we'll provide a service where we will write things down for you, but you can't have a pen and paper, you know, or you don't get the vote, but you can advise us on who you would like to run the country or whatever, you know, you kind of like put these layers in between to ensure that nobody could ever do anything bad. You don't give them the power at all. That's the difference between, you know, for example, ChatGPT, which is a great product, you know, and I like it and I use it, right? But that's a product where you're not using a model. You're using a layer on top of that, which you kind of can ensure the safety of, right?

• 42:08Speaker 03

  Because I don't have access to the model itself versus like Lama three, where I can actually download the model and use it for whatever I like.

• 42:17Speaker 03

  So if you create a situation-

• 42:18Jacob Effron?

  To make a really efficient fine tune, right?

• 42:21Speaker 03

  Uh huh, yeah.

• 42:21Jacob Effron?

  So that happens to be your purpose.

• 42:24Speaker 03

  Yeah, so if you, yeah, so if you say you're not allowed to release a model unless you can ensure the safety of it, then in practice, it means you can't release models. You can only release products on top of them. But models in their raw form are much more powerful, much more powerful. Because you can fine tune them, you can study their weights. You can, you know, look at their gradients and activations, you can really control them. You can do clever caching with them, you know? So anybody who has access to the underlying model, you know, has huge advantages over somebody who doesn't. So therefore, as soon as you say like, well, nobody's allowed to release such a model, which is what happens in practice, if you say you have to ensure the safety of it. You've now made the models themselves an extremely rival risk good,

• 43:07Speaker 03

  a jealously guarded secret only available to big states and big companies. Something you fight over, something you compete with, you know?

• 43:17Speaker 03

  And so now, what have you done? You know, you've taken something which is this jet jewel use technology like a piece of paper or like a calculator, which lots of people could have been using to like improve cybersecurity or improving, you know, vaccines or, you know, doing all these things that are defensive in nature. And in fact, instead you've made it an extremely rival risk good that a small number of super rich organizations compete for access to, you know, and have to race to make bigger ones and better ones. And nobody has the ability to, there's no transparency, you know? So nobody can be studying how they work and what they can and can't do and so forth. So that's basically the issue.

• 43:58Speaker 03

  It's, it's, it's exactly the kind of legislation that my analysis last year showed can't work. And at any time you try to ensure the safety. Of a dual use technology like a model, then by definition, you create a highly competitive space, massive centralization of power and reduce transparency.

• 44:22Jacob Effron?

  Super interesting. And I guess, you know, going back to the conversation we're having earlier, Eric, I'd be curious, like, you were talking obviously about, you know, the the kind of foundation model companies and how taking all this massive fundraising rounds and then going in a lab and building a model for a year before getting it out there is maybe not as, you know, not the classic lean startup approach. I'm wondering, like, if you were running one of these foundation model companies, like, what would you do differently?

• 44:47Speaker 01

  Well, it's actually not the foundation model labs that I have really the issue with because their mission is to try and discover whether AGI is possible to build. And

• 44:57Jacob Effron?

  That's fair.

• 44:57Speaker 01

  To whatever, to whatever degree they commercialize anything is just to show the progress necessary to raise the money, to do the research, to do the thing. And I'm glad, I'm glad there's people pushing the frontier. There's just an unbelievable reservoir of applications that don't require AGI to unlock that we're not building and we're not focused on because the fund, I mean, it's really the fundraising gravity has pushed so many entrepreneurs away from utility and into science fiction and speculative stuff. And so I think we're just, we're missing out on a lot of actual practical uses. And in fact, you know, one of, one of, one of the investors in Dancer AI is a major AI safety advocate. And one of the things that I think was really a key realization for

• 45:40Speaker 01

  him was that if we live in a world where the only models available are frontier models, and, you know, we could debate whether they're at AGI now or they're going to get to AGI, I don't want to get into all that. But just, you know, whatever the closest thing to AGI that's available is really the main thing that people are using, then you're actually, when you, when you deploy those models, you're asking, you connect them up to real world systems, you're just really asking for a lot of unsafe things to happen. That in theory could have been done with smaller models that are, you know, that are properly fine tuned and are basically safer by definition. And, you know, some people, of course, think that, that there is no safety worry in the front, you know, like, I don't want to get, it's a very controversial point, maybe many different

• 46:22Speaker 01

  views about it, but my point is that everyone can agree that there are ways to build a huge class of valuable applications that are intrinsically safe. And if nobody provides those options, then people will default to doing the unsafe thing, you know, and so we don't want to live in a world where there's no open source, where there's no open source research, where there's no ability to fine tune, where there's no ability to, to build these cool models, build that, build applications out of small models. It doesn't, I don't think that actually advances the cause, the cause of safety.

• 46:50Jacob Effron?

  Yeah, super interesting.

• 46:51Speaker 01

  Yeah, but, but all that said, having spent some time with the foundation labs and knowing the challenges they face, as they're getting bigger, some of them have become quite schizophrenic, where you have this like safety research AGI thing that was, it started out to be, but then you're building this unbelievable commercial apparatus. And in a lot of cases, the commercial people have only the faintest idea what the safety agenda is, or in some cases, I think basically don't care, are just there to make number go up. And that's, I think that's actually where it can get quite dangerous because now you have like, you have the rival tensions within the organization that can, that can rip it apart and you can lose the coherence and the alignment that's needed to keep the company together. So, so certainly one thing that I would do if I was in those, in those labs would be

• 47:35Speaker 01

  to try to reestablish the connection between the research and the customer, as we've been talking about. I don't think that's very realistic. I think they're not that interested in doing that, but if they viewed it as their responsibility to see that their customers were successful in the deployment of the technology for, you know, and they could, I was, I would have to take a point of view about the applications you want to affirmatively see be able to be built and go, you know, take an old page, page out of the Toyota production system playbook and like go and see for yourself, go sit with the customer and really like make it your responsibility to see that they succeed. I bet they would learn a ton from doing that. I hope more of them embrace that.

• 48:11Jacob Effron?

  Super interesting. Well, this has been a fascinating conversation. We always like to end our interviews with a quick fire round where we get your quick takes on, on a few common questions we ask our guests. And so maybe to start, would love, I guess, for each of you, one thing that's overhyped today in AI and one thing that's underhyped.

• 48:26Speaker 03

  Overhyped agents, underhyped resource efficiency.

• 48:32Speaker 02

  Why do you think agents are overhyped?

• 48:35Speaker 03

  This is the reason I mentioned earlier, the, you know, the things that most people are trying to do with them are not compatible with the mathematical foundations of the models.

• 48:46Speaker 02

  Next up is what AI startup outside of your own are you most excited about right now?

• 48:53Speaker 01

  I mean, there's so, God, there's so many to choose from that are so interesting.

• 48:58Jacob Effron?

  Yeah. Where are you putting your $500 budget toward?

• 49:01Speaker 01

  Yeah, totally. I mean, I, I, listen, I'm trying to buy, I want to buy more AI products. Please, everyone listening, please sell me some AI products. I'm, I'm, I'm dying to buy them. I mean,

• 49:10Jacob Effron?

  Everyone had to answer, you got 500 bucks to put them, say, you get started.

• 49:14Speaker 01

  I, I really do. I actually really, I mean, it just feels so lame to say that I really like perplexity, but everybody, because everybody does, I think. We haven't talked, we've been about, about cost as a, a degree in kind change, but so is speed. And the speed at which it, it operates is really, is really unbelievable. Actually, I'll give you a more original answer than that. Jeremy, you answer, I'll, I'll keep thinking about it.

• 49:39Speaker 03

  I'll go with Grok with a Q.

• 49:42Speaker 03

  For the same reason, they're fast and

• 49:46Speaker 03

  taking a different approach.

• 49:49Speaker 03

  And it's an approach which I think is again, foundationally sound. I don't think they count as a startup anymore, but I just want to say also Meta are the most important organization for AI right now.

• 50:04Jacob Effron?

  I guess what's one, what's one breakthrough? You know, obviously there's lots of things that are talked about that might happen on, in GPT-5 and beyond. What's, you know, you obviously feels like you have your kind of current beliefs on the capabilities of these models. What's one breakthrough that would change your mind on some of this stuff?

• 50:21Speaker 03

  I think there's, there's, there's two, you know, the, the first big issue right now is just

• 50:28Speaker 03

  massive energy requirements, which is just putting a, just a, a direct, you know, economic or even physical obstacle as to what we can do. The models are already very good.

• 50:47Speaker 03

  But it just, you know, there's just a lot of resources required. So

• 50:54Speaker 03

  a breakthrough in

• 50:57Speaker 03

  The amount of energy or other resource requirements would be huge. Even bigger would be a breakthrough in this kind of like planning and reasoning capability, something that goes past subgraph matching. So, you know, in particular, Jan Le Cun and his group have been working on these, they're called JEPa-based models for a while now. They are, you know, quite a different approach. There are also other folks working on what are called diffusion models for text, again, quite a different approach. The current approach, which is like, all right, pick a word. Okay, got it. Now pick your next word.

• 51:42Speaker 03

  Okay, got it. Now pick your next word. Like, that's literally how they work. It's called autoregressive.

• 51:48Speaker 01

  Well, when you put it like that, it sounds less exciting.

• 51:51Speaker 03

  I mean, they're very good at picking words, right? But the actual important word might be like in 50 words' time, you know? But if the previous 49 words didn't head you in the right direction, that's going to be a problem, you know? So these other approaches like JEPa and diffusion might get us out of this autoregressive approach and allow us to move into a land where we're doing more than, more than pattern matching. Although, you know, it turns out pattern matching is way more powerful than I think a lot of people intuitively expected, but it has its limitations.

• 52:28Speaker 01

  The most interesting thing to me about LLMs has been the change in my perception of what human intelligence is all about. And, and the fact that far more of human intelligence is obviously encoded in language and linguistic processing than we previously thought. And one of the wildest facts of the LLM revolution is that every weirdo post-modernist theory that's ever been cooked up in a, you know, an English department anywhere, in a literature department anywhere in the world, every, every obscure theory, and they're all going to turn out to be empirically testable. Who would have thought that's like? Totally wild. And the thing that I'm waiting for, like, I think that would really blow my mind, but I don't view it as like implausible, is that something in this, we

• 53:09Speaker 01

  will, someone will have a breakthrough and realize that, that the problem is not in our understanding of the scaling of LLMs, the problem is in our understanding of cognition itself, in human cognition. We don't, since we don't understand how it works, we've been trying to emulate it in this super brute force way. And the metaphor that I always keep coming back to is, have you ever seen one of those YouTube videos of someone who's built a calculator out of Minecraft blocks? And it's like, it's just, it's so amazing. Like, wow, I can't believe that Minecraft blocks are a general purpose computer, and they're turning complete, and you can compute anything you want. But like, it's basically the most brute force possible way you can think of to make a calculator, except the thing that I'm about to describe to you. Imagine if I said, look, I'm not going to do that. I'm just going to try every possible combination of Minecraft blocks until one

• 53:51Speaker 01

  of them happens to be able to do calculator level computation. I would, I would eventually find it if I put enough, throw enough GPUs at it, do enough math, it can be found. So like, is there something like that lurking in here where we're going to wake up one day and realize that we brute forced our way to an extremely, the most inefficient possible way of discovering some critical algorithm in cognition? And actually, if you could do it directly, then you would be, it would be a breakthrough. Now, people have been looking for that thing for a long time and haven't found it. So maybe, maybe this is the best we can do. But that would be the thing that would shake my, shake my beliefs quite a bit.

• 54:26Speaker 02

  I'm sure a lot of folks will want to learn more about you both and Answers. You know, where's the best place for them to go to?

• 54:34Speaker 03

  Answer.ai.

• 54:36Speaker 02

  Easy.

• 54:37Speaker 01

  How did you get Answer.ai? That's like a great domain. I feel like that must have been a competitive process.

• 54:42Speaker 03

  No, I had it for ages. I, I

• 54:44Speaker 01

  Really?

• 54:45Speaker 03

  Yeah, yeah. I grabbed a bunch of dot AI domains many, many years ago.

• 54:49Speaker 02

  Nice.

• 54:50Speaker 01

  I'm surprised you didn't just fund the company off that.

• 54:54Speaker 03

  One of the ones I bought, but I, but I actually sold was emergent.ai. And emergent.ai for many years just had a single sentence on it, which was, which was just, AI has not emerged yet.

• 55:09Speaker 03

  But then somebody

• 55:10Jacob Effron?

  It would be a scary day when you went to the site and that had changed, so.

• 55:13Speaker 03

  Exactly. That's what it was there for. That was my little canary. Yeah.

• 55:18Jacob Effron?

  That's awesome. Well, seriously, thank you two so much. I feel like this has been a fascinating conversation. I'm sure folks will want to pull a lot of threads and it's obviously a privilege of our job to get folks like you on. So really appreciate the time.

• 55:29Speaker 02

  Thanks for saying so.

• 55:31Speaker 02

  So that was a interesting conversation. What did you think, Jacob?

• 55:34Jacob Effron?

  Yeah, I mean, talk about getting two legends on the podcast between Jeremy and Eric. I feel like we could have done two hours with with with one of them. So to get them both on is is a real treat. And I feel like they have really interesting perspectives. I mean, Jeremy is such a polymath, right? You know, he's obviously super detailed in AI regulation. And then he had founded a healthcare AI company before and, you know, was super deep in in obviously in the education space through his work at Fast AI. I think, you know, I'm sure it's a great personality to run a research lab because there are so many different directions that that he can go in, which I think is what's so interesting.

• 56:06Speaker 02

  Yeah, I mean, I think it's an interesting combination between him and Eric. You know, a lot of the research labs we see are just hardcore AI people and, you know, they go in to sort of solve AGI. But here you have Eric, who is all about customer facing and in learning about what the customer needs. And so it's an interesting marriage of the two.

• 56:22Jacob Effron?

  And I love their idea that basically there's there's just so much more we can build today with today's capabilities. I mean, I think, you know, their point that, yes, everyone is obsessed with what's the next frontier and, you know, what's reasoning and planning and, you know, agents that take action. But like, just with the capabilities of the models we have today, if we make them cheaper and more accessible, there's so many applications waiting to be built. And I know we we certainly feel that way. I feel like I keep saying 2024 feels like it's the year of applications. You're really starting to see folks that have built on top of models with GPT-4-like capabilities for a longer period of time, just in every vertical, discovering lots of really cool applications. And it's cool that they, you know, legal and education are obviously two of the most interesting ones for them to be spending time on.

• 57:01Speaker 02

  Yeah. What were your thoughts on the health comments as the health expert?

• 57:04Jacob Effron?

  You know, healthcare is not for the faint of heart. I agree. And no, it's true. I mean, I think that I think so much of what Eric says really resonates on the lean startup side where it's like, you don't want to build something and then bring it to someone and realize their incentives are warped to not even use that product. And I think that's a classic problem we've seen in some areas of tech, healthcare, one of them, where I think a lot of the best entrepreneurs you see in the space really understand the end incentives of of different folks in the ecosystem. And then, you know, the challenge is can you still build a mission driven impactful product that meets those, you know, and fits into those incentives. But the magic really happens when you're able to do that. And so I think, you know, I love their push to like, go talk, go have the conversations and talk to the end customers before you build the product.

• 57:48Jacob Effron?

  You know, because I think that that's going to just guarantee much more impact.

• 57:52Speaker 02

  Yeah, no, I mean, I agree. And I think it's super interesting, their approach to more of the application side of things that we haven't really seen in a lot of the other research labs. You know, the fact that they have three lawyers in house, working on the legal side of things. I think it's super interesting. And, you know, to your point, you know, legal and education are just two areas. We're very interested at Redpoint, but it's such a clear value prop where you can go in and talk to tons of customers and really dig into the pain point. It's very easy and clear to do that.

• 58:17Jacob Effron?

  Yeah, no, totally. And I like this idea that law was just, you know, language in, language out. You know, I think our job is probably analogous to that as well. And so, yeah, but I think, you know, certainly, I don't know, just really interesting, you know, interesting approach to how they're tackling the space. It's certainly different and I'd encourage folks, you know, they had, you know, there's a great piece that we'll link to of just the history of research labs and why they thought it was the right approach here. And it's definitely different, which I think is, you know, it's cool to see, you know, people take this transformative technology and then structure organizations in a ton of different ways around it. And we're seeing all sorts of innovative things from the kind of nonprofit, open AI board governance to, you know, to this.

• 58:59Jacob Effron?

  And a lot of experiments being run about what the right way to, to, you know, use and responsibly use this technology.

• 59:05Speaker 02

  Yeah. And I think, you know, it wasn't on the application side, but I do think the conversation around the regulatory aspects of AI and just hearing Jeremy's perspective on how, you know, the more we try to make it safe, the less safe it becomes. And really striking that balance. It was counterintuitive, but it's super interesting.

• 59:21Jacob Effron?

  Totally. I know that's something we really, you know, obviously we're huge proponents of open source. And I think, you know, when you get something out there and get the most eyes on it, that's, you know, when, when ironically, it ends up being, you know, the safest. And so I think that certainly, certainly makes a lot of sense. It was definitely a really fun one. Yeah. Got some other great guests coming up in a few weeks. So we'll look forward to seeing you then.