David Luan: Why Nvidia Will Enter the Model Space & Models Will Enter the Chip Space | E1169

1. 0:00 – 1:03

   Intro

   1. DLDavid Luan0:00

      What OpenAI realized before basically everybody but DeepMind was that the next phase of AI after Transformer was not going to be about research paper writing. It was going to be about, "Let's choose a major unsolved scientific problem and just try to solve it." The second way of improving model performance is just starting to be tapped now, and that's also going to absorb a boatload of compute. So because of that, I actually am not worried about diminishing returns to compute over time. I think every tier one cloud provider existentially needs to win here.

   2. HSHarry Stebbings0:29

      Ready to go? (instrumental music plays) David, I am so excited for this. I've wanted to do this one for a long time. I've heard so many good things. So first, thank you so much for joining me today.

   3. DLDavid Luan0:51

      Yeah. Thanks, Harry, for having me on it. I've, um, got to watch some of your cool previous episodes, so it's a, it's a real honor to be on here.

   4. HSHarry Stebbings0:57

      And that's very, very kind of you. I really do appreciate that, man. But you, you've been at some incredible companies as a training ground, so to

2. 1:03 – 5:06

   Lessons from Google Brain & Their Influence on Building Adept

   1. HSHarry Stebbings1:03

      speak, one of which was Google Brain, and I just wanted to start there. When you think about your biggest takeaways from your time with Google Brain, what would you say ones two are, and how do you think that shaped how you think about building Adept today?

   2. DLDavid Luan1:15

      Google Brain, um, uh, uh, was, and also now is part of DeepMind, is a really magical place. It's, uh, it's the, I think during the, uh, peak days of AI, AI progress on the research side, right, where every day there was a new paper that came out that would just change the world, that, like, 2012 to 2018 or so era of Google Brain was just, like, incredibly dominant. They did an amazing job picking talent, like the people who invented Transformer, the people who invented the diffusion model, people who did, um, all of these new optimization techniques that we all take for granted today. They were all at Brain at the same time. Like, truly the Bell Labs of, of the era. And I think I learned a lot about how to make, um, how to make pure bottom-up, like to see what good pure bottom-up basic research looks like at Google Brain.

   3. HSHarry Stebbings1:55

      What does that mean, what pure bottoms-up good basic research is?

   4. DLDavid Luan1:58

      Uh, yeah. So, so I have this worldview, um, of, of AI progress as being a part of a couple different phases, right? And I, I like to think about pre-2012 as basically being pre-history. Of course, um, all of the, uh, OGs in the field would probably not like that if I characterized it that way, but before 2012, like, most of the things we tried just didn't really work, right? Like, you, you had things like, um, uh, a sheep being identified as cats and dogs, and, uh, chatbots that, um, barely said anything coherent, et cetera. But I think, like, there was a p- uh, there was a period between 2012 to, like, 2017 or 2018, where deep learning went from something that, um, people didn't believe in to being, like, the dominant paradigm in the field. Right? And so during that 2012, 2018 era, the way people made progress, um, what I mean by bottom-up basic research is you hire the most brilliant scientists. They come to work every day with, like, no, um, near-term objective they're being held accountable to, um, and they just work together and they think about, you know, "Hmm," like, "I wonder what it'd be like if we could solve this, like, open, uh, technical problem in AI." Like, "How do we go, um, how do we create a model that better understands how to generate images?" And they just go work on that out of their own curiosity and drive and maybe some interest in glory and fame through papers. And they do that for, uh, they, they do that for, like, six months or so, and then out pops out this, like, research paper that gets posted to arXiv and goes to a p- a, a journal that, um, that, like, just solves the problem. Like, that's, that's huge. Right? And so, m- the reason I call it bottom-up is because it's just driven by the natural, uh, interactions between all, all these researchers in a, in a setting, and they figure out what they wanna do.

   5. HSHarry Stebbings3:35

      And then what's the next wave, then? You had that 2012 to 2018. How do you categorize the next wave?

   6. DLDavid Luan3:40

      Well, I think what happened was in 2017, the Transformer came out. Um, and I remember, um, I was actually... I was running engineering at OpenAI at the time, and uh, um, I was working really closely with Ilya. Um, and Ilya and I were just sitting around and, and he was just like, "Look, this Transformer thing is, is, is, is real. It's gonna be the next most important thing. Let's get all of our teams looking at how we can use this thing." And so the thing that most people in the general public don't know about is we didn't invent Transformer at OpenAI. It was invented at Google. Uh, and, uh, and, and, like, but what Transformer did though was it was the first time you had a model that was sort of generally applicable to any machine learning task. Right? Like, back in the day, if you wanted to understand images, you used a convolutional neural network, right, like AlexNet. If you wanted to understand... uh, if you wanted to, to, uh, generate, uh, text, uh, you used RNNs. If you wanted to beat humans at Go, you used, uh, you used, uh, uh, like, um, m- uh, tree search or RL. Right? So you had these... all these different, like, m- different, um, models that you would use to solve problems in AI. And then Transformer kind of became, like, the universal model, and, like the, like the, like, base element of AI that came out at that time. And once Transformer came out, um, in a weird way, you kind of stopped needing to go, uh, uh, uh, stopped needing to go make, like, super low-level breakthroughs in modeling, because it kind of just worked for everything. And then you get... got to go take that thing to solve really,

3. 5:06 – 6:49

   Why It Took 6 Years for ChatGPT to Emerge After Transformers

   1. DLDavid Luan5:06

      really big problems.

   2. HSHarry Stebbings5:07

      The show's been successful because I'm not afraid to ask stupid questions. (laughs)

   3. DLDavid Luan5:10

      Shoot.

   4. HSHarry Stebbings5:10

      Um, which is very uncharacteristic of VCs. But, uh, 2017, Transformers, what a breakthrough.

   5. DLDavid Luan5:17

      Yeah.

   6. HSHarry Stebbings5:17

      But ChatGPT seems to be the consumer breakthrough that we all waited so many years for.

   7. DLDavid Luan5:22

      Ah.

   8. HSHarry Stebbings5:22

      What, w- what was the reason for that chasm between Transformer breakthrough and consumer adoption breakthrough with ChatGPT?

   9. DLDavid Luan5:30

      Yeah. Yeah. No, that's a really good question. It's kind of like ChatGPT was like, was like the frog that ultimately became boiled. Right? Like, Transformer was a huge breakthrough, and every incremental year from 2017 to when ChatGPT came out, um, language models just got a little better and a little better and a little better. Like, I remember, um, Alec Radford, um, uh, and a couple others, and I did GPT-2. Right? And GPT-2 came out in, I think, 2019. And I just remembered, um, you know, this th- this... Finally you had this, like, pretty smart generalist...... model that you could just say, "Hey, like, write me a newspaper article about insert celebrity being arrested in LA." And it would just do a perfect job, and you'd be like, "Oh, they were in, like, they were in, like, the Neiman Marcus store," et cetera. Um, I thought that was so much fun. And, um, and, um, but, like, the thing is, like, two things had to happen. One, the models were getting increasingly smart, but there's, like, a minimum viable smartness where you're like, "Damn, this is a compelling experience." And th- the second thing was it needed to be packaged up in a way that consumers could play with. So if you go look at the lag, right, ChatGPT was really just GPT-3 with, uh, instruction, uh, r- basically, like, more chat tuning. But GPT-3 API came out, I think, over a year before ChatGPT did, but only developers could play with it, so there's no viral aha moment for consumers. So the, the packaging and the intelligence had to exist in order for, uh, for that virality moment to happen

4. 6:49 – 9:57

   Takeaways from OpenAI

   1. DLDavid Luan6:49

      with ChatGPT.

   2. HSHarry Stebbings6:49

      We mentioned ChatGPT there. Before we dive into the meat of the show, I do just have to ask, you mentioned your time with OpenAI. It's such a transformative place to be. You mentioned working with Ilya there. Wha- what's one or two of your biggest takeaways from OpenAI that really informed how you think about building Adept?

   3. DLDavid Luan7:04

      Well, the first one, actually, is, and going back to the, like, eras of AI discussion we were having, right? The, um... Well, OpenAI realized before basically everybody but DeepMind was that the next phase of AI after Transformer was not going to be about, you know, research paper writing. It was going to be about, "Let's choose a major unsolved scientific problem and just try to solve it," right? And so, like, that led us to go build a culture. Instead of, like, loose collections of federations of researchers, let's put a giant team around, how do you solve, like, robot hand control? Let's put a giant team around beating humans at, like, one of the most popular video games on the planet, right? Let's put a giant team around scaling GPT until this thing is like a generalist reasoning and chat engine. Like, that's just a totally different framework from, like, this very academic curiosity-driven research. And I think that that's the right framework. And, uh, and I think that's a big part of how we build Adept now as well.

   4. HSHarry Stebbings7:57

      So it's the focus of large groups of scientists on specific real-world problems, not on scientific paper creation?

   5. DLDavid Luan8:05

      Exactly. So it's like the switch from, like, building, um, a, uh, hiring, like, a thousand people to go sit around thinking about how to put together small rockets versus, like, creating, like, the Apollo project, right? Um, it's m- a lot better to say, "Hey, our goal is to go to the moon, and, uh, we're gonna hire however many people it takes to go solve going to the moon." It's very different than just, like, a giant mass of people organically self-organizing to do that.

   6. HSHarry Stebbings8:30

      I'm interested, you said it abou- earlier about kind of the transition of models. You mentioned GPT-2, moving to GPT-3. When we think about model performance today, people are starting to say we're seeing diminishing returns, that more compute does not need- lead to better performance. I interviewed one of the most prominent people in AI the other day, and they said that A- uh, that OpenAI was actually, um, disappointed by the lack of performance that throwing more compute did to their latest release. Do you think we are seeing diminishing returns now, more compute does not lead to more performance?

   7. DLDavid Luan9:05

      I don't think so. And here's, here's, here's why I don't think so. So, um, it all depends on what access you use, right? So the way that, um, the way that giant model scaling works, historically, if you go look all the way back to GPT-2 to GPT-3 to 4, et cetera, um, f- wha- the way it's worked is that the, um, you know, every... Let's say, just to use a reductionist analogy, every incremental GPU you throw at the problem actually does have diminishing returns. But every, um, every, like, uh, power of tw- every doubling of GPUs you throw at the problem has very predictable, consistent returns. And so it's kind of like a l- a, a, like a logarithmic curve versus a straight line, right? Depending on what access you use to go look at it. So put another way, um, for just scaling up a base language model, um, you need to double the amount of compute for that language model for it to be predictably, consistently smarter.

5. 9:57 – 16:06

   The Key Bottleneck in AI Model Performance

   1. DLDavid Luan9:57

      Does that make sense?

   2. HSHarry Stebbings9:57

      It totally does. So it's like, okay, actually, there's a lot more room for improvement with the increase in compute availability that we should and will have.

   3. DLDavid Luan10:05

      Yes.

   4. HSHarry Stebbings10:05

      I had Alex Wang on the show, and he was like, "It's not algorithms, it's not compute, it's data that is the bottleneck to AI model performance."

   5. DLDavid Luan10:14

      Yeah.

   6. HSHarry Stebbings10:15

      How do you think about that? Is that true?

   7. DLDavid Luan10:17

      So let me put it a different way. I think that, um, I think, uh, I am not a lawyer, but e- but even so, I will say it depends. Uh, and so I think the, the way, the better way to go think about it, in my view, is, like, there's two parts to model scaling with compute. Um, one part to it is you simply make the model bigger, and then you throw more data and more GPUs at it. If we go look at CPUs in data centers, right? For a long time, we, we had Moore's law, right? Every year, the, every year, chips would get better at some predictable pace, and everybody says, "Ah, Moore's law is gonna die. You know, we're at three nanometers or whatever. There's, like, no more nanometers left for..." But what actually happened is you go look at the amount of compute available, um, uh, even, like, for, for, for chips, it's actually, um, continued to trend up because now what we do is we build systems that have multiple chips in them. So we have, like, both the scale-up of a single chip and the scale-out, and as a result, every year, we s- humanity has more and more compute available to it, right? It's the same thing with, um, with giant model scaling. Like, the base model itself, um, even if the base model itself stops scaling at some point as you throw more compute at it, there's a whole new way to go make models smarter that is just being tapped right now. And that whole new way of making models smarter is not just making the base model larger, but it's by having the base model, uh, basically, by s- like, having the base model collect data for itself to learn how to get smarter. So let me give you a concrete example of this, right? Concrete example of this is, right now, let's say you wanna tr- uh, you wanna train an LLM to get better at, um, at solving math problems. The way you do it is you collect lots of, uh, you collect lots and lots of, like, positive solutions to hard math problems, and you throw it in the dataset, right? And you're like, "Hey, like, model, good. Go get smarter at this thing." But a much better way to solve this problem is you give the model that you're training...... access to a theorem-proving and math environment, right? Like, uh, like, uh, like give it a Jupyter Notebook. Give it, um... There's a, there's a theorem-proving library out there that a lot of people use as well. Like, give the model direct access to those tools, and then say, "Hey, I want you to experiment. I want you to go try solving this problem, and then reflect on it. Like, did it, d- did you do a good job? Like, is this problem solved? If no, try again." So now what you get to do is you get to have the model play with the, the simulated world basically to collect positive and negative data how to solve math problems, and then that makes the model smarter. So the second way of improving model performance is just starting to be tapped now, and that's also going to absorb a boatload of compute. So because of that, I actually am not worried about, um, about, um, the diminishing returns to compute over time.

   8. HSHarry Stebbings12:54

      Why is it just starting to be tapped now, and what does that progression look like of its own development?

   9. DLDavid Luan13:00

      Everything's an S-curve, right? So the giant model scaling, base model scaling S-curve, for the last couple years we were here. We were at the sharpest point of improvement, right? You just, you, um, you, everybody could, you know, you could double the cost of your model from $100 million to $200 million, and that would be the fastest and easiest way to deliver a smarter thing to, to, to, to the world. And now if you're getting to billion-dollar training grounds and $2 billion training grounds and $4 billion training grounds, it's really freaking hard to go get more money to go, uh, to go make the base thing bigger. And so because of that now, the critical path for model improvement is, is, uh, is shifting over to this, uh, to this, um, uh, sort of, uh, broader sort of simulation/slash, um, synthetic data/ like, RL loop sort of path. Uh, I think it's just a natural consequence of the fact that it's so expensive to just keep scaling.

   10. HSHarry Stebbings13:50

       Is it like reinforcement learning of its own datasets? Like, where it just continuously repeats the same things until it gets it right? Is that a good understanding of it?

   11. DLDavid Luan14:00

       That is a good understanding of it. So, I think, um, I think a good way to think about it is historically, for the last couple years as we scaled up LLMs, we've just been doing more unsupervised learning, right? Like, um, you get more data, more, more, more, um, more smart, uh, journalists writing articles, feed it in there, and that makes it smarter. But the problem is, like, a model trained that way is only as good as the smartest data in the training set. Like, it cannot discover new knowledge, because its job, the way the models are trained, is to do what a human would do in that situation. But the underlying thing is if you wanna go solve, like, really big problems, like solve un- like prove unproven math theorems or, like, be able to, like, help you solve a creative problem at work, those problems are by definition things that are not in the training set, because it's a, it's a, it's a, it's either like a superhuman thing, or it's a novel situation.

   12. HSHarry Stebbings14:53

       Is that why we haven't seen agent progression in the way that we wanted to or hoped we would? Because a lot of the tasks that people do are not actually codified in data. They're codified in conversations in rooms, in whiteboards, but not in data.

   13. DLDavid Luan15:08

       But not in data, yeah. I mean, I think that's a key, that's a key insight. Um, it's like, like I kinda think that chatbots like ChatGBT and stuff and, and agents, um, are kind of becoming different species of technology a little bit, right? Like, um, like, I think they'll be useful in very different ways, and, um, and what they need to be used for is super different, right? Like, just one concrete example is, um, is, is the hallucination problem. Having hallucinations in chatbots and in, like, image generators is, like, a really good thing, right? Because it gives you, it gives you, um, a- like, a starter tool for, like, getting to, like, solve the blank page problem, right? Like, and, like, gives you, like, little bits of novelty and creativity. But, um, but they, but agents on the other hand, like, if you want something to go, like, consistently, I don't know, like, do your taxes for you or, uh, handle all of your shipping containers or something like that, you do not want that thing to go randomly hallucinate and, like, make up stuff along the way, right? And so, like, these things are speciating in an interesting way right

6. 16:06 – 20:17

   Understanding Minimum Viable Capability Levels & Model Scale

   1. DLDavid Luan16:06

      now.

   2. HSHarry Stebbings16:06

      You mentioned to me before the minimum viable capabilities levels and how that is a function of model scale.

   3. DLDavid Luan16:14

      Yeah.

   4. HSHarry Stebbings16:14

      Now when you said that, I didn't have a clue what you meant, and so I was hoping that you'd be able to unpack it for me. (laughs)

   5. DLDavid Luan16:20

      Yeah. Okay, so the coolest, the coolest thing, the reason why I love working in AI is that, like, for the first time y- you're, as an engineer or researcher, you're c- it feels like you're, like, uncovering, like, like, unknown secrets about how intelligence works every day. It's, like, very different from programming. As a programmer, I show up to work, I'm like, "Here's the thing I wanna build. I know I can build it. I know if I, um, if I, uh, am clever enough, I can, I can, I can solve the problem, and I know exactly the behavior of the system that I've built will be." But the cool thing about AI is that, is that every day, you come to work, and you make some tweaks to the model, and what you get on the other end is actually somewhat unpredictable. Like, you kinda feel more like a gardener than an engineer. Uh, and, uh, and I think what's really cool about it is that, like, you know, um, as these AI systems have gotten bigger and as the architectures and datasets have improved, what the model's good or bad at, you can't totally predict ahead of time. You have some estimates for, for, for things. But just going back to the early days, right? When we were training GPT-2, we trained GPT-2 in various different sizes, and, um, at the smallest size, the model was just, like, unable to do three-digit arithmetic. But as the models got bigger and bigger and bigger, we didn't change anything else. We just, um, had it look at more data, and then we made the model bigger. And then, like, at a particular size, there's just an aha moment where from, like, where it, where it went, went from not being able to do, like, three-digit arithmetic to being, like, very good and predictably improving at getting three-digit arithmetic better. And that, like, aha moment, we couldn't know about in advance. So that's what I mean by, like, minimum viable capabilities, a- and how it's a function of model scale. Like, if you, um, uh, there are things that we really want these models to be able to do, like be really useful agents or to, to, um, help us discover new things in science or whatever. But it's hard today to say, "Hey, you know, if I just spend, like, $2 billion in compute on this model and have the right data, that'll happen for sure."... I think that's why it's so cool to work in the field.

   6. HSHarry Stebbings18:18

      When we think about, like, actually improvements in models, they lead to improvements in performance, and I, I, kind of, think there's three ways of doing that, one of which is, like, a breakthrough in reasoning. How do you think about the likelihood of a breakthrough in reasoning, what's required for that, and whether that is a reasonable expectation?

   7. DLDavid Luan18:36

      Reasoning is one of the problems in the field right now that I think, um, a bunch of us, sort of, have similar ideas for how to solve, but it actually requires some new research to be done. So, um, in a, in a weird way, working in AI is pretty funny these days, because, because, um, the giant model scaling problem is so known, and it's really a function of resources, and so if you, uh, and so you kind of don't feel like you need to be a genius to go make new progress on just pure model scaling. But I think pure model scaling does not deliver solutions to reasoning. To me, the definition of reasoning is being able to, uh, is being able to, like, compose existing thoughts to discover some new thought, right? Um, and I think to go do that, that's not something that's trained into the capabilities of LLMs by simply asking it to regurgitate the internet's worth of data. The way we're gonna solve reasoning is back to what we were talking about earlier, of like, taking theorem proving as an example. You wanna give the model access to a theorem proving environment and have it try things, um, in the same way that, like, you know, a, a human mathematician would sit down and be like, "Well, you know, here's the things I know to tr- be true about the world. How do I compose them such that I can prove the thing that I wanna prove?"

   8. HSHarry Stebbings19:49

      Is it not the model providers who will be the ones solving reasoning? Or is it actually the end, kind of, consumers or vendors who will be leveraging proprietary data sets to then utilize that to solve reasoning? Which one's which?

   9. DLDavid Luan20:02

      I think the general capability of reasoning will need to be solved at the model provider level, um, and that's because what you're actually doing is you're not just using the model to reason, you are trying to improve the model's ability to reason, which means the model itself needs to change.

7. 20:17 – 33:26

   The Future of the Foundational Model Layer

   1. DLDavid Luan20:17

   2. HSHarry Stebbings20:17

      Does that mean that we're not gonna see the commoditization of models? Everyone talks about this commoditization, "We're just gonna switch between them. Uh, it's gonna be a race to the bottom." Does reasoning mean that actually that won't happen?

   3. DLDavid Luan20:30

      No. I, I actually think that, um... I, I think that, like, solving these reasoning skills are on the roadmap of every LLM player. I do think there will not be that many LLM players. I think, I think that, um, there will probably be... my guess is somewhere between five to seven long-term steady state LLM providers at maximum scale, just because of the costs involved. Reasoning is just another expensive thing that these companies have to get right, but I think they will all solve it, because I think the way to solve reasoning is something that many of us in the field, kind of, uh, have a pretty strong suspicion on.

   4. HSHarry Stebbings21:01

      What is your strong suspicion on how to solve reasoning?

   5. DLDavid Luan21:04

      Ah, it's what we were talking about earlier. It's tr- train a base model, give it access to a wide range of different environments, uh, to go, to go solve hard problems in, and have the model try how to solve those problems, and use that and combine that with, um, with, sort of, human input on wh- whether it's doing a good or bad job, and I think that will solve reasoning.

   6. HSHarry Stebbings21:23

      Why has no one been able to solve memory? People often talk about this, and respectfully, it seems a confusing one to me, 'cause it's like, computers have memory anyway. Why, why in AI is memory such a challenge?

   7. DLDavid Luan21:35

      You can kind of think about memory as being two different things, right? You kind of have short term working memory, and then you have long term memory. I think people have made really good progress on short term working memory, right? Like, um, if you could look at Gemini, Gemini's context length is like a million... it might even be more now, I just don't quite remember, like a, like a million tokens long, which is so cool, as you can feed it, like, giant s- snippets of video and be like, "Hey," like, "write me a step by step of, like, every- everything the person cooking on this, uh, on this, in this particular video did," and it'll do it. Like, that stuff is insane. Like, that's making good progress, and the reason that's been hard is for computational reasons. Um, but I think this, this sort of longer term memory problem, this goes back to, like, another thing that I believe in, that, that's why I'm excited... um, um, I'm slightly less excited about model building and slightly more excited about application developers, because the underlying thing that everyone's realizing now is that LLMs themselves are not a product. Like, an actual product is this entire software system that uses LLMs in it. So for example, like, um, uh, what we should be doing is we should be finding ways in which, uh, in which, like, end application builders can be themselves responsible for, uh, for how to build a long term memory about user preferences, right? Like, I don't know, just to choose a random example, I'm, I'm, I'm working... let's say I'm building a company that's working on a, sort of, consumer travel assistant, right? Like, I should just be able to tell that thing, "Hey," like, "I, I freaking hate..." This is a true story by the way, "I hate aisle seats because once someone dropped a suitcase on my head on a flight and I got a concussion. Never book me an aisle seat again." Like, that kind of, like, long term memory, I think application providers should be able to handle as part of a bigger system.

   8. HSHarry Stebbings23:16

      I mean, I, I love that in terms of also, uh, LLMs not being products in themselves.

   9. DLDavid Luan23:21

      Yeah.

   10. HSHarry Stebbings23:21

       You mentioned five to seven cool providers winning. What will separate those that win versus those that don't? Is it purely a game of resources and cash?

   11. DLDavid Luan23:30

       I think it's a game of how much you existentially need to win. I think every cloud provider, um... sorry, every tier one cloud provider existentially needs to win here, right? Like, let's just look at, like... let's look at the dynamics involved. It's, it's one where, you know, as these models get smarter and smarter, they kind of become the base computing primitive. Like, today, the base computing primitive is, like, nodes on EC2, or, like, uh, storage, right? But in the future when more and more software is just, like, the logic of software is actually just handled by a, by an LLM, nobody cares anymore about what the base computing primitive is. All you need to do is access these models and compose these models to go solve things for customers. So then whoever controls the model layer controls, uh, all of the underlying compute.And so, like, right now what's happening, right, is, like, is, like, um, if you don't have an offering here then, uh, that is state of the art, then you're just gonna be cut out of this particular game. I think this is also actually an area where I think it's really important for companies like NVIDIA to go up the stack, right? Like NVIDIA clearly killing it right now on, on, on chips. But what's happening is, every one of the major clouds is working on, um, uh, and every major LM provider is working on a strategy to, to, um, have their in-house chips, because that way they have better margins. And so then at the end of the day, if you're like a developer or you're like an end user talking to ChatGPT and from one of them different providers, do you care whether the backend is an NVIDIA chip or an AMD chip or an in-house chip from Google, right? You, you don't really care. And so therefore, like, like there's like a really key point of the interface that the LLM gives you tremendous leverage on everything downstream.

   12. HSHarry Stebbings25:15

       So do you think we'll see the ownership of the vertical stack? We saw Apple talk about their own chips being a prominent part of their new releases. Do you think we'll see NVIDIA, you know, really move into the model layer with prominence, but also the model layer move into the chip layer with prominence and both try and eat each other's lunch from different ends of the spectrum?

   13. DLDavid Luan25:32

       That's my expectation. To me, like what's interesting about AI from a business side, right, is like it forces the question of like what companies or offerings are gonna be bundled or, or integrated, and which ones are gonna get unbundled. And I actually think that there's gonna be a really strong vertical integration pressure between model builders and chip makers.

   14. HSHarry Stebbings25:51

       Can you just unpack that for me? Keep going with that thought.

   15. DLDavid Luan25:53

       I love the topic of chips. We can stay here for a while. It's just, it's so much fun. It's like the most interesting thing, uh, uh, that's happened in some time for that industry. Like the, um, uh, so, so I- we were just talking a minute ago about how important it is, right, for, um, uh, for model makers to control their chips, because that way, you know, if it really is a scale and resources game, if company A, let's say choose Google with TPU, TPUs are great, right? Um, with TPU has a 20% cost advantage compared to company B using chip Y, right? Then Google will just be able to just th- the better cost of model training will let them go bigger, let them invest more in post-training tricks like the ones we were talking about earlier and have an advantage. And so then company, company B is like gonna be really pressured to go find some way to go do that, go, go do that themselves. Similarly, if you're a chip maker, um, if, uh, if you become like it's just too easy to be com- to be commoditized by these in-house efforts if you don't also own something at the model layer.

   16. HSHarry Stebbings26:57

       Is it easy, like when you think about NVIDIA and what they do, is it easy for them to be commoditized? Is there not such sophistication that actually, uh, it's incredibly hard for these people to move into the chip layer and take that lunch?

   17. DLDavid Luan27:11

       Okay. I actually think we're saying the same thing. It is incredibly hard. Like NVIDIA is killing it. It is incredibly hard. But it is possible. And if the economic returns are high enough, people will do it, right? So I just think Google TPU is a great example. Like the, the, um, um, I am, I'm a, I'm a massive NVIDIA fanboy. Jensen is incredible, like brilliant guy. Uh, I, I think, um, and I think NVIDIA has like executed so well here. I think we also have to give props though to like, I think the TPU team when I was at Google was like sub 500 people. Um, and their budget was a shoestring budget. And yet somehow every generation they taped out, um, quite good chips that were then used to train Gemini and, and PaLM and are used by third parties now and all of that stuff. And like there is such a strong will to ensure that Google has its own first-party chip, um, that I think that's like the, the, the counter-example to like the, uh, uh, the perpetual chip dominance.

   18. HSHarry Stebbings28:07

       I had someone say to me that actually Apple are the kind of the, the dark horse in the race, 'cause they own obviously the consumer and the end device, and they can actually run models offline on everyone's device without reliance on anyone.

   19. DLDavid Luan28:21

       Yeah.

   20. HSHarry Stebbings28:21

       How do you think and feel about that?

   21. DLDavid Luan28:23

       Thinking about the Apple, uh, think about like the Apple advantages in, in this particular space. I think there's like two areas of like extreme power and leverage that you get in machine learning right now. One is the ability to, to, to run smart models for free at the edge, and the other one is to have the absolute smartest models possible. And so I think Apple has a massive advantage on the former. And when we go think about like what, um, like whether that'll be enough, I think that's a really hard question to think through, because I kinda think about it as like concentric rings of model capability, right? To give some concrete examples, like a, like a one billion parameter model that is otherwise trained to be state of the art kinda has like this set of capabilities that it's like perfect at, and then it's like kinda okay at the next rung up. So like maybe the very minimal set of pers- of capabilities is like, is like, is this tweet positive or negative, right? Like you don't need, you don't need GPT-10 to go tell you whether this tweet is positive or negative. Like a pretty small model can do a perfect job at that. And so things like that will always run on the edge. And so then if you're a giant frontier model provider, like you're just not gonna be able to monetize these like tiny skills that, um, are just gonna live at the edge. But then con- conversely, like, um, a billion parameter model is probably for some time not going to solve like be able to like, uh, uh, create a 3D part for me for my car, right? That's probably gonna be the like GPT-10 problem. And so I think as a result, I think Apple is just going to completely crush at everything that looks like something that's really private, something that's, that's, um, fine-tuned on your own particular data, but doesn't require massive reasoning capability, and that will all run at the edge.

   22. HSHarry Stebbings30:03

       Can I ask you, I was quite shocked by Apple's, uh, partnership with OpenAI in terms of the looseness that they tied to it. They continuously said, "Oh, but we'll actually maintain relationships with others," and they very much left the door open to switching between different providers. I almost thought it was a net negative when I heard it. How did you ... I'm just intrigued. How did you interpret that when you heard it?

   23. DLDavid Luan30:26

       I am extremely impressed with OpenAI, I think in terms of their technical, their technical delivery. I think the degree to which GPT-4o was like, I think relatively underhyped, relative to what I think the true scientific improvements have been in that model, um, is, is, is, it is a pretty big gap. Like, I think, like, like, we're, we're moving towards a world where we're gonna be training these, like, universal models that take any input in, right? Audio, text, video, uh, you name it, and then generate any, any output out. And all of humanity's knowledge will be encoded in one of these models, and GPT-4o is a much bigger step towards that than people realize. Um, so I think that Apple cutting that deal with OpenAI, I, I mean, of course I'm not privy to what actually happened, but I think at least part of it is a recognition that I think OpenAI's on a different trajectory compared to others on actual model progress. But at the same time, it also really strongly hints at a commoditized future. Like, to the same extent today, um, as a consumer, I no longer care whether my computer, my desktop at home is powered by an AMD or Intel CPU. I think trying to create a way in which Apple owns the interface and Apple owns the end customer, and then the, like, big brain LLM smarts is just, like, one hot swappable thing, is brilliant for them.

   24. HSHarry Stebbings31:46

       What do you think the end state is for the foundational model, uh, before we move on to the application layer? Like, do, do they just get bought, you know, you've got a couple of cool ones which is really anthropic, Mistral, who have raised a lot of money, uh, billions, and bluntly will not have the resources to compete continuously in the tens of billions of dollars needed. What happens to this layer?

   25. DLDavid Luan32:09

       I think what happens is, um, all of the tier one clouds will have, um, their own effort that will do well, because it has to do well. Um, and they will do whatever it takes to go ensure that they have the capital, and data flywheel, and talent to go do that. Then I think for the independent companies, and I would say that Adept is very different, because we're, w- what we do is we sell an actual end user facing agent to enterprises, which is a very different business model than selling models to, uh, developers. But companies that sell models to developers will either need to effectively be the, like, first party effort of one of these big clouds, or they have a short window between now and commoditization to build such a big economic flywheel that they can afford to stay independent.

   26. HSHarry Stebbings32:56

       How could you build such a big economic flywheel, just so I'm understanding? Like, an amazing enterprise go-to market that generates five billion of free cash flow?

   27. DLDavid Luan33:04

       Yeah. I think it would have to looks like something like that, um, and I think that, um, I think that right now, that's why I think, um, of the companies, of the independent foundation model companies, besides Adept, I'm more excited about places like OpenAI because they have ChatGPT to help do that. Um, uh, whereas I think if you're a pure play model seller, um, I, I think

8. 33:26 – 35:53

   Adept’s Focus for Vertical Integration for AI Agents

   1. DLDavid Luan33:26

      it's very difficult.

   2. HSHarry Stebbings33:27

      Maybe it's the pain and the lack of anesthetic, or maybe it's just my naivety-

   3. DLDavid Luan33:31

      Sure.

   4. HSHarry Stebbings33:32

      ... but I'm not, not afraid to ask this one. Would you say Adept is a foundation model company, or would you say that actually you are not? Uh, how do you think about the positioning?

   5. DLDavid Luan33:41

      We're actually just focused, we're really, really focused on this particular problem that we're trying to solve. Like, we're trying to build an AI agent that you can delegate arbitrary work tasks to, right? And so then everything we do stems from that. So, what we are not doing is we are not trying to just train foundation models to sell them to other people. What we're doing is we're building, like, a very vertically integrated stack. Going back to our previous discussion of where will vertical integration happen versus not, I do think that in the agent space, it's extremely important that you own the entire stack from, "What is the end user interface?" I think, like, as we were talking about the Apple example earlier, owning the interface gives you tremendous leverage in this era of AI, to, "How do you, uh, make agents that are reliable enough to be used at work?" All the way down to, "What needs to happen to the foundation modeling layer to enable this whole end-to-end system to be maximally performant?" That's what we do, is this vertical slice.

   6. HSHarry Stebbings34:31

      How do you think about the variation of agent requirements based on a per-industry basis?

   7. DLDavid Luan34:36

      Uh, that is-

   8. HSHarry Stebbings34:36

      You know, if you

   9. DLDavid Luan34:37

      Mm-hmm.

   10. HSHarry Stebbings34:37

       Do you know what I mean? Like-

   11. DLDavid Luan34:38

       Yes.

   12. HSHarry Stebbings34:38

       ... it's so varying, it's completely different.

   13. DLDavid Luan34:40

       Yes. And that's what our advantage is. That's what our advantage is, is like, is like, you know, we get this question all the time, right, with Adept trying to build, build something that lets anybody, like, like, we wanna be the system of record for workflows in enterprises. Like, any employee at any l- large company should be able to teach Adept, "Hey, like, here's how I do this particular thing," right? Like, "Here's how I handle, um, here's how, here's how I handle fetching all the data for an insurance claim," right? And they should just be able to show Adept that, and then Adept should be able to do it for them. And, like-

   14. HSHarry Stebbings35:07

       Mm-hmm.

   15. DLDavid Luan35:07

       ... that generalization, all of those edge cases and variability, is why the only way to solve that is to have vertical integration of model with use case. And it's also why I think we'll do better than companies that are just focused on a vertical, like, a particular narrow problem. Because every u- like, uh, I was talking to, um, Parag, who used to be the CEO of Twitter, we were just hanging out the other day, and he's like, "Dude, every enterprise workflow is an edge case." And he's absolutely right, and that's why you need to control everything.

   16. HSHarry Stebbings35:36

       What, what does, what does he mean by that? Can you unpack that for me?

   17. DLDavid Luan35:39

       I mean, just even looking at something as simple as, like, "I wanna add a new lead to Salesforce," right? You go, "Let's go outside and find ten different companies who all use Salesforce, and look at how they've got it configured," and it all looks completely different

9. 35:53 – 40:24

   The Distinction Between RPA & Agents

   1. DLDavid Luan35:53

      from each other.

   2. HSHarry Stebbings35:53

      Yeah, I get that. Can I ask you a dumb question again? Is this not what RPA was always meant to be? You know, I'm friends with, uh, Daniel Dynes from UiPath.

   3. DLDavid Luan36:02

      Yeah.

   4. HSHarry Stebbings36:02

      Wonderful dude.

   5. DLDavid Luan36:03

      Yeah.

   6. HSHarry Stebbings36:04

      I always swear this is what RPA was. So can you help me understand the distinction between, like, traditional RPA, which is, you know-

   7. DLDavid Luan36:10

      Yeah.

   8. HSHarry Stebbings36:10

      ... what we've seen with UiPath, and this new era of agent that we see today?

   9. DLDavid Luan36:14

      Yeah, totally. I mean, this is, uh, this is, this is really, this is, this is a good question. It's actually a question used to cause me a lot of heartburn, because I found it so hard to explain to people why agents were gonna be different than RPA. Best analogy I've got is, um, RPA is very useful for high volume tasks that always look the same.So, um, example, uh, the analogy that I would give would be, like RPA is a little bit like, um, little bit like, you know, when you go to a factory floor and there's robots roaming around everywhere?

   10. HSHarry Stebbings36:42

       Yeah.

   11. DLDavid Luan36:42

       What those robots do is, there's like a literally yellow line painted on the floor, and the robots like follow that line. They go from cell to cell, and station to station, and they pick up stuff. But what agents are, are agents are meant to be, agents are meant to be constantly thinking and reevaluating and planning at every step to solve your goal, and it's much more like full self-driving. The difference in utility between those two things is, is, is fairly large. Of course, there's many areas where you don't want, uh, something that can have variability, and therefore you should use RPA. But I think the majority, like I just think in five to ten years, people are gonna use their computers by giving them high level goals, right?

   12. HSHarry Stebbings37:19

       Well, will the largest enterprises in the world run RPA and agent-based systems alongside each other?

   13. DLDavid Luan37:25

       I think so, yeah.

   14. HSHarry Stebbings37:26

       Why are RPA players not best placed to provide an agent solution to existing customers?

   15. DLDavid Luan37:32

       I think it's just really fundamentally disruptive to their business model. Like, um, the way that, um, the way that a big corporation, um, uses UiPath, right, is like, um, is like there's a, there's a, there's a big, like big plans around a process transformation that needs to be done. Sometimes like a, like a, like an Accenture or something comes in and then maps out what the processes are like, sometimes with a process discovery, um, uh, thing. And then RPA engineers go and build those workflows, and then six to nine months later you hit play on this thing that then automates some invoice processing every night or something like that. Right? Like, this new model of like you just put an agent in there and the agent observes what the, the end user does to go do that job, and then that becomes like a thing that you can then just invoke with natural language. It's like really disruptive to the business model, um, and I think that, um, I think that the best way to, uh, to, to run circles around incumbents is to do something as a different business model than what they have.

   16. HSHarry Stebbings38:29

       How is your business model different?

   17. DLDavid Luan38:31

       The way that we're doing things is we are addressing, um, we are addressing use cases initially that are really, uh, that are really painful that get our foot in the door, but we're really focused on how do we make the end user be able to teach it any new capability. Like, I should be able to dump in my, like to Adept, I should be able to dump in my like standard operating procedure for this new thing my team does. Or I should be able to show Adept like ten times and give it corrections on how I, um, on, on, um, on how I enroll a, a, a new nurse into a healthcare portal, uh, in, in the US, and then the model should be able to do that for me. And, um, and basically like we're working on something that's ultimately very self-serve over time.

   18. HSHarry Stebbings39:09

       Everyone speaks about kind of we're gonna sell the work and not the tools, is like the, the hottest statement. Um, and the end of price per seat, and we're all moving to a consumption-based pricing model. Do you agree with that statement? Do you think we're all slightly over-emphasizing like the end of price per seat, and how do you feel about this kind of fundamental shift in, in business model and pricing that AI could bring about?

   19. DLDavid Luan39:34

       I think in places we're definitely gonna see that become true, but I actually think in knowledge work that's, I think in knowledge work the most valuable things to do will not be priced that way. And, and, and here's why. Um, I think that like the definition of price per work assumes repetitiveness, commoditization, cookie cutter, you know, like no creativity, right? Um, but the way that, like I think what these AI systems are gonna do, especially AI agents are gonna do, we are basically gonna like give people the ability to go, to go do new things and have way more leverage on their time and like give them more opportunities to be creative. And so then ultimately what we're building is like a co-pilot or a teammate, and co-pilots and teammates don't charge you a price per work. They ch- like, like you really pay them based on their ability to augment your ability to go, to go do new things,

10. 40:24 – 42:46

    The Co-pilot Approach: Incumbent Strategy or Innovation Catalyst

    1. DLDavid Luan40:24

       right?

    2. HSHarry Stebbings40:24

       You mentioned that that kind of co-pilot approach. I had the guests on the show say that the co-pilot, and I think it was Miles Grimshaw at Benchmark, who will now Thrive, said actually co-pilots is an incumbent strategy, is leveraging existing distribution, and it's an incumbent strategy. Is that fair, or do you think actually it's not giving due credit to the co-pilot approach?

    3. DLDavid Luan40:43

       I think both of these two things can be true. Like I think it's co-pilots are a great incumbent strategy because it lets them morph their existing software business model when something like kind of looks the same while getting in on the AI thing. Um, but, but even separately from that, I just think like where are these systems going to be most useful? Like, I just feel like everybody in this field has this vision, right, that like AI is gonna take all jobs. The like pricing by work thing is just a corollary if AI is gonna take all jobs, right? Because then it's like, all right, maybe you price by work on invoices, and then next month you price by work on like consulting decks, and then before you know it you price by work on like being AI CEO of like David Co. or something like that, right? Like, that's not, I don't think that's how this is gonna play out. I think the way this is gonna play out is, is that what we're gonna have is humans fundamentally be the drivers of these agentic systems that, um, like, that like basically give everybody a tremendous amount of leverage on their own creativity. And like, how can that be built without a co-pilot style approach, is like, is like my question.

    4. HSHarry Stebbings41:45

       What does that do to the org structures of teams, David, do you think? Does this mean much, much smaller companies? Does this me- wh- how do you think that actually plays out?

    5. DLDavid Luan41:53

       I think the main way it plays out is, um... Actually this is something I'm gonna steal from, um, our angel investor, Scott Belsky, who, uh, ha- has just thought about this so much. Um, he always calls it like this collapsing the talent stack thing, and the idea is basically that, um, uh... Uh, or this is my interpretation on it, like projects and teams where the same person is simultaneously the PM and the designer and/or the engineer or the go-to-market person or the marketer or whatever, like the more that like those different skill sets are smooshed in the same person, the faster that thing moves and the more effective the thing becomes. So I think what it's gonna do is it's gonna make humans at work much more like generalists, and it's gonna have, like cause us to create larger and larger...Oh, sorry, uh, like, uh, uh, or giving people sort of like larger and larger scope over various different, like, areas or different functions today, while they ultimately supervise like a cohort of like AI co-pilots that are the specialists.

11. 42:46 – 46:53

    Enterprise AI Adoption Budgets: Experimental vs. Core

    1. DLDavid Luan42:46

    2. HSHarry Stebbings42:46

       One thing I, I did want to touch on was when we think about the rollout, I think we overestimate enterprise adoption. One, are we still in experimental enterprise adoption budgets? Or do you think we are moving into core enterprise adoption budgets?

    3. DLDavid Luan43:00

       I think, um, you know, Harry, we should rewatch this podcast in 10 years, uh, and, and see, and see how we feel. Uh, but I think, like, you know, when we talk about AI, like, AI is so freaking broad. It's a little bit like us asking, um, you know, maybe in the early days of the internet, um, like, uh, the, the, a generalized thing about the internet as well. Like, it's just I think it, I think there are some use cases that are clearly hitting PMF within an enterprise. But for the most part, like, just when we go to enterprises to go, to go, to go sell them stuff, like, they've got so much stuff that's still on-prem. They've still got workflows running on mainframes. And it's, it's 2024, uh, and so I think, like, even if technologies like cloud, which we probably look at as from a startup lens as being so freaking mature, still (laughs) doesn't have full, full adoption in enterprises is, like, I, I think that stuff is really, is, is really interesting. And I think as a result, like, we're gonna be on this adoption curve for enterprise AI for a very, very long time.

    4. HSHarry Stebbings43:54

       So we are still in the experimental budget phase?

    5. DLDavid Luan43:56

       I think it's, I think the majority of it is, like, extremely experimental. And one of the things we do, for example, is we really try to not sign deals that are coming out of experiment budget 'cause we want quality revenue basically.

    6. HSHarry Stebbings44:06

       D- do you think we grossly overestimate enterprise adoption in the short term and underestimate it in the long term?

    7. DLDavid Luan44:12

       Yeah. I think it's true for, I think it's true for most new technology, but definitely, but definitely here.

    8. HSHarry Stebbings44:17

       I, I lo- Alex Schultz, the CM at Meta, uh, told me and recently released a piece about kind of hype cycles in new technology, and actually he states his kind of concern that AI will replicate autonomous driving in the way that we got so excited about 10 years ago. Everyone's gonna be unemployed, eight million truck drivers. But my question to you is, are we gonna see that similar plateauing where for 10 years actually kind of autonomous cars didn't feel like it was progressing? Do you... How do you think about that?

    9. DLDavid Luan44:48

       I've never worked in self-driving.

    10. HSHarry Stebbings44:50

        Mm-hmm.

    11. DLDavid Luan44:50

        Um, but if I, I'm gonna, I'm gonna apply a mental model and you tell me, you, you, you tell me if it, if it lines up. I feel like in self-driving what happened was there was an a-ha moment where, you know, you could get the thing to work at all. And then you're like, "Okay, well, now it works 60% of the time. How do we get this to 99.99999% of the time?" And every day you show up to work and you just play Whac-A-Mole on what's not working, uh, and you just, like, hope and pray that this converges to that, like, 99.99999 thing. That's not true for AI right now. That's, oh, sorry, that's not true for, uh, specifically what I'm about to say is only applicable to building smarter and smarter models and agentic systems that ultimately help you do work. That's, that's the thing that I'm trying to talk about. Like, for building that, that's not how, that, that's not how the, the underlying dynamics are right now. Like, every day we go to work and there's, like, actually brand new scientific things we wanna try that just dramatically improve the performance of a model. And, um, some of those bets don't work and some of those bets really work. Um, I think, like, the reasoning that we talked about earlier is an example of one. Think another example of one is, like, this, like, universal multimodality that GPT-4O is. Like, those, those breakthroughs are, like, visible, and as a result, I think the ca- the, what's gonna prevent this for what I think has a hope of preventing this from just being a hype cycle that falls flat like AB, is that those things are, those shoes are yet to drop. And as they do, the capabilities of these models are gonna continue to improve, and that on top of that they're already... It does not a technology that you don't, that you, like, need to get to that level of reliability before it can be deployed. It's already deployed today.

    12. HSHarry Stebbings46:27

        And speaking of deployment, I think, and the kind of enterprise adoption, I tweeted actually that, um, you would see AI services companies, uh, really people who help in the implementation of AI in large enterprises, be bigger than the model providers themselves in terms of revenue, uh, and we've seen that actually come out as being true.

    13. DLDavid Luan46:46

        Yeah.

    14. HSHarry Stebbings46:46

        Um, I had some famous people call me an idiot, which actually made me quite pleased-

    15. DLDavid Luan46:51

        (laughs)

    16. HSHarry Stebbings46:51

        ... when those revenue numbers were revealed.

12. 46:53 – 49:32

    AI Services Providers vs. Actual Providers

    1. DLDavid Luan46:53

       Yeah.

    2. HSHarry Stebbings46:53

       Uh, how do you feel about implementation providers, AI services providers being bigger than the actual providers in the next five years? Do you think that it's right that the biggest players to come out of this cycle will be the AI services providers?

    3. DLDavid Luan47:06

       don't think so because I think the third bucket of, the third bucket of, like, economic upside is still early, and I think that bucket is, like, um, is the companies that then turn w- the use cases that have product market fit into repeatable products. Right now, right, imagine your, your, your very large company X, right? And you need capability Y, and you've got the base model over here, right, that's pretty darn smart, GPT-4 or, or, or Gemini or whatever, right? Um, and there's a giant gulf in the middle. Like, the first... I feel like in, in, um, in, in every one of these cases, the first people to go fill that gulf are, like, sort of consulting-y service providers, right? But then the moment that gulf starts getting filled and you start seeing, "Ah, okay, like, this is the really useful thing for an enterprise," then people just go productize that thing and then that becomes a startup. And so then that becomes eventually a company that's a conduit between the base intelligence and the customer. So today that might be true for services, but I feel like a lot of these things will be turned into generalizable products. And then when they do, those companies will then be the real economic winners.

    4. HSHarry Stebbings48:10

       One other concern that I have, I have two other concerns, but I want to touch on both of them with you 'cause, uh, they keep me up at night and I have enough wrinkles, David. Uh, one of my-

    5. DLDavid Luan48:18

       I love that you also write them down as well, so it's not like

    6. NANarrator48:20

       (laughs)

    7. HSHarry Stebbings48:20

       Do you know what? I, I, it's important my memory goes, okay? It's, uh, the challenge of being old. Uh, but, uh, regulation, Europe, you know, specializes in it. One concern I have is that we could regulate ourselves into oblivion around data usage, data collection, and actually we don't see the progression of these models and AI in a way that we want to.How do you feel about that? How likely is that? What would you like to see happen in regulator environments? Help me sleep at night.

    8. DLDavid Luan48:46

       I think my main concern right now is actually one of regulatory capture. The, in the same vein we were talking about earlier about how there will be only a few sort of frontier model companies that can exist steady state, I think, um, I think the, the move to go pull up the ladder behind them is already beginning. Um, and I think that, um, uh, lawmakers don't really understand this technology at all, and so their default instinct is sort of listen to the most credible source, and usually those credible sources have, have, um, alternate, uh, uh, sort of ulterior motives here.

    9. HSHarry Stebbings49:17

       So what happens in that case?

    10. DLDavid Luan49:19

        Becomes harder for the general field to go build on open source. It'll become harder for new companies to get started that have new AI ideas they wanna go train and scale up. Um, I, I think what really happens is just another concentration

13. 49:32 – 54:18

    Open vs. Closed AI Systems for Crucial Decision Making

    1. DLDavid Luan49:32

       of power moment.

    2. HSHarry Stebbings49:33

       Ironic, given Lina Khan's (laughs) position. Um, ha, okay. That is a concern for me. You mentioned kind of the ability to build on open. The other concern that I have is actually, you know, we had Alex Wang on the show, and he said, I think, that, you know ... I th- what was his statement? His statement was, "AI is more powerful than nuclear weapons, and in the hands of the wrong people, especially AGI, it could be probably the most lethal weapon used ever, and for that reason, we should probably have more closed systems." How do you think about that debate of open versus closed and whether for some of the most crucial decision-making AI systems it should be closed?

    3. DLDavid Luan50:10

       I think two things. One, I think the broader set of concerns about use and misuse and safety are, uh, are, are extremely important, and I think that, like, what, what was a good thing about all of this is that people are having these discussions more, more openly, which, um, which I, I really strongly appreciate. I think with a lot of these systems you can already see, like, clearer ways to go, to go misuse them, right? Like, spin up a bunch of, uh, spin up a bunch of servers, take the best code model you have out there, use th- like, use them to go try to, like, find vulnerabilities in software systems. Like, that's, like, if that's already happening, it's gonna really start kicking into gear. So, like, things like that I think make me very concerned. At the same time, I think that, like, um, AGI is just a really difficult thing to reason about, because the way that many people define it is, like, almost defining it as an infinity, a- and, like, reasoning about infinity is, is really hard because you multiply infinity by 0.00000001%, and that's still infinity. Um, and so, uh, and so I think it's a very brain-breaking, uh, thing. And so I think a better way to go look at it is to look at the path dependence, like how will this technology actually be developed in the next five years? And I think in next five years, open will always lag closed, and because open will always lag closed because open just has fewer resources behind them and fewer incentives for people to go, uh, uh, to go make things to be open as these things become more and more expensive. I view open really as a way for the rest of the field to keep up with the biggest incumbents, um, and therefore I think it's actually pretty darn important.

    4. HSHarry Stebbings51:40

       Uh, you mentioned kind of AGI being kind of infinity there in people's minds. You said bu- before to me that the last step is human/computer interaction, and that's the last ingredient to AGI. Before we do a quick-fire, what did you mean by that? I didn't get that one either.

    5. DLDavid Luan51:55

       I personally find a world in which sort of increasingly generally intelligent systems run around with their own agency and goals, uh, and, uh, and, um, not involve, um, not involved what, what, what humans most care about to be not a world that I really wanna live in. And this goes back to the, what you were saying about, like, selling AI by work versus as a, as a, as a software tool, right? Like, I, I would much rather live in a world where we have sort of these, like, like, AI teammates, um, and assistants that we interact with instead. And then I think the question becomes, um, the, then the question becomes, how do you find the right interface between smarter and smarter AI systems and people? And how that interface is defined actually changes a lot about what training data you collect, how can humans align these systems towards the preferences of what humans want. Also, ultimately, like, how these models are even built and what their architectures are. And so in a weird way, like, the way the field is moving is, let's make models smarter, and then let's make use cases smarter, and then let's go put them in people's hands, and then, like, let's figure out what those means for people. Like, it's kind of this waterfall sequential method, which I don't think is a very good way to develop the technology. I think we should start back from ultimately how do, how should humans use these things and then, um, and then, like, create the whole solution, um, end to end that way. And so that's why the HCI problem, like, people just aren't spending enough time thinking about. Like, chat is obviously not, n- is no- obviously not it.

    6. HSHarry Stebbings53:16

       You mentioned that kind of, uh, the wrong way to think about it. What questions do you think people are not asking enough that they should be asking more?

    7. DLDavid Luan53:24

       As these models get smarter and smarter and they sort of know more and more about the world and have more and more ability to do things in the world, um, how do you interact with them? How do you supervise them? How do you give them corrections and teach them to be more aligned with what you want? I think questions like that from, like, a, from, like, a, um-

    8. HSHarry Stebbings53:43

       I'm sorry. To be blunt, is that not-

    9. DLDavid Luan53:44

       Yeah.

    10. HSHarry Stebbings53:44

        ... just basic prompting?

    11. DLDavid Luan53:46

        You know, when you go work with a coworker, right, you don't just, uh, you don't just, you don't just talk back and forth. You, like, actually, like, share the same canvas. You'll go use a whiteboard. You'll go, like, uh, maybe look at the same thing on the computer together and then go try to solve the problem together. Like, human-human interaction is, like, way, way, way, way richer than, like, writing a set of instructions, uh, and then being like, "Do this thing," and then when it fails, you come back, and you tweak the instructions a little bit. Like, that's just not a very, like, especially as these systems become more and more smart, that's just a really bad way to go interact with them, and I, we're just not putting enough time to figuring out the right way in which we interact

14. 54:18 – 58:17

    Quick-Fire Round

    1. DLDavid Luan54:18

       with these things.

    2. HSHarry Stebbings54:18

       Listen, I've peppered you with questions. I wanna move into a quick-fire, so I-

    3. DLDavid Luan54:21

       Yeah.

    4. HSHarry Stebbings54:21

       ... say a short statement. You give me your immediate thoughts. Does that sound okay?

    5. DLDavid Luan54:25

       Yeah. Sounds great.

    6. HSHarry Stebbings54:27

       So what have you changed your mind on most in the last 12 months?

    7. DLDavid Luan54:31

       It's actually a little bit what we were talking about earlier. I actually think agents and chatbots are gonna speciate and turn into two different products.

    8. HSHarry Stebbings54:36

       How does that look?

    9. DLDavid Luan54:38

       I think it's gonna look like you're gonna have these, like, rich... You're gonna have these, like, uh... You're gonna have these, like, rich interactions with these increasingly smart systems that can do things on your behalf, and then you're gonna go have other systems that you talk to for, like, therapeutic or fun use cases.

    10. HSHarry Stebbings54:53

        What's the biggest misconception people have today of the next 10 years of AI?

    11. DLDavid Luan54:58

        Biggest misconception is that this is just gonna be something that, at every step, takes another human capability and fully au- and fully automates it. Like, the implicit goal of AGI right now is replace human work, but like so much of human work, like, I just don't think will be neatly captured by AI. And instead, it's gonna be... It's gonna be something that, like... It's just, like, A- AI will be a tool to level up human intelligence.

    12. HSHarry Stebbings55:21

        What's your vision for the future of agents? If everything goes to plan, agents in five years time are... Agents in five years time. I mean, uh, it's kinda gonna be like a non-invasive, like, brain-computer interface basically. I think that's what an agent will be. Like, all of us are gonna be up-leveled. We're going to, you know. It's gonna, it's gonna feel like the same transition from, like, DOS/command line to the GUI but from GUI to agents. We're gonna interact with them at a, at a high level, at the level of goals. And they're basically just gonna let us, I think... They're gonna let us, like, have like basically, like, new kinds of thoughts. Like, the ability to go, to go, like, like, reason at one level of abstraction beyond what we all do today. You're writing a pre-mortem on why that does not happen, why agents are not that in five years time. What is the most probable reason why that does not happen?

    13. DLDavid Luan56:11

        I think one way it won't happen is that incumbents have so much... Like, fundamentally, like, agents are a reframing of where, of like how software is bundled, right? Like, today, we've, we, we, we, we, we bundle software in these, like, functional ways, right? Like, you've got, you've got Notion or Google Docs for, for your, for your docs, and then you've got Salesforce for sales, and then you've got, like... Um, you've got, like, Workday for HR and all of this stuff, right? But the work that we do fundamentally spans all these different domains, and an agent sh- should bridge those domains. Otherwise, you can't become a s- a, a, uh... You can't become a higher level of, uh, thing. So, if we're locked into, like, end walled gardens by incumbents, then that vision will not happen.

    14. HSHarry Stebbings56:54

        This sounds super awful VC mindset, but when you look at, like, UiPath today being like a $67 billion company, do you not think there's bigger opportunities to go after? Like, that took 17 year, maybe more, 19 years and is a $7 billion company with billions in revenue. You're a super smart, super ambitious guy. That feels like a lot and a long time for actually a value capture that, if done well, is still like, "Mm."

    15. DLDavid Luan57:24

        Well, I think the, uh, the question is what percentage of work done today is addressable by RPA? It's very little, right? But what's the percentage of work done today that's addressable by agents? It's like 1,000X that, maybe. 10,000X that? I don't know. Something in that order of magnitude. It's just a... It's a very different market. It's like a... It's like saying, "Should we work on self-driving," uh, uh, "when self-driving didn't exist?" and then looking at the market for, um, those, uh, autonomous rovers and warehouses.

    16. HSHarry Stebbings57:54

        David, honestly, I, I... Shows like this are why I love doing this. So, thank you so much for being-

    17. DLDavid Luan57:59

        Thank you.

    18. HSHarry Stebbings58:00

        ... so brilliant. I so appreciate you putting up with my lack of smiles. Uh, but this has been fantastic.

    19. DLDavid Luan58:05

        No, you're, you're doing awesome. The fact that you're able to do this after a wisdom tooth removal is insane. I had so much fun. I thought you asked great questions across business and tech. Uh, and, uh, yeah, I'm excited to see how this all, how this all plays out from here.

Episode duration: 58:17