999: What's Left to Build When Software Is Free, with Chip Huyen

23:14OK, if the Skype query is like this, then maybe you can use the data from a week ago. If not, it can. Yeah, it's just like it is very, it's still very much manual, like the way I see how things are being defined. But anyway, so by the way, when I mentioned the system prompts, like the last time I looked at the system prompt of this model was about four months ago. So things may have changed. So I'm not sure how they are doing web search nowadays.

23:41Jon Krohn:Yeah, I'm sure it's a fast evolving space, especially if it's consuming tons of tokens. All of these frontier labs are trying to keep their compute down as much as they can to get the same quality of result at a lower cost. So there's got to be optimizations going on all the time. I think so. It's looking at how people are spending money on their agent. So I think there's a part of the actual reasoning, actually a lot cheaper, but a lot of them are more like tool use tokens, right? Like web search, file search, like one big category of tool use. The other category of tool use is like I see like productivity.

24:20Tool use, like when you connect it to like Gmail, Slack, Asana and like they're going to have to like do like tool calls and then analyze the results of my tool calls. And then of course it's also always like valuations. Yeah, some people were telling me that they were paying like 20 % of their token costs. Some neighbors? Oh, no, some company.

24:43Jon Krohn:Oh, some company. Some neighbor. You know, my neighbor is not into AI now. Oh, really? In San Francisco, you have neighbors that aren't into AI. That's amazing. Yeah. Fantastic. I'm going to fast forward a little bit in your book because earlier in the conversation, we were talking about how interested you are in inference. Yeah. So I want to give you some time on that and explain to our listeners why that's so exciting for you. It gets its own chapter in the book. And for our folks who are watching their token bills balloon, what are the highest leverage moves that they can do with their own LLMs in production in order to be more efficient?

25:21Honestly, I feel like inference. I'm interested in it because I'm a nerd. But I think for a lot of people, I think if you don't control the model, then you can't really do much to optimize that. Yeah.

25:33Jon Krohn:Yeah, if it's your own, I guess if it's your own, then you're not so worried about token costs themselves. I guess you're maybe just worried about how often you have the model up or how many GPUs you need to have running. Maybe we're starting to get into something that's really for people that have a large number of models on it, a large number of GPUs on it anytime doing inference. I think one thing people can do for token cost if they're worried about is that I try to look into what owns the usage and what kind of usage that can be offloaded to smaller models. Not everything needs to be sold by very big, very massive, expensive models.

26:16So if something that is small, like, for example, like going back to the classic example of customer support, like if you can see a certain category of queries that can be answered just by mapping to FAQ, then maybe you don't need to send them to big models. Yeah.

26:35Jon Krohn:So bringing that back again to the idea of great systems design for your AI system and bridging that with your book, which actually we're going to move on to, I just have one last question for you related to your books, and then we're going to get onto what you're excited about right now. But your final chapter in your AI engineering book is about closing the loop with user feedback. And it seems like a good user feedback design is critical to having a great AI system in production. Do you have any thoughts for us on that? So I think feedback, user feedback, serve many purposes, right? One thing is that if you serve people, if you charge people money for any service, you kind of want to know how happy they are about the service, right?

27:27So, of course, you need feedback. Feedback is also very good to uncover the area for improvement, right? So, I think if you look at feedback, you need to understand. So, it has maybe like for this certain demographic, somehow the users are not extremely happy. And you want to understand why. So we try to look into the data as detailed as possible, like slice and dice by users, by use case, by locations, such as you understand different patterns in there. So another feedback is also good for tuning model. So a lot of evaluations to... So user feedback is also one type of feedback, right? But the ultimate goal is so that we have a way to automate.

28:16automatically evaluate our systems in production.

28:20Jon Krohn:Agent feedback. How do our agents feel about our AI system? You really need to ask the agents. Exactly. Yeah, so I think it's like, okay, so let's say you build this very complex and very amazing systems, and your engineers sway on their life that, okay, this is amazing. But you're like, yeah, I trust you, but maybe you should just put something to see whether it's actually working, right? So you kind of want to build an evaluation system to see if the responses make sense. And a lot of people using LM as a judge, right? They use another AI model to look at the response and evaluate. And one thing that is actually really, really hard that I see that a lot of people spend so much time on is how to write the guideline to have the AI models evaluate, like output score that makes sense, right?

29:12So I think we discovered that pretty early on, there were a lot of case studies when somebody was like, okay, we spend 80 % of our development time just to write guideline for our AI judge, like I am as a judge. It's painful. Yeah.

29:26Jon Krohn:You have two chapters in your book on evaluation, in your AI engineering book on evaluation. And in it, you describe things as getting slippery as soon as the outputs that we're evaluating are generative. because it's very difficult to come up with a really rigorous test. Like it used to be the case that you could have tests in software development where you were like, if it isn't character by character, this output, there's a problem. And now obviously with Gen AI, you can't do that anymore because you'd have such a wide range of responses. So that's kind of what we're touching on here, right?

29:58Jon Krohn:You need to spend a lot of time figuring out what your evals are to ensure that you're getting the kind of result that you want. Yeah, I think like bringing back to what you just mentioned, in the past, I saw where we had test-driven development, which is like a very interesting program, extremely useful for a lot of use cases. For AI, I think the simple calling is like evaluations-driven development. When you like, there are only development applications that you can't measure the output for, right? Because I think like there's some questions sometimes when I give talk to a company, I ask people, like, okay, which one is worse?

30:37Like, having a system in production that, okay, like not having an AI system in production or having a system that nobody knows whether it's working or not. Yeah, so I think it's happened quite a couple of times. It was like, okay, we deploy this AI system. And some people was like, okay, like, we save so much money, we make so much money. And I was like, okay, how do you know? And so we're like, we guess. Yeah, so it's tricky. So, yeah, so going back to the guideline for the AI judges, it's painful. So, for example, like, how do you even explain to it, like, okay, this is a good response, this is a bad response, and why is that a good response?

31:17Like, a lot of people, like, we look at something, we can tell that it's bad or it's good, but we don't really want to sit down and write down our reasoning. and but like those um and i think like one of the most tests i have with um a lot of people is it's like okay after you write out the guideline for the model use yourself go and follow that guideline and go through some examples and see like okay like whether you're good right i like or and then i all i give it to the co-worker and like beg the co-workers like just follow it

31:45Jon Krohn:so you're saying that in today's day and age humans can still have taste yeah that is more useful than an LLM. That is something. Wow. So I think it's like, it's important to not just assume that AI can do everything. Like AI can follow instructions, but if the instructions are like trash, it's not going to be good, right? And I think something you talked about earlier in the episode, there's one of the absolutely most important things to having an LLM being able to do what you want or an agent to be able to do what you want as an individual or in an organization and an enterprise, whatever, it's having the right context.

32:24Jon Krohn:And so you might think you could have, you know, your AI system that's doing evals. You can't just trust the outputs because what if, you know, you think that you wrote a great prompt, you think that you have the evals set up properly, but if you don't go through and read them individually, there could be some key piece of missing context, maybe some product requirement that came from your users or the client. And only you as, you know, the stakeholder, this human who's been in all those meetings really knows what's important and can give the final sign off. Yeah. I think that like is, it's very important to like have a good understanding of what your instructions are doing.

33:04So, so a lot of time you don't need to read it line by line, but like just be aware that there's something you need to pay attention Like, for example, like I threw it into like another model and like just analyze for me finding contradictions in my own prompt. And it does that pretty well. And also have explained a lot of weird behavior that I felt. So, so for example, like so, so another applications when I use AI for like to do to do like deep research and generate a summary of the research. and then all the summaries somehow always focus on very specific things very weird and was like why so it was at first like okay is this is the ai models like bias why are you so insistent on like this new york office of this company like one of them and then it's it turned out like in some like weird part of the prompt i give an example of like okay if you if somebody asks about a company find its offices, maybe check whether it's in big city or New York and stuff.

34:03And it totally overindexed on that. So it was like, OK, and so it removed that example. And it works much better. So you just be aware of what you're asking the model to do. And I think it's one of the things that also useful for having good guideline is that you can also use it to synthesize data later on if you ever want to need more, fight more and more data. For sure. For sure.

34:28Jon Krohn:Yeah, that's, I mean, that's what we were talking about. Oh, yeah. I mean, you can be using, we didn't actually talk about synthetic data specifically. So we were talking about, I was talking about earlier in your episode, I'm taking up way too much of your time, but I was talking about how when GPT-4 came out, that was the first time that I could, that I felt confident enough about an LLM that I could be using it to label data. Yeah. But you could also use GPT-4 and any of those kinds of frontier models since to actually be synthesizing not just the labels but the training data as well yeah and you've got to be very careful in doing that because you can end up you might not get the breadth of the sample space that you want so you have to be really careful how you see that training but it can be very effective yeah like i personally don't fight tune models uh anymore i think i did try like early on but then at some point was like okay this model just keep getting a lot better yeah so i spent a lot more of time on like tool use and like web search and stuff to make my things better.

35:27But yeah, so I actually don't use a lot of data. But I do use AI to label a lot of stuff. Because a lot of my applications require AI to label different stuff, categorizing things, taxonomy, and that evaluations, obviously. Yeah.

35:44Jon Krohn:All right. I think it's time to go on to, you know, we've been talking about what you're excited about now. It's not fine tuning. So let's move on to what you're really excited about. It's actually, it's going beyond software. It's going into hardware, into physical systems. And so at CES this year, Jensen declared the chat GPT moment for physical AI is here. Is it? That's what he says. And so I'm here to ask you about that. Is that just hype? Or do you think that this is a really exciting time for you, maybe for our listeners, to think about getting into physical systems? I'm not sure I have the street cred to argue with Jensen at this point.

36:32Jon Krohn:I think you're the next. It's the two of you. You are the two people that the world looks up to the most on this kind of stuff. No, no, that's crazy. But yeah, so something is like, I'm not sure that you have gone through that, but I feel like a lot of my friends and I, we went through this, like we call existential crisis, right? And we're like, holy shit, like what do we do now? AI is coming for our job. So we thought a lot about, not just like what to build, not just like how to build things, but like what to build. Because I do think it's like AI is getting incredibly good at a lot of things.

37:05and building things is actually not, the process of building things is not that hard anymore. So let's say you have a great idea, right? You write a spec for it. You can ask AI to review those specs, improve those specs, and then it inputs the spec into AI and it can create some amazing website applications which is super cool. But then you build something and then what, right? So it has an experience of like, so it has a side project about two, three months ago and I put it online. And it's like, it's not like mind blowing anything. It's like something chill, something like, it was useful for me.

37:42Like I think one of the things I really love about AI nowadays is it just like allows me to build very good micro tool. It makes my life so much easier. Sure. It's very easy to build.

37:51Jon Krohn:Can you share some of your favorite micro tools that you built with us? So the things I did was like good AI list. So like every day you just like crawl, like crawl GitHub and it's fine based on like a bunch of keywords I give it and it finds all the repos related to this keyword. And then it analyzes repos and tells me what's interesting about it. And then it's categorized and also rank them. So it's helping to resurface things that might be relevant to me. So I put it out there. And it got, I think, 300 ,000 views in a week. And the next day, someone emailed me. It's like, oh, I love what you did.

38:30So I used AI to replicate exactly that. and here's that and it was just like i'm not sure how i feel about it like you know where i'm flatter but it was just like is that like copy like yeah i don't know like what is that right so it just made me so realize it's like anything that anything that exists today software can be copy like replicated when there was the leak of the cloud code repo about a month

38:57Jon Krohn:ago at the time of us recording, somebody, obviously, you can't, that's copyrighted, you can't publish exactly the same code, but somebody recreated it in Rust or something, use Claude code to recreate the Claude code code base in Rust and then publish that in GitHub. And so, yeah, it's pretty wild what you can replicate so easily today. Yeah, so I do think that it's like, it's an interesting feeling. Because I feel like, you know, on the one hand, AI mix is, it's like, it's very exciting because it allows me to build anything I want. And not everything, of course, there are some things that's harder to like build than other, right?

39:39I'm not going to be like Google search in like a webcam. But AI is also getting like, if you look at the complexity, there's a level of complexity of tasks that I can do. Like that level of complexity is like going way, way up. It's insane.

39:52Jon Krohn:With the Mythos release. So you know the people at Meter, M-E-T-R? Yeah. And so they have, I'll put a link. It's a very cool organization. Very cool organization. I should have someone from them on the show. If you know anyone, let's talk. And so they, I love, I always, every talk that I do these days, I always have the latest meter chart because they take the latest and greatest frontier model. And then they benchmark it against how well it can replace a human on a software development or a machine learning task. And Mythos broke their evaluations because they didn't have reliable tests that take humans longer than 16 hours.

40:32Jon Krohn:Imagine how long it would take. It was very easy when meters started going and they're benchmarking, okay, let's find some tasks that it takes a human a few seconds. Let's find a task that it takes a human a few minutes or a few hours. but now that they have to be coming up with tests that take 16 hours, 24 hours, it's going to be no time before it's 50 hours, 100 hours. How do you even find and pay humans to do that as a benchmark? That is scary. And exciting. Yeah. So Mythos broke it because it was the previous front, like top performing model was Opus 4.6 in the meter evals. And that was averaging, like it was able to do on average 50 % accuracy on a task that would take a human about eight hours.

41:12Yeah.

41:12Jon Krohn:And then Mythos, they're like, okay, they put the dot at 16 hours, but they also put this disclaimer that like, but we don't actually have good evals past 16 hours. So it might be even much more. Mythos might be able to handle on average a 30-hour task. We just don't even know. It's crazy. Yeah. Wow. Yeah. So yeah, things are improving fast. Yeah. Exactly. So is that what you're making? Yeah. Improving for machines. Yeah. So, yeah, back to the land of the humans. So, it just made me think about, okay, so it was saying that the cost of building software now is approaching zero, right? It's just code generation is cheap and nobody really cares about, like, in the past, people compared how many lines of code you write today, right?

42:01It's quite meaningless nowadays. And, yeah, so the cost of building software is zero. So, what does it mean? like it doesn't mean that the value of building software is also zero right because if the cost of it is zero then obviously um but yeah so so so maybe so think about like okay so what is the next frontier um and i think that a lot of people uh things that um we still have a lot of open ended problems in software i'm not saying that like is is also i think like but i think it's like we are on a trajectory where things are just being sold at a much much faster rate right And it's so hard to predict how long things are going to take.

42:41So a lot of people start looking into the physical world. How do we get AI to be able to perform in the physical world? And I think it's like, even from my perspective, and I could be wrong, is that I see the AI agents in the digital world actually have a lot in command with AI in the physical world. So I think if you look at the, so I was looking, so when we were working on AI agent, we were trying to create a mental model of what it's like. And I think it's like an agent is something that is interact with the environment, right? It can perceive the environment and act upon it. And then I get feedback from the environment, right?

43:21So in the digital world, like a lot, so the agents work in the digital environment. They can work in a browser. It can work in like a computer, like a terminal. It can work in a VS code. And the actions can be read, write, web browsing, like send queries, right? So a lot of that is like a digital agent. In the physical world, the environment can be the road, right? The agent can be a car with actions like turn left, turn right, break, and things like that, like accelerate. And on the actions, so you can actually map them. But the challenge with doing physical world is that the physical worlds are not very well described.

43:59So in digital world, ideally, not everything is like that, but in a lot of APIs, you have pretty good documentations. You have descriptions of the functions. If you send a request like this, you receive responses in this shape. You have status codes. You have arrow descriptions. It's pretty interesting. Whereas in physical world, we don't really have a good description of like, okay, if you squeeze this much force into the X, it's going to break. It's going to break in that direction, right? So like as humans, we learn to operate in the real world because we have observed over time. Like we learned just like, okay, if we, this, we kind of like some kind of haptic intuitions.

44:41Like we don't, we don't press too hard something. Like we don't step on a child because we know that the child is going to get hurt. Don't step on a child.

44:48Jon Krohn:Yeah, this is very important. Yeah, we learn that. Yeah. Everyone at home, remember. So, yeah. So, the physical world does not have this. So, the AI, it can maybe reason, right? It can know that, okay, if I want to do this, I should be able to, like, make up, come up with a plan of actions that can help me achieve that task without causing the consequences I don't want to cause, right? So, I do things that's like AI is getting really good at reasoning. It can come up with a task and come up with a task and then it can come up, sorry, it's given a task, it can come up with a plan to solve the task if it has a good understanding of the physical environment.

45:33And I think there's a lot of initiative around the world model, like how to build a model that encode physically accurate information about the world so that AI can operate in it.

45:45Jon Krohn:World models, that's the word. Yeah. Yeah, so we've had big fund races. I think Fei-Fei Li was the first with her World Labs. Yann Le Ke now with AMI Labs, a billion dollar first VC round, which is wild. David, I'm forgetting his last name. There's also a Google. David Ha. No, David Silver. David Silver. From Google DeepMind in London. He raised like a billion dollars for his world model company too. Who are you saying? David Ha? Oh, David Ha and Smith Schuber. were the authors of the proof of concept world models, I think, in 2018. And it was one of the first. Like, world models is not a new concept, right?

46:27I think we've been trying to model the world for, like, a very long time. But the ideas of, like, using neural network to, like, build that, I think, is more modern. Yeah, the world models.

46:38Jon Krohn:Yeah, it's really exciting. And so there are hard problems in robotics. Simulation to real-world transfer, scarce action data, sub-second latency, The cost of being physically wrong, like stepping on a kid. So it seems like that is similar to the kinds of real-time machine learning problems that you were obsessed with at ClayPod. So I think a lot of – so I think there's a different part of it. So I think, like, there are many challenges with AI in the physical world. So one of that is the AI part, right? They had to make AI capable of understanding the world, coming with a plan of actions. So, like, the hardware part, right?

47:23So, I think I actually realized this talk by the Unitree CEO. Do you know Unitree?

47:29Jon Krohn:Unitree? Unitree. Evening Tree? Unitree. Unitree. Yeah. Oh, Unity. Oh, no, Unitree. Unitree is a Chinese company. Oh, no, I don't know that. Yeah, no, I think they're one of the very cool robotics companies. Yeah, so they just fight for IPO, by the way. and one of the very rare robotic companies that claim to be profitable. So I'm actually very excited about the IPO. So the CEO has a great talk when he talks about robotic intelligence and he distinguished between two parts, right? One is like reasoning and the other is movement. So what that means is that like reasoning is like, okay, so robot look at the thing, given a task and I think about how to achieve the task, which is actually very similar like reasoning for like digital agent with the caveat that it has to like the physical agent have to have a good understand of the physical world.

48:24The other part is movement. And what that means is that like the robot may come up with a great plan and say, okay, to do that, let's just walk over there and like open the door. But then if it's just like go one step and fall over because it tripped on the wire or something,

48:38Jon Krohn:you would think it's stupid, right? So it's very important to have the robot to do like be able to do a lot of movement. I think that part is doing a lot of increasing performance in the last few years. So I think he was talking about how they had this really, really cool robot doing kung fu. So they have a lot of fleet of robots, and they're just performing kung fu. And it's like six feet tall or something? Like human-sized? So they have two humanoid. One is a G1, and the other is a H2. so the H2 is very imposing it's like 6 foot something but it's actually very hard to work with I don't think anyone I know is working with them actually I have one friend who has an H2 and he was like oh do you want to take it and the reason is that they got one but then they found it's too heavy they wanted to just give you a robot but then I don't know what to do with it what am I going to do with a 100 something pound robot it's heavy, I cannot carry it and it is falling on you So you're not ideal.

49:40It's not great. But it's a G1 scooter. It's like five foot or something. Oh, that's too pretty big. It's a lot easier to work with. So yeah, so they have this demo of robots doing kung fu. And instead, it's like, OK, to do that demonstration, they program the robots to do 20 movements, right? So like, OK, do this, I don't know, punch or something, whatever. And all of that is pre-captured. And then the robots just use different motion to combine them to do the demo. But then he was talking about like, we actually get, he believes that in the next six months, they can also do like instant, like arbitrary motion generations.

50:20So the robots can do like different motions without having being like pre-programmed, like pre-captured. So I think it's very exciting. So I think like, okay, there's two parts of robot intelligence, right? One is the reasoning, which I think this AI is getting like pretty good at. We're understanding that a lot of people with a lot of VC money is trying to solve. And there's the motion parts that I think is like getting very exciting. So, yeah, so I feel like all the pieces are moving, I hope, in the right direction.

50:47Jon Krohn:Very exciting. I can't wait to see what you do with this. Obviously, it's stealth right now, but it's going to be exciting. For our listeners who can't raise a billion dollars in VC money in the first round, what recommendations do you have? You were kind of talking about this earlier, this kind of existential crisis, which I feel as well. You know, it ranges from across everything I do. You know, I had, I did a PhD in AI and that gave me a real moat around my career. You know, other people couldn't create a machine learning classifier or, you know, understand problems with labeling data or these kinds of things.

51:20Jon Krohn:But now all of those kinds of things, a machine can do, you know, no problem. For the podcasting too. I mean, it lowers the barrier to entry. You know, you don't have to be a great writer to come up with great topics, you know, to script episodes. so you know there's there's a lot of people in a lot of industries who would feel like the mode is going away from them i'm wondering so obviously focusing on physical systems is one way to you know create a bit of a moat for yourself because you know hardware r &d cycles are going to be longer than software and there's lots of different ways that hardware can specialize and anthropic or open ai aren't going to next month just all of a sudden have a robot that does that too So there's a mode in physical AI.

52:01Jon Krohn:So maybe that is just the answer, but are there any other ways? Yeah, what other ways do you recommend? I guess you also have the systems design idea right at the beginning of this episode. What other tips do you have for our listeners on how they can try to future-proof themselves a little bit in this time? Well, nothing. this is really funny because um i i have um i have a friend who is an economist like he's one of the smartest people i know and he does like consult a lot of governments and coming up with like technical uh policy tech policies on like um yeah how how how to get their nations like stay up to date in the air era and he was straight up telling some of them it's like yeah there's nothing you can do it's like you're you don't have enough budget for it like just don't do anything so it was like yeah some some people do have a very very pessimistic views of the world um but um i think like i'm more on the optimistic side i do things that's like ai can solve a lot of problems but i also think that like there will never stop being problems for me like for us to solve so for one thing like doesn't matter like how many ai models are there or how good AI models are, I would never stop being angry at people on the internet.

53:20There would always be people that piss me off. There would always be customer services I'm unhappy with. There would always be things just like collaborations. It's not quite straightforward. So recently, there's a founder, and I really like the founder. He's very smart. So he came to me and he pitched this idea of another Asian orchestration framework. And then he told me that all the problems that he has seen with a lot of companies is that there's not enough communications between product and engineering, which is very classic. And he was like, okay, and my agent orchestration platform is going to solve that.

53:58And I was like, I don't think that's a technical problem. I think usually when product and engineering people don't talk with each other, that requires people's solutions. It usually, you don't solve that by like, okay, here's another tool. We magically make product people and engineer people get along. So there are a lot of people problems that it's not quite like so. So I think there's several categories of tasks. I think it's like, are not entirely completely clear for us to like how to solve. Like one is like human AI collaborations, right? So a lot of AI tools nowadays kind of built upon legacy systems and things about how we interact with old software systems.

54:47So just an example of coding tools. So originally, we have a lot of coding tools that are just part of VS Code because VS Code existed. And then we have a lot of coding tools as part of the terminal because terminal has always existed. But then this made me think, wait a second. why are they why are like editing like note taking apps like VS Code and Terminals why do they need both of them like why you know like why I mean I'm just going back like why is it different stuff and another thing is why is Terminal so hard to use so a lot of engine for me like I have used Terminals like for I know in school and stuff and for work like I use it but I'm not crazily happy with it

55:33Jon Krohn:a big Vim person. No, okay. So I have friends who are crazy Vim person. Like I have friends who just don't use VS Code or like PyCharm or anything. They just like straight up code in their terminal. So you think it's a lot faster with all the like key, right? So in theory, you could use terminals as a... As an IDE. Yeah, as an IDE. Yeah, but so, but like, yeah, okay. So we know that terminals exist, but like because of like coding tools like Cloud Code, a lot of product people or like people who never used terminal before are suddenly exposed to terminals. And then it was like, okay, why is this so hard to use?

56:07You know, it's like, it's very painful. And I think it was like, okay, terminals, maybe that terminals are hard to use on purpose because terminals are actually very powerful. Like terminals basically give you access to a control plane for you to like control the computer. You could easily like remove like RF, like everything, right? Like, yeah, like you can just do that. So maybe like you make it hard to use so that only people who are willing to get used to it use it so they are less likely to make mistakes. So, yeah, so I think maybe, but it's also like a legacy thing, and I think, like, why don't we have something, like, in between?

56:43Like, yeah, we have something that can be both very, like, can give you, like, access to the file system, access to the computer, like a control plane, a terminal, but also easy to use, like an IDE. And a thing like that is, I'm going to talk about it, And actually, OpenAI and Anthropocene introduced a bunch of services like desktop apps, right? Like the Codex desktop app, which is basically the same idea of like, okay, very easy to use interface, but give you access to a lot of things the ways that a terminal can. So I think this is evolving. And also there's a bunch of how to access your agent when the computer is not working.

57:20You probably have seen people complaining about like, okay, like why I have to keep my computer open all the time because my coding sessions, like my cloud coding codecs are doing their things. Yeah, yeah. Yeah, so I usually just get into an Uber and it's just like, can my computer open? I look like a freaking nerd. But I was just like, it's just making me think like, there's no reason my computer should be open, right? Yeah, because it's totally run as a cloud. And then we need like something that can access through the phone. I think like people, a bunch of people are building it. Like, okay, now you've won your things on the phone.

57:52Now you need some kind of sandbox, right? because like how do you share context between the phone and the computer? So anyway, I'm going to like very much, it was like genres like, shut up, she's talking with you all.

58:05Jon Krohn:Well, I'm just wondering what, this has all been very interesting, but what is this all, what is this all, are you saying that there's lots of lots of problems that we can still solve? Yeah, so I think it's like, I use it, not a great example, obviously, but I was saying it's like, it's one thing is like, we don't quite have a good understanding like what is the optimal way for humans to use AI. So human AI interface is one big thing, right? Another category of problem is it's like... I see. You know what I'm saying? She's not done. It's just like just the first. It's one of my first in the 20 points.

58:38No, this is really important. This is good. Yeah. Another thing, if you will, let me say it, is how AI interacts with the world. So we have a lot of techniques to make AI good at using tools, right? But I think for AI to interact well with the world, we do not just want to improve AI. We can still make the world more AI-ready. Right? How do you, for websites, for apps, we can make better documentation, better APIs that agents can call, better security, making, OK, this kind of actions are dangerous. So maybe you should have less permission to AI and stuff like that. But how about physical world?

59:19So I'm not sure you've seen this very cute video of a food-delivering robot. A foot what robot? Food-delivering robot.

59:29Jon Krohn:Food-delivering, right, right, right. Yeah, so it's a very tiny robot. Not a foot-delivering robot. That would be weird. Do you have extra feet to deliver? I'll take six feet, please.

59:41So the robots are very cute. And then the robot just couldn't cross the street. So the robot had to ask the pedestrian, like, hey, can you press a button for me so that it turn green so that I can't cross? And the pedestrian was like, what the heck is going on? So I actually talked to someone who worked at one of those robot food delivery companies. And he told me, like, the hardest part is just, like, how to get the robot in track of the world. So, like, actually some cities have this, like, streetlight API. so that the robot could connect to the Streetlight API so it can turn, it can press the button via the API.

1:00:21Jon Krohn:Oh, because it can't press the button to say that I want to walk. Yeah. So I'm just saying that part of how you could provide tooling to make the wall easier for AI to operate in. Sure, sure, sure. Yeah. Nice. Well, these were lots of great ideas. We do need to start wrapping up the episode a little bit because our next guest has actually arrived, a friend of yours. So we're going to, you're going to chat to him in a moment, but we do have actually for you, we have some audience questions. And so I want to get at least one or two of those in because we got a huge response to you coming on the show.

1:00:58Jon Krohn:It's maybe, you know, one of the biggest responses we've ever had about a guest coming on. All right. So our first audience question. So I posted on LinkedIn that Chip would be on the show. We had this tons of questions come in. And my first question, there were way too many to ask, but I'm just going to ask a few that I think are some of my favorites that might interest the audience the most. So Rahichia Valpuri, who is a business systems analyst for CIBC, a big Canadian bank. She's in a suburb of Toronto called North York. It's a nice city. Yeah. I'm from Toronto. Did you know that? Oh, that's why you're so nice.

1:01:35Jon Krohn:and um so rahisha says that she loves your writing so much so that when she's reading your books i kid you not there are moments when she finds herself smiling and she wonders if you're working on a new book oof so i've been playing with this idea of like writing a novel from an ai perspective with an ai narrator so it's kind of like try to explain how ai works from the ai perspective It's a crazy idea. I'm not sure I wouldn't do it. But I do want to get back into crib writing. I think my physical AI could be very interesting as well, but that also evolving quite a bit. So I'm trying to finish one blog post about physical AI.

1:02:16Hopefully it's going to be out by the time the episode is air.

1:02:21Jon Krohn:Nice, yeah. That should be easy to find from your website, which we'll mention the URL of in a moment. And then, yeah, some of the questions we actually addressed in the episode. So someone named Jing Xu, who is a, she is a very frequent listener and she makes lots of posts tagging me. And so I really appreciate that. Keep it up, Jing. Thank you. It's always great to get so much interaction from you. So she works as a director of data science at Elevance Health in Chicago. And she had lots of questions about how Vibe Coding, so Gen AI models, how that has changed. What is expected of AI engineers?

1:02:56Jon Krohn:And I think we talked about that a lot in this episode. So that was kind of even your very long answer near the end of this episode was basically about opportunities for us. You are free to cut it off. Okay. No, I can't. I can't do that. So, yeah, I think that we've kind of covered a lot of the really big questions. Okay, here's one very last one. So this is from Brian Willett, who's a retired research scientist. Very easy question for you, Chip. When will AGI be achieved? Oh, is that like – so it's the best definition. of AGI, do you follow the lawsuit with OpenAI? I think the contract between Microsoft and OpenAI is contingent on some definition of AGI.

1:03:37They already evoked that. So by the definitions we are...

1:03:43Jon Krohn:Legal definitions, we have AGI. I mean, it depends on whose legal team, right? But I think by the contract, I think Microsoft evoked that we are in AGI. Wow. I guess we'll see what happens there. We'll let the lawyers decide. Do you feel better or worse that we are in AGI? I mean, it's a difficult thing. There's a paper I did an episode a few years ago, a Five Minute Friday episode, on the five levels of AGI. And I think this is an important thing. And I think there's lots of areas where we do have amazing capabilities now in text-to-text, text-to-code, code-to-text. You know, in a lot of ways, in a lot of domains, you know, we have now exceeded the average person's capability on those kinds of tasks.

1:04:29Jon Krohn:But there's all kinds of things you brought up, like, you know, squishing an egg or stepping on a child. There's all kinds of things that AI systems don't need to improve on. So, you know, yeah, difficult to define exactly, but I guess we'll see what the lawyers say. So, yeah, so that brings us to the end of this great episode. It's been so nice to be here with you in person. And before I let you go, I know that you're famously a voracious reader of books. I'm a good influencer. I'm not sure that means anything to you. Yeah, you're a big, I mean, you're a huge AI influencer. And now you're going to be a big literary influencer because I understand you have an Instagram channel dedicated to your books.

1:05:07Jon Krohn:Not that you write, just the ones that you read. So I really like books. I like, I like, I like reading. So I have this Instagram account I created recently, like Chips Lip. which is really hard. I just realized they're very hard to pronounce. But like Chips Library, short for it. Chips Library. Yeah, I was showing my friend, my account the other day, and he was so disappointed. He was like, this is the most overhyped book account I've ever seen in my life. Because I feel like, okay, I finished reading a book and I want to post a review. And then when I get busy, so I just don't post a review.

1:05:39So yeah, but I try. So yeah, so I like, so I try to focus on books. So I think of as both fun to read and teach me something new.

1:05:52Jon Krohn:Fantastic. Do you have a new book recommendation for us? Since when you were last on the show, we did that one remotely and you were on your walking pad. And I had to ask you before we started recording, I had to ask you to stop walking on the walking pad because like the camera was moving, you could hear it. But behind you on your walking pad at your standing desk at home, you have a big bookshelf with lots of books. You pulled a number of books off the shelf to tell us. Do you have anything new that people need to read? Oh, so last year, I think I realized a book like Apple in China. So it's about the Apple supply chain in China.

1:06:27It's fascinating. It shares a lot about like how both Apple strategy and how China, just Chinese government thinks about technology supply chain as a comparative advantage. And I think like it's actually placed out quite interesting as in robotics nowadays. it's just really hard to find an American company that is as fast moving as a Chinese company in terms of robotic hardware. So I try to buy some robots, right? And I go to a bunch of American companies and a lot of them just don't have robots to sell. Whereas if you go to a Chinese company, you're just like, okay, which one do you want? Here's all the options.

1:07:06It's like, whoa, it's amazing. So it's very interesting. that book is I still like very much into like like urban design because I feel like city is a system and I'm thinking about how I approach that and there's a book that I found very amusing recently it's a book on parking I think it's like paved paradise I know parking right I think do you know that like on average there like you have to allocate six parking spots per car in the city

1:07:36Jon Krohn:really I know it's quite crazy that's wild Yeah, yeah, yeah. Because you have one other apartment, one other work, shopping mall or something like that. Oh, I see. So it's kind of interesting. And so I tried to read about it because I do think the self-driving cars are going to change a lot of that. Sure. Because a lot of cities nowadays are designed around parking space. Like having been to LA, it's extremely spread out because it needs a lot of parking space. And which also makes it worse because there are a lot of parking space, a lot of empty space. So now you got things just further apart.

1:08:10So more people need cars. It's just like kind of like vicious cycle cycle. So I think like self-driving cars are going to make things so much different because people don't need to park anymore.

1:08:19Jon Krohn:For sure. So I think I find that book's very, it's amusing. It's interesting. Yeah. Just like a bunch of like other books like that. Yeah. Nice. I'm sure we could go on and on. Pave paradise. Put up a parking lot. We will end on my little jingle. how should people follow you after the episode? You have over 300 ,000 people following you on LinkedIn. It's incredible. So you can follow Chip obviously on LinkedIn. I wish it means something, but yes, thank you. It definitely means something. I'll tell you that for sure. Anywhere else that people should be following you? We got Chips Lib on Instagram.

1:08:57Jon Krohn:We've got your LinkedIn account. It's not an account. If you follow it expecting to see the latest Asian tech news, You want to be so disappointed. But yeah, so like, I think that's just classic Twitter, LinkedIn. I'm trying to start a sub stack. I've been trying for like two years. I have a coming soon and people just keep tagging me like, you said coming soon a year ago. I know. When? I know. All right. Well, we'll see maybe next time you're on the show. We'll have your sub stack to talk about as well. It'll be vibrant by then. thank you so much chip for meeting with me in person in san francisco it's been such a great episode such an honor to meet you in person and yeah hope it's not too long before you're speaking to my listeners again yeah no thank you so much for having me again and congratulations for like 10 years a thousand episodes i know this is amazing i know yeah the very next episode will be episode a thousand thank you for being episode 999 what a special spot yeah no this is amazing Thank you.

1:09:56Jon Krohn:Wow, what an episode with Chip Hu Yen in it. She covered the sharp line between machine learning engineering, which meant building a model from scratch by collecting data, training and deploying, and AI engineering, which treats powerful models as a service you call instantly, making time to market dramatically faster. She talked about her start simple philosophy, meaning reaching for prompting first, RAG, retrieval augmented generation next, and fine tuning only as a last resort. She talked about how the hard part of physical AI isn't reasoning, but the fact that the real world has no documentation.

1:10:28Jon Krohn:And one fix is making the world more AI ready. Like cities exposing a streetlight API so a delivery robot can change the light itself instead of begging a pedestrian to do so. And she talked about how AI is making the act of building software nearly free. And in that new paradigm, the value of software hasn't dropped to zero. Instead, the durable problems worth solving are increasingly people problems and physical world problems that no model can simply copy. As always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for chips, social media profiles, as well as my own at superdatascience.com slash 999.

1:11:11Jon Krohn:That's fun to say. Thanks, of course, to everyone on the Super Data Science podcast team. our podcast manager Sonja Breivich, media editor Mario Pombo, partnerships manager Natalie Zajski, researcher Serge Massis and founder Kirill Aromenko. Thanks to all of them for producing another fantastic episode for us today for enabling that super team to create this free podcast for you. We are deeply grateful to our sponsors. You can support the show by checking out our sponsors links, which you can find in the show notes. And if you yourself are interested in sponsoring an episode, you can get the details on how by making your way to johnkrone.com slash podcast.

1:11:48Jon Krohn:Otherwise, please help us out. There's lots of other ways you can do that by sharing this episode with folks who would love to listen to it, reviewing it on your favorite podcasting app or commenting on YouTube, subscribing if you're not already a subscriber. But most importantly, I hope you'll just keep on tuning in. I'm so grateful to have you listening and I hope I can continue to make episodes you love for years and years to come. Join us for a very special episode episode 1000 coming up in a few days um it'll have both kyrill the original host and founder of the show as well as me and lots of people dropping in from all over the world to ask us questions uh including listeners just like you all right until next time keep on rocking it out there and i'm looking forward to enjoying another round of the super data science podcast with you very soon

1:12:43Thank you.