972: In Case You Missed It in February 2026

0:00Jon Krohn:This is episode number 972, our In Case You Missed It in February episode. Welcome back to the Super Data Science Podcast. I'm your host, Jon Krohn. This is an In Case You Missed It episode that highlights the best parts of conversations we had on the show over the past month. I'm taking my first clip from episode number 965. It was a technical interview with Will Falcon, co-founder and CEO of Lightning AI. In this excerpt, Will explains how he converted his wildly successful open source project, PyTorch Lightning, into a startup with over$500 million in annually recurring revenue. You know, you're in this great ecosystem, both with this commercial relationship with Facebook, but also all the open source stuff that's encouraged by Facebook, by NYU.

0:47Jon Krohn:What were the conversations that led you to founding a company? Was it because you'd already done it with NextGenVest? So I did not want to start a company, first of all. I think if you've done startups, you know they're hard and I just come from like, you know, grind. And so I was very happy to like become an academic for a bit, like read. I mean, it was amazing. I was like reading books and doing research and like I love science. I just love thinking about things, going into really hard, impossible problems and like going as deep as I can. So if everyone's thinking ever like should I do a PhD or not, that's what I would ask you.

1:19Like, do you love science so much that that's what you want to do? because if all you're doing is for a career step, you're going to hate your life like two years into it. But I loved it. It was amazing. And so I did not want to leave. I mean, the problem is PyTorch Lightning took off and I wasn't trying to make that a thing. It just took off and people started using it. And so then VCs started pinging me. They started emailing me being like, hey, we see your open source project. It's working, this and that. And I was like, hey, I'm good. I just sold a company. I just want to focus on research.

1:50And they're like, I'm more interested. Yeah. Exactly. And you know, now at the time I think about it, like all these things have names today, but what I was doing back then was pre-training world models. Today that has a name. Back then we didn't call it that, but I was training me pre-training models on four, 8 ,000 GPUs, single person, doing all the things that the pre-training folks are doing, like OpenAI and whatever. Like how should you scale the thing? What gradients should you use? Like what's a distributed strategy? How do you prevent faults? And like, what if the loss function is this?

2:17So most of our research at FAIR, at least for me, was more theoretical. So it's more like, how do you change the math function? So every model has a loss function that like optimizes something. And so it's more like, how do you change the math function? Like, do you add a regularizer here? Do you use this one or this one? And so I think what I learned was cool. And I have lost that skill a little bit, I think I need to get it back is when you leave undergrad and you've done math right you're really good at reading math it'd be like learning a language and you're really good at hearing it but speaking it is something you have to practice and speaking it is more like like let's say you wanted me to build a model to model this world here that we're in right now i'd have to figure out how to model the wall and like what does gravity look like and like how does the lighting hit and this and that and that's a math equation so being able to like translate that to math that was a skill that we developed a lot and that was what fair was really Oh, really?

3:13Yeah.

3:13Jon Krohn:That is a cool skill. Exactly. And so then that's what we train models with. It wasn't, oh, here's a model, train it. No, it's like, what is the training algorithm itself look like? And if you want to have better representations of the world, how do you accomplish that? Do you make things come together or separate? Do you embed images or not? And if so, how do you do it? Do you use this type of embedding or this type of function? And what's the similarity between them? And what if you add a penalizer and regularizer and this? So most of pre-training today is actually not that. it's more like the engineering of pre-training we were not just doing that but we're also doing the math and that's fair that's why fair was what it is these are like math people who are also really good engineers and they can solve these problems together right and so that was called pre-training um and so yeah i mean i think the vcs were kind of onto something i guess very early and um you know they asked me to come in for meetings and then like i flew out and like a thursday and then by monday i had like 10 term sheets and it's like i kind of got out of hand and so i was like well i I guess I'm starting a company.

4:11And so I called my advisors and I was like, hey guys, so I got all this money that they're offering me. Should I do this? And they're like, well, your research is going super slow. So probably. No, I mean, I've been telling them. I was like, hey, listen, like, cause I kept getting dragged into coding every night. And so I was like, research, run, you know, submit my training runs. Okay, it's running. And then I'm like, okay, go merge this pull request from PyTorch Lightning. And then it was like stuff that I wasn't doing. People are like, oh, can I have like this function for RNN? like the way you do a loss function for an RNN, right?

4:42Like backprop through time. I was like, I'm not using backprop through time, but like now I have to implement it for you. So I started working for people. I was like, this is terrible. I almost shut down Pitesh Lighting in September. I was like, this is distracting. And so, so my advisors were already kind of like, Hey, like your research is a bit slow. And I was like, I was transparent about it. I was like, Hey guys, like I just keep getting dragged into this.

5:01Jon Krohn:We're going to keep on looking into training algorithms and all that essential pre-training work we do with an excerpt from my interview with Tom Griffiths. In episode number 969, I asked Tom, who is professor of both computer science and psychology at Princeton, as well as author of the bestselling new book, The Laws of Thought, how he might adapt our understanding of human intelligence to guide designs for AI systems. You frame human intelligence as rational adaptation under tight limits. So explaining biases, heuristics, and few-shot learning, so learning from a few examples, which is what LLMs don't do, as optimal use of limited time, limited computational resources, and limited bandwidth for communication.

5:41Jon Krohn:So it seems like that, I didn't even really get to a question there, but the question that I was going to have, you kind of already started now getting into, but maybe with that context, you can develop on your answer even further. The question is, how should those constraints guide the design of the AI systems that we now build? I think one thing is, it's not necessarily the case that they have to. So I would think about those constraints as a good way of characterizing what the difference is between human minds and our AI systems, right? So human minds solve a set of computational problems that are characterized by these constraints.

6:15We live only, you know, a few decades, you know, give or take. We have to do all the things we're going to do using just a couple pounds of neurons. And we can only communicate with one another by making weird honking noises or wiggling our fingers on a keyboard. Those are the kinds of constraints that humans operate under. And so as a consequence, we have to be able to learn from small amounts of data, because we're just not going to live long enough to see lots of data. right it's not like alpha zero which is or um uh alpha go which is trained on many human lifetimes of go games right you just can't you can't do that human go players are going to be limited by what it's possible to learn in a human lifetime um uh you have to be able to recognize the structure and problems so you can use previous solutions that you found or come up with good sort of efficient strategies for solving those problems using the limited computational resources you have Right.

7:20So like, unlike Deep Blue, a human chess player can't search 100 ,000 positions per second. They can search much less than that. And they have to focus on finding the good ones to search. And you have to you can't overcome those constraints by, you know, being able to transfer the data that we've experienced or pool our computational resources directly because we have to operate under those bandwidth limits. And so we have to come up with good conventions for allowing us to do things like work together and to pool data, things like writing, you know, language in the first place, but writing, you know, science, starting companies, right?

7:59These are all mechanisms that we can have for being able to coordinate the resources that are available to us. And so that set of things really characterizes in some sense what it is to be a human as an intelligent system. And our AI systems are not subject to any of those constraints. right? They can learn from any human lifetimes. They can add more compute when more computers needed. They can transfer state or share the data that they train from. All of those things are much more straightforward for your AI systems. So one consequence of that is that you might expect that the AI systems are just going to be a different kind of intelligence from human intelligence, right?

8:37Because they're not operating under the same constraints, they're not solving the same computational problems. They're going to be able to solve problems potentially really well. And in some cases, they're going to solve those problems better than humans, but they're not necessarily going to solve them in ways that are recognizable to humans or ways that are the same as the solutions that humans find. And so that sets up sort of two possibilities. One possibility is we say, that's fine. We just want to engineer the best systems we can. We should just keep on doing what we're doing, keep going in that direction.

9:06And we're going to get this kind of divergence between what human intelligence is like and what AI is like. And that's OK. The other approach you could think about is one of saying, oh, well, actually, we want some of those attributes of human intelligence to be inside our AI systems. In particular, it might help us make AI systems that just make more sense to us, right, in the kinds of solutions that they find. And so if we want to do that, then we really need to understand the things that are going on on the human side and think about how it is that we transfer those over. And so we talked about meta learning as a trick for building systems that have inductive biases that are maybe more like humans.

9:40that's certainly an avenue that you can go down if you want to do that but engineering inductive bias is hard you know scaling is much easier than engineering inductive bias and that's the reason why we're in the current sort of moment that we're in an ai is that that's a technique that sort of works and it's an engineering problem we know how to solve it um uh but engineering inductive bias really requires having an understanding of the problems that we're solving and what reasonable inductive biases for those problems are. And it's something that people are still trying to figure out how to do well.

10:11Jon Krohn:I'm excited to see how the complexities of inductive bias based on human experience actually get implemented into AI models. I don't expect it will be long before we start to see real development. In my next clip, I'm turning to episode number 963. In it, Anshie Bart, a member of technical staff at Amazon's HEI Labs, came to the podcast to tell us about their latest product, Nova Act, and what it could do for AI developers. You drive the vision, strategy, execution for how developers engage with Amazon's next generation AI products. And the latest exciting AI product you're promoting is something called Nova Act.

10:48Jon Krohn:So tell us about Nova Act. Yes. So Nova Act is a service that we just recently launched. We had a research preview going on since March last year and are super excited that this past December at our AWS reInvent conference, we launched this as a GA service. Nova Act helps you to build UI automation tasks at scale very reliably and helps you to really kind of start prototyping and putting it in production really fast. So you can get started, which I love as a developer, really fast in a playground experience and then, you know, iterate on it, debug it. And once you're ready, push it onto the AWS side and run it there reliably and safely at scale.

11:34Jon Krohn:Really cool. And it's free to get started, right? Absolutely. So one of the key things, again, I'm excited as a developer is we want to make it really easy for you out there to get started, right? Like in this industry, we know the speed to delivery, speed to shipping is kind of the mode. So you don't want to like, you know, spend too much time spending up the right environment and infrastructure and integrations. You really just want to validate your ideas, right? A lot of startups, like you just want to go build the ideas you have, validate them really quick, iterate on them. So you can get started really quick in our playground experience to especially do that.

12:11And then you iterate there and then you can move into, you know, the next step, for example, if you want to customize more in IDE environments. So we really want to keep also the surface area really kind of, you know, where you're doing your day-to-day jobs as a developer.

12:25Jon Krohn:Really interesting. Would you be able to walk us through the typical user journey through a Nova Act experience? Obviously, I'll have a link to Nova Act in the show notes for this episode so people can go there and describe to them what it would be like to us to experience it as we go there for the first time and we're playing around in the playground all the way through to deploying. Absolutely. So you can go to nova.amazon.com slash act. And then there's the playground experience. There's also links to all the dev tooling we're offering, IDE extensions, SDK downloads. But really kind of the first step is you go into the playground free of charge.

13:03You don't need any AWS account or anything. So you just go in there. You provide a website. For example, you're going to, let's say, a booking website or a specific event, maybe sign up site. Right. Conference season is starting soon here in the Bay Area and everywhere. So you just put in the website and then in natural language, you can decide and put in the actions to take. Right. Let's say I want to sign up to an upcoming meetup. Right. So maybe I put in the Luma website and I'm going to say, hey, search for a specific meetup. Maybe I want to join, you know, an AI performance meetup. I look for that.

13:42And then you can also like, you know, put the actions in to fill out. like click the RSVP, click the join it and have it fill out the form for you. And on the same playground, you see an embedded UI, the browser environment. So you don't have to set that up any manual way. So you can see it right there. So you can observe how Nova Act, how the agent is going to that website. It's performing your actions. You see also the reasoning traces. So what it is doing, which is exciting, especially important for developers to be able, you know, to debug, to troubleshoot, to really see what's happening there.

14:18And you can tweak it if it's not getting it at the first time. So you can optimize your instructions, see it. And then once this is performing well to what you want to achieve, you can then download the scripts on the background. It's writing a Python script that captures all of those steps. So you have the ready script. You're using natural language. It converts it to the code for you. and then we have IDE extensions, for example, or just an SDK. You get that into your preferred IDE of choice. You import that script and then you're basically back in your coding world as a developer, right? So you can customize it, you can tweak it in there.

14:57And also the IDE extension has this embedded live preview. So, you know, a lot of times when we're building those automation workflows, you have like a separate window popping up that shows you the browser. and we got a lot of developer feedback that that's just a little bit too much, right? We want to stay in the flow when we're building something. So we included that in your IDE so you can really stay in there, have a unified window with all the troubleshooting, the debugging, the traces, so you can keep really close eye, develop, customize. And from the IDE, you can then, if you want to, also connect to your AWS account if you have one and for example, deploy it there to run it in production.

15:35Nice.

15:35Jon Krohn:So that sounds like a really easy way for somebody to be experimenting with developing an agent because you can go to nova.amazon.com slash act and then be able to use your natural language to describe what you'd like your agent to do. Watch it, do that task, and then get the Python code, use it in whatever environment you're comfortable writing your code in, and then that allows you to easily scale up whatever you're doing. It sounded like there was also, it sounded like you mentioned also being able to productionize on AWS with this solution. Right. So a core kind of motivation for us, you know, talking to a lot of customers, talking to developers out there.

16:10We've seen so many flashy demos, right? Like you go to any meetup, especially here in the Bay Area, but I assume now in other parts of the world in a similar way. And everyone has a flashy demo, right? And all the kind of fun stuff that agents can do. But what we observed and the feedback we've received as well from customers is that on average, those agents work maybe 60 % of the time. And to be honest, an agent that is reliable 60 % is 0 % useful, right? Especially if you want to productionize this, you need to really have reliability that you can trust those workflows and that agent to complete that one task.

16:48So this was the core motivation for us, kind of really the P0 to make sure Nova Act can reliably 90 % and more deliver on those workflow executions.

17:00Jon Krohn:Awesome. So how do you do that? How do you get that kind of confidence? How do you go from a 60 % reliability to better than that? What kinds of tooling do you have in Nova Act to ensure that? Yeah. So I want to talk a little bit how we train Nova Act, right? Which is exciting. At least I find it super exciting. So in the past when we trained, you know, AI models and things, it's a lot about imitation learning, right? You collect data and you show it that data. Now, as we're moving from kind of the chat-based conversational models that we all know and work really well into this space where we're building agents that need to take an action, that doesn't work anymore, right?

17:42Because the agent is not just predicting a next word, next token anymore. The agent needs to predict the next action to take. So what we're doing is we're building out those reinforcement learning-based web gyms, as we're calling them. We have actually one in the playground. You can actually play around and observe one. And those web gyms are replicas of very typical UIs. So maybe a form filling UI, a shopping workflow, et cetera. And we're letting the agent train in those web gyms. So imagine hundreds of those gyms, like typical UI design elements, they are typical tasks to do, and then give the agent like thousands of tasks, right?

18:24And they basically self-play in there. So this is similar like in the past, you know, how AI learned to play chess, how it learned to play Go. It's really kind of a trial and error approach. So the agent goes in there, tries to do that form filling. It might fail a couple of times, right? But it self-corrects. It does it again and it learns a better way to achieve the task. So this way, the agent understands to reason through this UI and understands how to accomplish the task and helps it also to generalize well, right? Because UIs change. So that's super important. If you're building that agent, yes, you have a specific site.

19:03Maybe you're in those gyms you're training it on. But then again, you want to generalize if the website changes, if the checkout button moves, if the sign-in button moves somewhere else, if it's using different icons. because a lot of UIs are really designed for us humans, right, to navigate. So, and for us, this is a simple task. If you think about maybe you go and write an email. And depending on which email program you're using, which application, sometimes the button says draft or new or create, right? Or it's just a little pen icon to create a new email. For us humans, we learned that. We understand that it's not a hard task for us.

19:41But if you're sending an agent to that environment, to that UI, the agent needs to be able to generalize and understand and reason in a similar way, even if, you know, some buttons and some tools says draft, the other one says create. So this is really exciting. So you're training the models and Nova Act is trained like this on those web gems to then really kind of be able to generalize and navigate those web pages, even in real life as they're changing the structure and then they look.

20:10Jon Krohn:On this podcast, I'm always going on about how Claude Code is mind-blowing, but now Claude Co-Work is making my jaw drop as well. For example, I recently wanted to quantify how healthy my sales pipeline is for my AI consulting business. I simply asked Claude to estimate my sales for the coming quarter, and it brought info from relevant Google Sheets and my Gmail to create a professional spreadsheet of clients with estimated revenue for each one. Well, this might have taken me a day. Instead, it was done flawlessly with Claude Co-Work in minutes. Claude is the AI for minds that don't stop at good enough.

20:41Jon Krohn:It's the collaborator that actually understands your entire workflow and thinks with you. Whether you're debugging code at midnight or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. Ah, and you'll appreciate that I can ask Cowork to show me data such as my sales spreadsheet, and it provides an interactive chart right in the conversation. For problems worth solving, get started with Claude at Claude.ai slash superdata. That's Claude.ai slash superdata. And check out Claude Pro, which includes access to all of the features mentioned in today's episode, Claude.ai slash superdata.

21:14Jon Krohn:And in my final clip from last month, we're sticking with this potential of creating self-reliant AI systems. In episode 967, I chatted with Praveen Murugaisan, the VP of engineering at Samsara, a publicly listed Internet of Things company with a billion dollars in annual revenue. In this clip, Praveen fills me in on how quantum physics might be the catalyst for creating AI agents that can operate free from human intervention. Something else, another emerging technology that you've talked about in a blog post we caught is talking about quantum computing. So I wanted to get into that a little bit.

21:50Jon Krohn:In an article you noted, and we'll have a link to this article, it's in something called Diginomica. And so we'll link to that so people can read it in full. But you note that connected operations platforms already do many things the transport, you know, transport reimagines and that quantum computing may continue to enable this by crunching, quote, unimaginable amounts of data. So you talk about something that would today, you know, take a traditional compute platform. Ten septillion years can be done in minutes with quantum. So, yeah, tell us about what Quantum might be able to do in your space around routing, predictive maintenance, scheduling.

22:37Jon Krohn:And yeah, go ahead. Yeah, I think for when you really think about it, right, like a very practical example is like routing and scheduling, you could say. Right. I give you an example, which is like pretty much everyone who studied computer science has gone through like the traveling salesman problem. Right. It's an NP-hard problem, which means it just scales exponentially in terms of computational needs for you to solve it. And anybody who's trying to solve that problem effectively uses optimization techniques. So in a simple sense to say it's like a traveling salesman problem where you're basically saying for routing, a vehicle has to visit 20 stops.