955: Nested Learning, Spatial Intelligence and the AI Trends of 2026, with Sadie St. Lawrence

23:09And so much higher growth in what we're seeing for AI prompting, AI engineering skills, particularly compared to data science skills. So check out the LinkedIn jobs report. It's a great way, I think, just to kind of see what's trending. Think of it kind of like Google search keywords, but for what's happening in the job world and what's happening on job descriptions.

23:34Jon Krohn:That is really interesting. Big implications for all of our listeners who are trying to have the most relevant skills of the moment. And actually, so this is going back to the beginning of the year, but I did do an episode number 856 talking about the fastest growing job in most developed countries in the world. And that was AI engineer. So more in that episode on the kinds of skills that you can be developing listener in order to be super employable. and I don't know how you feel about this Sadie but to me it's kind of obvious that AI engineer is kind of like a sub-specialization of data scientists that has emerged and you know not really a completely distinct area but yeah that's kind of how I perceive it.

24:24I personally think the data scientists created a family of jobs so when I got into the field in 2014 it was thought that you were like this unicorn and everybody knew we were putting way too much on like the world data scientists because you're supposed to do everything from data engineering to data visualization to building being an ml engineer and building out these models also taking it to production so doing ml ops like those jobs that i just described really didn't exist when the data scientists got created the data scientist was that and then we realized oh most people aren't a full stack data scientists.

25:03And we can actually move faster if we split this into particular roles like MLOps, ML engineering, et cetera. What I'm seeing happen is not only did the data scientists breed out a family of jobs and five different job descriptions, but now we're seeing it change, like just a rebrand from ML to AI. And what's funny is I was just doing some, I was reading the book Thinking Machines, highly recommend that book to anyone. And And they do a whole history on AI. And they didn't want to call it AI back in the 2010s because they didn't think anybody would take it seriously. So to avoid that, they called it machine learning.

25:42And now we're back to the term AI. So I just I found it humorous that like they came up with the term machine learning as a way to get away from AI. But here we all are back to AI to begin with.

25:54Jon Krohn:Wow, that looks like a cool book. So it's a 2016 book by Luke Dormell about the history of AI. Yes. I like that. A big portion of it is about NVIDIA and really like their leadership in it. That's a different book. That's a different book. Okay. Gotcha. Gotcha. Gotcha. That's by Stephen Witt. Yes. It turns out a bunch of people have named their book Thinking Machines colon something more. And yeah, so this one is Jensen Huang NVIDIA and the World's Most Coveted Microchip by Stephen Witt. Cool. I'll have that for folks in the show notes. Yeah, it's surprisingly a great, if you really want a history on AI since the 90s, to me, the book is more about a really more detailed history of AI and the hardware that supported it to get us to this point.

26:44So great read.

26:46Jon Krohn:Nice. Fantastic. All right. So that wraps up your predictions for 2025. We ended up going off on a lot of tangents and doing some forward-looking stuff. Before we get to your 2026 predictions, we're going to do what we did for the first time last year and that I personally love doing. I think our audience loved it as well. Do reach out to us on LinkedIn and let us know how much you like or don't like the next segment so that we can evaluate how much we emphasize it or maybe expand it in the future, or maybe stop doing it, but had a lot of fun. So last year we gave out awards. Well, you know, you and I picked winners.

27:25Jon Krohn:We didn't actually create awards or send them to anyone. But we, you know, we awarded four things for the past year. So there's our wow moment of the year is number one. Number two is the comeback of the year. Number three is our disappointment of the year. And number four is our overall winner in AI, I guess, data science over the past year. So let's do wow moment first. Do you want to do yours first or do you want me to do mine? I'm ready. So let's go. So my wow moment for the year was Nano Banana, particularly because that's why I thought I felt the biggest change in image generation from last year to this year.

28:07And particularly for me, I'm remodeling our house. And so to take an image of a room and say, what would it look like with this tile floor sample and throw in the sample and like how accurate it is was really mind blowing. And particularly as you know how those models work, you don't have a ton of control over them. And so the fact that it was able to replace the floor with my tile, I think also because I got a lot of use out of it this year. It definitely has been my newest sidekick, but also like a ton of help. What's been your wow?

28:44Jon Krohn:Nano Banana Pro did pop into my head as my potential wow moment of the year, Sadie. It is very impressive, but I spent some time reflecting on this and trying to think over the whole year. And my big wow moment came in the spring when Ed Donner and I were developing a full-day talk, a full-day workshop for the Open Data Science Conference East, ODSC East in Boston. and the demo that we came up with and Ed developed and delivered. So we kind of over the course of the whole day, we taught people what we perceive as the key agentic AI frameworks that people need to know. And so we started off by teaching the OpenAI Agents SDK to show people how you can get an individual agent up and running with guardrails, doing the kinds of tasks that you want it to be doing.

29:42Jon Krohn:Then we taught them MCP, Model Context Protocol, for equipping their agent with tools that they can use. And then the third thing is we introduced them to Crew AI so that they can get a team of AI agents now working with their tools, thanks to MCP, all working together on some particular task. and what Ed did at the end. So we kind of like over the course of the day, the sophistication of the hands-on demos that Ed delivered got more and more complex. By the end of the day, he had a team of software developer agents. So there were four of them. If I'm remembering correctly, there was a front-end developer, a back-end developer, a tester, and a project manager.

30:31Jon Krohn:And so each of those four agents had different context to work with, had different tools that they could use, kind of different specializations as things like the front end, back end developer distinction would suggest. And Ed used that crew of agents to create a trading platform. And this trading platform worked. So it was a trading platform where you could send buy orders, sell orders. It would simulate the stock market. And then you could buy or sell any given stock on a real stock market. Well, you weren't trading on a real stock market. It was simulated, but it was all the stocks, say, on NASDAQ available to you to trade with.

31:18And you're using the real-time price information in order to be able to trade these simulated portfolios.

31:25Jon Krohn:and then he created a crew of trading agents named after famous traders and each of those agents had that famous traders kind of style and then those traders used the software platform that the crew of agents had built and it's that kind of, like it blows my mind that that works. Like this was a robust, sophisticated trading platform. And the team of software developers of fake, these aren't real people, people. These are agents. They created this thing over the lunch break. So he, you know, he set it off running right before we took a 90 minute lunch break. And when we came back in the afternoon, the software platform worked.

32:18Jon Krohn:And that's, that's crazy to me. Now, if I can add a like a honorable mention to a wow moment, that would definitely be lovable. And, you know, what they've done with their platform is very similar to what you're describing with the team of agents. It's really that but for front end development. I mean, I redid our website for HMCI really with one prompt and then there was some, you know, so you can go to our website and let me know if you like it or not, because it literally only took like one prompt. But overall, I like, it's just incredible. Like I'm still, I just, I feel so old, John. I'm like, I remember the day when I was editing code for my MySpace, where I was like even using Squarespace and felt like it was so much further ahead in terms of like drag and drop website development.

33:09And now it's like one prompt with a team of agents. And you have a whole website, a whole platform, and you're good to go.

33:15Jon Krohn:So I've got it up here. The HMCI.AI website looks pretty good. Pretty good. Built by a bunch of agents and thank you, lovable. Wow. Yeah, there you go. Really cool. All right. So, yeah, so that's our wow moment of the year. Congrats to the winners. Wow. We're going to skip their speeches and move right on to the next category. Great. So the second category for our 2025 awards, Sadie, is comeback of the year. And I think you and I actually let it slip to each other before we started recording. So most of these, we don't know what the other is going to say, but for comeback of the year, it's just so obvious.

34:02Jon Krohn:I don't know how anyone could argue with us on this one. And we have the same one. Do you want and tell the audience what it is? Yeah. You know, not only do we have the same one, but this was our same one from last year. So we have not changed all this. So any of your guesses, but we are both solid on Google. I mean, not only did they come back last year, but they're coming back even stronger than ever this year. And I think just not only the suite of tools that they have, but you know what I've been most impressed with was how they've integrated into search and integrate it into G Suite. And just I think they've done a really great job of integrating it into their existing products in a really thoughtful way that doesn't distract from the old way of working and has honestly, in my opinion, been really seamless.

34:49Like I continue to find myself using the AI answers more and more and more. And that's becoming my standard. And I think a lot of people last year were really curious of like, what is Google going to do? Are they going to eat their own business? Like, how's it all going to turn out? And so beyond just like, you know, talking about Nano Banana and their new model developments, they're doing a really great job of integrating it into their existing products, which kudos to Google because that's no small feat.

35:17Jon Krohn:And something that I've mentioned on air, have I mentioned this on air? I'm not 100 % sure. I think I have. I've definitely talked about this in my personal life, my professional life personally, is that the other thing about Google that I think they have a big advantage on relative to some of the other frontier labs like OpenAI and Anthropic is that a lot of people already trust Google with their Google Office Suite, Google Drive, Gmail. You already have a lot of data in there. And so, you know, for me personally, you know, I have the most cutting edge models and subscriptions from all three of those providers.

35:57and I use them for different purposes,

36:03Jon Krohn:it doesn't make sense to me to connect Anthropic or OpenAI to my Google Drive and give all of that information as access when I already have a cutting edge frontier model that does. Why should I give myself that extra risk? Exactly. And I think back to, yes, they took their time And I think it was really important for them to think about how it integrates with their existing suite. I will say that's my biggest regret was starting my new company is with Women in Data, we use G Suite. And with HMCI, I switched over to Microsoft. Don't do it. Don't do it. If I can make any recommendation, I should have stayed with G Suite.

36:47So that's my biggest regret in the new company is like, why did we switch over and start on Microsoft? So maybe it's not too late.

36:55Jon Krohn:Are you doing Teams meetings? We are doing, yes, it's horrible. I know, I know, what was I thinking? I thought I'd be cool and try something new. I wasn't smart, but speaking of Google, I would highly recommend if you haven't watched it, the Thinking Game documentary. So it is on, I know it is on Prime, but it's about just DeepMind, and it starts from the beginning days when they started the lab, and then Google taking it over. So, um, yeah, if you really want to get them again, I must be into history lately because if you want a history on deep mind, it's a great documentary. It looks like it is available in full on YouTube for free.

37:39Amazing.

37:40Jon Krohn:Which doesn't surprise me because it seems like it's a Google product to have created this film. So yeah, it looks like, uh, yeah, 10 days ago at the time of recording, it became available an hour and 24 minutes, 25 million views published by Google DeepMind. I will have a link to that in the show notes so that anyone anywhere can watch it for free. All right. Great recommendation there. So yeah, so comeback of the year. Congratulations again, Google. Maybe next year it will be somebody else. How much, how many times can you come back? All right. Disappointment of the year. it's interesting because you earlier in this episode you used a company name and the word disappointment in the same sentence so is that this one may be controversial i think you're gonna disagree with me on this john but i'll explain myself and so i'm just gonna say for me actually agents were disappointing and i wrote a substech this year called agents of disappointment.

38:48Take her off the air.

38:49Jon Krohn:Take her off the air. Where's the abort button? If all of a sudden I get muted, you'll know why, right? All right. And that brings us to the end of the episode. Yes, exactly. No, so I wrote a post this year as my best performing stuff called agents of disappointment. But really what I was talking about was the divide between like the hype of agents and really where I saw this come into disappointment was from the enterprise companies. You know, we have the obviously Salesforce has been talking about its agents force for some time. You have SAP. You have all of these enterprise companies who have been talking about how to implement agents into your existing enterprise tools.

39:30And it it just doesn't work right. Like or at least a lot of companies aren't set up for it properly or in my mind, don't know how to think about how to structure them properly and what to have agents do. And so from that standpoint, from like an enterprise agent standpoint, I think the hype and the practicality of the implementation, the divide between those two was too great. And so that was my disappointment of 2025.

40:00Jon Krohn:I totally get it. And I don't disagree with you. It makes a lot of sense to me. There's too much talk about agentic AI relative to the impact that it's making. No question. I do think that a lot of it is related to people not having their data silos set up in a way or, you know, their security set up in a way where they're comfortable with it. But, yeah, a lot of tinkering with agents, not nearly as many enterprise deployments. But I do think it will come. That is not my disappointment of the year. My disappointment of the year is Apple. I think you were disappointed with Apple last year. We got to go back.

40:41Let's look.

40:42Jon Krohn:Oh. I think, yeah, because Apple, did they announce Apple intelligence last year? Was that a this year thing? I don't know. Time is weird in the AI world. No, I think you're right. I think they announced Apple intelligence in the autumn, Northern Hemisphere autumn of last year. And it was disappointing. But I mean, I guess it's kind of just like Google. It's like - And it's still disappointing. It's still, because it's like a year later. And what can I do additionally with AI in my phone that I actually use? not much. I once made an emoji and then didn't do it again. You know, just like used it built in Gen.ai to send an emoji over iMessage.

41:25Jon Krohn:But I don't know. I don't really need to do that. Like there's, yeah, it seems like things like just having Siri be able to understand what I'm saying to it in the same way that opening iWhisper can. I mean... You know who's actually even done better than that is, so I will use Grok in my Tesla. And so when I'm coming home from work and want to like brainstorm something or just learn a new subject, like I just push a button and it works seamlessly and the chat mode goes back and forth, right? And I'm like, okay, if I can do that in my car, why can't I do it on my phone? that seamlessly. So I hope that like next year, our comeback of the year is Apple because, you know, they've been in the disappointment quarter for far too long.

42:14Yeah.

42:14Jon Krohn:And they have a lot of potential because of how ubiquitous their devices are. There's a big opportunity for Apple if they can get it right. Now I've got to, I've also got to talk about Grok and XAI because I never talk about them on air or very rarely, not nearly as much as they deserve, given that they have come to the frontier. I mean, they, well, we haven't talked about our overall winner yet. So maybe, maybe XAI is your overall winner. So maybe I should save it. But I just want to say that, you know, you mentioning Grok, I even did recently, I did an episode on, um, on, on Google Gemini three pro and how that at the time of the episode coming out, at the end of November, it was the top model across most categories as well as overall on the LM Arena.

43:06Jon Krohn:And so, you know, I kind of in that episode make the case that it's the best model. But in terms of ELO score, statistically, even though it had a higher ELO score overall than Grock in LM Arena, the difference in the ELO scores wasn't enough for LM Arena to call that statistically significant. And so technically Gemini 3 Pro is actually in a tie with Grok, but it just, it kind of, it made my story, it dragged out, like I couldn't make the episode tight and about how, uh, how Google had been losing to open AI and then later Anthropic. And it just became too complex if I threw XAI and Grok in the mix, but yeah, they've really, they've done a great job.

43:50I think what's impressive. I have a slide that I've been talking about for years and I just go I think it's um rate my iq.org and it rates all the ai models in a curve and so I have screenshots from like 2023 to 2025 to now and it's just so fun to see like how few models there were and just how pretty much everybody has gotten up to like the top tier you know not even like the paid private models but even I mean like this isn't my overall winner, but again, like what's been happening in open source and what came out with DeepSeek, I mean, it's, it's really incredible what access we have now for even free.

44:33Jon Krohn:Yeah. It really does make you wonder with all these, you know, these crazy amounts of money that need to go in to create a frontier model at the next capability level. And then six months later, you have an open source model that can do the same thing does raise lots of questions about monetization. Uh, yeah. Especially if you don't have clear distribution channels at vast scales like Google does. Yes, yes, yes. Interesting indeed. But let's not make the audience wait any longer. It's time for our overall winner of 2025 to be crowned. Oh my God. The audience is losing their minds. Wow. They're out of control.

45:08Jon Krohn:Oh, wow. Um, okay. I don't know. We have no, I have no idea what you're going to say. Um, I don't, I'm not even a hundred percent sure I have a clear winner. So maybe we can, maybe we can kind of work together to decide on an overall winner between us. Do you want me to say what I'm thinking is my overall winners or do you want to go ahead? I have one, but so I've, I've two firms that I have as my overall winner. And so one of them that we've already talked about a lot is Google, but the other one, which, you know, know you've talked about them not too long ago in your book recommendation the thing that is powering all these am not not all of them i guess technically because google does have tpus but nvidia yeah nvidia is the big winner in ai they're the overall winner i mean look at their share price oh my gosh it's great it's crazy it's bonkers so they truly i would say i mean in the ai industry, they truly are just the overall winner in AI, right?

46:13Like that is it, right?

46:14Jon Krohn:Because all roads point back to NVIDIA. So I would say they are the overall winner and will continue to be for some time. I guess my approach to who the overall winner was is I really saw this year as the year of vibes. Like to me, this was the year vibe coding got coined as a term, you know, Where I do think agents were and are really useful is in coding. I think that's the one place that we can see them just take off and do their thing. So for me, maybe it's not the overall winner I would agree with you is NVIDIA. Maybe my theme, I'm going to add in an extra category into this. 2026, we're getting a new category here.

47:02What the theme of the year was. And to me, this was such a vibey year. Like it was just like everything was about vibes. You know, I didn't we didn't have like new model breakthroughs like GPT five was disappointing. I don't know of anything that people are like anticipating coming out. Like for the previous years, people are like, wait till we get multimodal and then wait till we get agents. I don't know what people are anticipating really next. Like it's just all about vibes. And so that's my thing.

47:33Jon Krohn:I like that. Makes a lot of sense. All right. our joint winners are NVIDIA and Vibes. Yeah, no, I totally get it. It makes a lot of sense. And also to go into a little bit more detail, a lot of our listeners might already know this, and you definitely know this, Sadie, but the reason why coding is something that is so great to be doing with LLMs is because it's something that it's really easy to train LLMs to be good at at a really high level because you know if the code executes, it works. You just kind of have the answer. So it's easier to simulate multi-step, very long problems. And that's why earlier when I was talking about the length of human task doubling that AI systems can do every seven months, that is only on these kinds of tasks like computer science or machine learning tasks, where you have a very clear sense of whether the thing worked and whether you can simulate data, simulate training data that allow you to continuously expand the length of what these models are doing for a lot of real world tasks in most industries to collect those data is just so crazy expensive to have humans creating those data that yes, it's happening, but way, way, way more slowly than those areas where we can simulate the data really effectively.

49:01Jon Krohn:Alrighty. Okay. So those are our winners. It's time for the predictions. What's going to happen in 2026, Sadie? I believe that like last year, you have five categories for us. I do. I have five themes and it's not 2025, but oh well, here we are. So number one, I think we're going to see, again, I know we're going to AGI and this is going to make it sound of like we're going back to narrow AI, but I think more like specialized industry models like AlphaFold. So I think we've reached the limit for now until we build out some of these hyperscaler clusters, but really the limit on data for training these like general purpose models.

49:45And, you know, back to your comment on why we see agents and LLMs do so well in coding is because we can actually measure it and see that it works. We need to have more specialized models for particular domains and then have domain experts evaluate those. And so I think we're going to start to see more development of specialized models for particular domains. And you could think of them as like mini models, but the way I think of it is like we have human intelligence, right and there's like a general intelligence that we all have that get us through day-to-day life and then we have people who go and get PhDs in particular areas or just our savants at a particular area and I think that we while for some reason nature has selected that path for us that we each have our own specialization I think we'll start to see that happen now in model development as well and so I'm interested to see how this unfolds but I think that's how we're going to make progress is getting more specialized, kind of like expect more alpha fold moments in particular domains.

50:54Jon Krohn:Nice. I like that. And it is interesting that you bring up the alpha fold example. It may not be coincidental, but that is an example of, while very narrow, artificial super intelligence because it's a humans, no matter how hard we try, you can't look at a sequence of amino acids and predict what a three-dimensional protein will look like. We just can't do it. But AlphaFold can, not for every kind of protein, but for a very broad range of proteins with remarkable accuracy. And yeah, so artificial superintelligence, yeah. So these kinds of very specialized industry models could become more and more prominent.

51:35Jon Krohn:We could see more and more examples like AlphaFold where you know, in these very narrow niches, a machine vastly outperforms human intelligence capabilities. But think about it. How do we get to like an overall general super intelligence? Probably by combining these mini models of particular domains together to get there. But right now, one of the hardest things about some of these general intelligence models is really being able to test them fully in their subdomain. And so I think that it's just splitting it up, maybe half of it is the training and then half of it is the testing in a way that we can actually evaluate them more properly.

52:11So nice.

52:11Jon Krohn:Makes a lot of sense, Sadie. All right. What's number two? Number two is continual and nested learning in models. So this is a new paper that just came out from Google. It is called nested learning, hence the prediction from it again. But back to Google being hot this year and last year and coming back with some of their research. I'm really interested in this paper because one of the issues that we have today is we train a model and, you know, other than it remembering what's in our context window, it's not continuing to update and learn for me. And so nested learning is a way to bypass that and get around it.

52:53And so I think we're going to see a lot more progress in this space with how do we get models to continue to update and continue to learn? Because that's what makes humans really great about being able to learn and grow and what makes our intelligence so strong. So why wouldn't we want that in a model as well?

53:11Jon Krohn:Makes a lot of sense. Really big innovation there. Yeah. That continuous learning is a big gap in most of the, you know, certainly all of the big LLMs from the frontier labs, they don't continuously learn. It's a huge exercise and yeah, that is a function of biological intelligence is continuous learning. So pretty cool. I like that one a lot. What's your number three prediction for 2026, Sadie? Number three, and it's my bias is coming in, but we're going to get back to research. So I don't think that any lab right now, any frontier lab has a clear path for what is going to be our next big breakthrough in AI.

53:54Again, back to my comment earlier, you know, before we kind of saw the roadmap, oh, wait till we get a model that's multimodal. Wait till GPT-5 comes out. I don't hear anybody talking about GPT-6. I don't hear anybody really talking about like what's the next thing they're really waiting for from an AI model. and I think that we have to get back into labs to discover it and I think we're waiting to have a clearer picture of what that may or may not look like and a lot of this came comes from the the recent interview with Ilya on the Darkesh podcast and just his estimates too that there's still a lot of labs that don't know exactly what that path forward looks like.

54:43And I think that we're going to run into some limitations with some of the hyperscalers, not necessarily from a compute standpoint, but from a data availability standpoint. And so it's time to get creative and get back to research. And I think is really just an exciting time because it means that new ideas are welcome. And so it's a good time to be in the space.

55:09Jon Krohn:Makes a lot of sense. Yeah. I mean, we, for a long time, the de facto scale that we were scaling up on was number of weights in a model, you know, 10Xing that, 100Xing it, you were getting magical capability improvements, you know, as notably done from GPT-2 to 3 to 4. and yeah and then more recently like in 2025 there was a lot of well 2024 even more so there's a lot of excitement about scaling inference time and so you know how long do these reasoning models reason for and those you know reasoning times did expand a lot in 2025 and so we saw a really impressive of results on math and chemistry, Olympiad kind of results.

55:58Jon Krohn:And just in general, these kinds of supposedly very hard benchmarks like humanities, the last exam have started to look tractable thanks to long inference times. But I think you're right. I think that some kind of orthogonal breakthrough beyond just scaling has got to get us to the next level. And I think continuous learning is potentially part of that. And I don't think this is going to happen in 2026. If it does, that would be a huge breakthrough. It would be something that allows algorithms to learn in a much more sample efficient way. So humans, even infants can learn from one example or even infer something from zero examples.

56:50Jon Krohn:So a child who isn't even old enough to speak but can kind of walk around, if there's an adult carrying a bunch of heavy looking objects and they're kind of walking into a closet door, that child can't even speak but will open the closet door. for the adult. You're just, the child is able to infer the intentions of the adult in a scenario that the child has never seen before. And so that kind of zero shot learning or one shot learning or few shot learning, you know, we have those terms in AI, but those terms only apply when you're kind of providing examples in a context after the LLM has learned from billions and billions of examples.

57:41Jon Krohn:And so, yeah, I think some kind of breakthrough that allows for way more sample efficient learning is critical. You must be peeping at the next prediction. Oh, yeah. Well, I have no idea what they are. Seriously, for our audience, I have no idea what Sadie's going to say next. All right. Number four. Sadie, what is it? Yeah. So I know there's been a lot of talk about robotics and I keep saying this prediction that You bought a new robot, didn't you? Yes. So it's supposed to come in 2026. I don't have a date, but I know this year when this episode comes out. So I'm very excited to take my robot for a walk.

58:22Jon Krohn:Wild. But I don't think that 2026 will be the year of the robot. I still am standing true to what I say, which is 2027 will be the year of the robot. But I do believe we will see more physical AI and spatial intelligence and particularly spatial intelligence, because I think we need to branch off into new data sets. And so, you know, things like world labs and simulation of environments. Right. I think that's going to be a new playground that we're going to explore as a new data set. And so really looking at how do we bridge the gap between real physical AI and getting a robot to work in space?

59:09Again, this is where the continual learning will come into with these models as well, right? Because when you go into a 3D environment, you're encountering a lot of things that, as all of us know, with driving or walking and different things that we have to continually update our model for. And so I see physical AI, spatial intelligence being a space that will be really popular, but mainly from the standpoint of just helping us to collect new and harder data that we don't have today that can finally help bring our AIs into the world and really start to expand the use of these models. Nice.

59:50Jon Krohn:I love that. I actually, it was after we had recorded last year's episode, but before we'd released it, I was at NeurIPS in Vancouver and Fei-Fei Li was one of the keynotes and like thousands and thousands and thousands of people, a huge auditorium packed full of people watched her talk about her company, World Labs, that you're describing right now. And this is, you know, It's very expensive to collect these big real-world data sets, but absolutely essential to be able to be training the machines of the future. Because right now, all these frontier LLMs, they are capable of being helpful inside of a computer.

1:00:32Jon Krohn:But if you want some robotics application, some real world spatial application, yeah, the hard work of collecting all those data and getting the machines going is key in getting your NEO able to water all your plants, if I understand that's a key application you were looking for from your robot. Yes, sir. I have a plant wall at the office. Well, actually two plant walls. And so I do need somebody to take care of these plants, which will be great. Nice. All right. And then fifth and final prediction for 2026, Sadie, what is it? Yes, drum roll, please. And I love to bring it to practical application for work, which is what will be the kind of new hot trending job.

1:01:17And I think we're going to start to see AI ops become a thing. So AI operations, think of this as like how are somebody who manages the GPU management model orchestration, right? Agent reliability, thinking about, think of it from like a function of DevOps in 2010 to, you know, the AI ops of 2025, 26 era. And so I think this is a job description that we'll start to see pop up. I don't know if it would be what I call the most popular one, but I think it's going to be a new trend that we'll see emerge here in 2026. Nice.

1:01:58Jon Krohn:I like that one as well. AIOps, quite practical for all of our hands-on practitioners out there listening. So yeah, so to recap, specialized industry models, continual and nested learning, research on the next big breakthrough, spatial intelligence, and AIOps. Sadie, thank you so much yet again for doing a predictions episode. I hope that I can wrangle you to do it again next year. I always enjoy this episode so much. Please, absolutely. That would be great. Maybe we can even do that partway through the year when you have your robot. Maybe I can be in California with you and I don't know. Yes, come out to our new studio.

1:02:42Once we get it set up, Neo will let you in the door.

1:02:46Jon Krohn:So cool. I can't wait to see it. I'm so jealous. Yes. And yeah. And so, as you know, I always ask for a book recommendation at the end of every episode. You already gave one in this episode, The Thinking Machine, Jensen Huang, NVIDIA, and the World's Most Coveted Microchip. Incidentally, turned out to be my overall winner of 2025. Do you have any other book recommendation for us or do you want to go with that one? I'm going to go with that one and then my book, Becoming an AI Orchestrator. Of course. Of course. Yes. They just pair really, really well as a Christmas gift. Yes. You know, those will be great together.

1:03:18You know, the other one I would say, I'll add a bonus one. The Innovator's Dilemma, I feel like, is a really good one right now, tying us back to the beginning of the episode where you talked about intelligence is on this exponential curve, and we mentioned that we need to rethink business models. I think that book does just a great job of highlighting how once you are an established business, it is difficult to reinvent your business. And so I think it's a really just a relevant book for everybody right now from a standpoint of how do we rethink our businesses with AI.

1:03:53Jon Krohn:Nice. I like that recommendation as well. Seems like something I need to be reading. Gosh, I wish I could read all the books that people recommend on this show. That seems like a really useful one for me to be able to sink my teeth into because it is a dilemma that I face all the time. All right. And yeah, of course, your book, I'll mention it by its full title again, Becoming an AI Orchestrator, a Business Professional's Guide to Leading, Creating, and Thriving in the Age of Intelligence on Bookshelves Now. Check it out, people. Get it. And other than your book, where else should people be following you, subscribing to you going forward, Sadie?

1:04:29You know, I'm having a lot of fun on Substack this year, mainly because it's direct to people's inbox. And I just started my YouTube channel doing weekly videos on just like deep dives in my favorite tools in AI. So I have one coming out on Gemini Pro this next week. So it's just, I'm having a lot of fun with those two platforms right now.

1:04:52Jon Krohn:Nice. Substack and YouTube. And I imagine people should be following you on LinkedIn. Of course. The good old streets of LinkedIn, come say hello. Just don't throw your bot on my comments. Yeah, I block you. You do that to me. You're blocked forever. See ya. All right, Sadie, thank you so much. I'm looking forward to an exciting 2026 in AI and robotics and consciousness research and book releases. So much fun sharing this time with you again, as always. It's always a great time. So happy Happy New Year everyone, and here's to another great year. Five years in a row, and I continue to love these annual Look Ahead Predictions episodes with Sadie St.

1:05:38Jon Krohn:Lawrence. In today's episode, we covered how her five predictions for 2025 largely panned out. Agentic AI dominated, AI integrated into major everyday devices like AirPods, scientific research expanded with AI, enterprise monetization remained crucial, and AI engineering skills overtook traditional data science skills in demand. My wow moment of 2025 was watching a crew of AI agents build a functional stock trading platform in 90 minutes. For Sadie, it was using Lovable to generate her entire hmci.ai website with a single prompt. We awarded Google a comeback of the year again for becoming the Frontier AI Lab to beat for the first time in years thanks to its Gemini 3 Pro and Nanobanana Pro models, while in Sadie's book, Vibing, was the overall winner of 2025.

1:06:27Jon Krohn:In terms of Sadie's predictions for 2026, number one was that we'll see more specialized industry models emerge, think AlphaFold-style breakthroughs in specific domains rather than big general-purpose models, partly because we've hit data limits for training massive generalist systems. Number two is that continual and nested learning will advance significantly, allowing models to keep updating and learning rather than remaining frozen after training, a key gap between current AI and human intelligence. Her third prediction is that labs will return their focus to fundamental research because scaling, like model weights, inference time, etc., is no longer the clear provider of the next big AI breakthrough.

1:07:04Jon Krohn:Number four is that spatial intelligence and physical AI will gain momentum as researchers explore 3D environments and simulation as new data sources to bring AI into the real world. and fifth and finally, AIOps will emerge as a hot new job category. Think DevOps, but for managing GPU infrastructure, model orchestration, and agent reliability. 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 Sadie's social media profiles, as well as my own at superdatascience.com slash 955. Thanks to everyone on the Super Data Science Podcast Team, Podcast Manager Tony Breivich, Media Editor Mario Pombo, Partnerships Manager Natalie Zajski, Researcher Serge Massis, Writer Dr.

1:07:49Jon Krohn:Zara Karche, and our founder Kirill Aromenko. Thanks to all of them for producing another fantastic episode for us today to kick off the year. For enabling that super team to create this free podcast for you, we're deeply grateful to our sponsors. They and you are what allowed this show to happen alongside us as a team. So consider checking out our sponsors' links in the show notes to support the show. If you ever want to support the show directly by sponsoring it yourself, you can get the details on how by making your way to johnkrone.com slash podcast. Otherwise, support us by sharing this podcast episode with folks who would love to hear about it, review it on your favorite podcasting app or YouTube, subscribe.

1:08:29Jon Krohn:But most importantly, just keep on tuning in. I'm so grateful to have you listening and hope I can continue to make episodes you love for years and years to come. 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.