AI in 2026: Reid Hoffman’s Predictions on Agents, Work, and Creation

What we will see more of in 2026 is a combination of parallelization, longer workflows, and orchestration. People will experience what it is to have their computer running separately from them, doing something productive for them as they're walking away to go get their coffee. Whether it's a Mac Mini running cloud code or codex. For a company to be a thriving, going, and growing concern, and evolving with the times, you will need to be recording every single meeting and using agents on it to amplify your work process. Reid, welcome to the show. It's great to be back. And as much as I try to avoid doing predictions, you're one of the few people that I will essay this with.

That is, I feel very blessed. Thank you for taking the time to do it with me. I think this is your third appearance on this podcast, and that makes you the most frequent guest. So I'm honored.

Feeling is mutual. I'm honored. Yeah, yeah. Okay. So we're heading into 2026. By the time this podcast comes out, it will be 2026. So for all of our purposes, it's 2026. And I think this time of year is such a good time to look back and look forward. So I want to start with a couple of pre-2026 predictions that you made and reflect a little bit on how things went in 2025 and what might be different about how you're seeing things. So the first one is we dug up a quote from you in 2017 that said you thought that the nine-to-five work model will be extinct by 2034. Where did that view come from? And how has that changed in 2025 as we've moved into agentic territory?

Well, let's see. So part of it was an extension of a very old set of thoughts of mine, which is a startup view, which is more and more of work and more and more of career will become entrepreneurial. It doesn't mean that everyone is going to start companies or everyone's going to launch new products or any of that. But it does mean that the kind of old career ladder, career escalator is no longer the way to think about it. It's no longer to be thinking about what color is your parachute and that kind of thing. It's actually to be thinking about your economic life, your work life, your job life as kind of the skills of an entrepreneur. And that's part of where that came from. And it wasn't meant to be nine-to-five is everyone's going to be working 996 or equivalent. Which would have been a good prediction maybe for Silicon Valley.

Yes, exactly.

But it's – and by the way, startups in Silicon Valley have always worked 996. It even, frankly, 997 for how they operate. And – but it's more the fact that actually, in fact, the way that you're going to be working isn't going to be this kind of clock here, hit your punch card at the door, be there, take your lunch break, come out at five. But actually, in fact, going to be, you know, running Claude Code on minis in parallel to what you're doing. You're going to be in a crunch where you're doing stuff and all of a sudden this week you're doing 120-hour week. And the next week you might be doing 40, or 10 as the case may be. And then this kind of entrepreneurial journey is actually more of what's going to be happening. And I think that we're still on track for. You know, here we are and 25 going to 26 is time of broadcast 26. If anything, when you begin to see what the impacts of the fact that we are going to be all of our work is going to be enmeshed in agents and in parallel and in management and all these, which we'll get into some depth on. That actually, I think, is part and parcel of it's not just nine to five.

Got it. So I think when I read that quote, I was thinking it's not going to be nine to five, meaning we might not be working that much. But you're saying it's more just an entrepreneurial way of working where it's suffused throughout your life. Exactly. And that actually makes me think. And by the way, that also can be in some cases, you're just not working as much. I mean, it's much higher range. If you're Tim Ferriss. Yes. Yes, he's already been doing that already. I know, right? Yeah. He's got to do a new four-hour work week. Yeah. The future is already here. It's just unevenly distributed.

That actually makes me think of one of my hot takes for 2026. So I think we can jump there real quick because I really want to know what you think. We've been on this trajectory of talking about technology and addicting technologies and social media and how social media breaks your brain. And I think we've put on this pedestal the act of creating things as something that can be inherently good and not necessarily addicting. And my experience with Claude Code right now is I'm addicted to it. It's, I cannot stop. I just want one more prompt. And I think, shockingly, the most addictive technology of 2026 and the narrative that we might be talking about at the end of the year is how addicting it is to just make things. And what's interesting is there's a certain class of people that know that already. And it's CEOs of startups who have that experience already because you're always looking at your chat, your Discord, your Slack or whatever. And you're always, oh my God, I need to do something else. But I think now that's a broadly distributed thing where everyone's just going to be prompting Claude Code.

So to one, I definitely believe it can be addicting. And I think it's actually addicting for a much broader range of people than normally think. And it's partially because most people just don't have the experience of succeeding at creating. And once you have that, the dopamine hit is you succeed at creating. And part of the thing that Claude Code, actually AI more generally, general AI more generally, it suddenly goes, oh my God, I can create something interesting. And I think that's the thing. It's actually a healthy dopamine hit. I mean, one of the things that's weird about the word addiction is you say, well, I'm addicted to breathing. And so it's, well, actually, in fact, that's a good thing. But it's not. And so addiction has this kind of negative overlay. But the fact that you get very committed to something, it's, oh, is it unhealthy for you? And actually, in fact, in the creation thing, actually, in fact, it's not unhealthy. And if you're, no, no, I'm actually going a little bit more obsessive. I'm going, I want to finish this. I want to make this. I want to make this really great. That's actually, in fact, part of where we explore our fuller potentials, our super agency, if you will. And it's that kind of thing that I think is actually really good. And I do think that it's part of the kind of generative AI revolution in ways that people go, the discourse is generally quite mixed and negative. And actually will be more intensely negative next year because of the transformations and changes. But it's part of the reason why it's so important for people to go, wait a minute, I can be so much more human doing this. And we can be collectively together. And so we need to sort out the fact that, yes, it's going to be a turbulently created future. But we can do amazing things. And so I think that kind of creative addiction, creative commitment, creative exploration is actually one of the really important things. And I think people have been discovering it not just with cloud code, but also through prompting these agents, creating images. And that's part of the reason why Sora went to the moon in a couple weeks because it's, well, wait, I can make something here?

I think that makes sense. I kind of want to know for one thing you said earlier is you think that there's going to be sort of a backlash, but negative sentiment towards tech will increase in 2026. Is that one of your big hypotheses? So tell me about that. And so I think that creative commitment, creative exploration is actually one of the really important things. And I think people have been discovering it not just with cloud code, but also learning through prompting these agents, creating images. And that's part of the reason why Sora went to the moon in a couple weeks because it's like, well, wait, I can make something here? I think that makes sense.

I want to know for one thing you said earlier is you think that there's going to be a backlash, or negative sentiment towards tech will increase in 2026. Is that one of your big hypotheses?

So tell me about that. So we haven't, while there's been a lot of discussion, the actual overall impacts of AI have been relatively more minimally felt. And most of the places where they're described as being felt are actually fictional. For example, AI is causing electricity prices to rise. And really, a little bit here and there, maybe in certain grids, certain power stations. But really it's old grids, old power stations, increasing cost of energy, net impact of tariffs and other kinds of things. If you actually do an analytic map to say, where are the data centers? That doesn't actually correlate to where electricity prices got up and not.

But that's going to be the meme. And so the meme is college students aren't going to be hired because of AI. The meme is electricity prices are going up because of AI. The meme is the price of eggs is going up because of AI. And so because there's a lot of people who go, I'm looking around for something to blame for things being troubled, bad, different than I would like. And it is going to be a very turbulent year. And so AI, it's going to be almost like the Farmer McDonald song. AI is going to be the way that this is going to play. And I think it's actually really important for people to understand it.

AI hasn't had any of that impact yet, but it's actually going to start. I think it's suddenly going to be like, hey, I used to be really competent at my marketing job, etc. I think it'll be, hey, I only want to hire when it's part of an AI transformation, a la Shopify and that kind of thing. It isn't going to be what a lot of the employment is. It's a reworking of the COVID disaster and misorganization, etc. But actually, it is going to start impacting. And then so it moves from 98, 99 percent fictional to 90 percent fictional. But that will intensify the desire to say a whole bunch of negative things.

I've been surprised so far at the creation of a Christmas record for my friends using AI. I haven't gotten a whole bunch of negative blowback of, oh, this is going to be terrible for artists and terrible for creatives and so forth. I think that will happen. I'm going to create some more records and I think that will be the case. And I actually think it's not the case. I just think you need to adjust to using it and creating that as a new basis for your creativity, for your industry, for your work. And that transition is going to be what's difficult. But I think next year is going to be much more negative on AI than actually this year in general popular discourse.

So to repeat that back, I think you're saying so far it's a meme like AI is bad. And the meme to a large extent is making a scapegoat for just anything bad. If you're laying people off, it's easy to say, because of AI, and that will probably continue. I do think that's true. That's just going to continue. And there will be increasing real negative impacts that people are going to have to deal with. So you're a programmer and you're coming into work and your job just totally changed.

You're not in the code anymore. And that's going to be upsetting to people. And that's going to lead to changes in the way organizations are run and who gets hired and all that kind of stuff. And that's going to make people upset. What do you think is the right move for big AI companies in an environment like that and how they should be talking about it, how they should be positioning? And to some extent, it's probably not even desirable to prevent any kind of backlash. It's normal for people to have bad feelings about new things. But how do you think what's the right way to deal with that strategically?

Well, the most substantive way is to make it pragmatically helpful to as many people as you can. It's part of the reason why podcasts you and I are doing and other things to say, hey, explore it, get a chance, use it. You can use it for personal things. If you have any kind of serious medical question, if you're not getting a second opinion from chat GPT or your favorite frontier model, you and your doctor are both making mistakes. And similarly, OK, how do I use it to help me with my work?

How do I use it to help me learn things? How do I help it help me be creative? And if you can't in each of those areas find something where it's actually seriously helpful, you're not trying hard enough. You're not looking. It doesn't mean it's everything. It's not the Swiss Army knife for everything yet. There are many limitations. But it is enormously amplifying. And so I think that's the reason why, everything from not just writing super agent, but creating holiday Christmas gift records is showing, hey, that's the kind of thing we can do now. Like everyone can do this. There's not using any tools that do this.

And by the way, not only can everyone do it, but as the people who get more expert, like people who are much better at music than I am, which is 95 percent of the human race, can then do much better. It's an amplifier for everybody. And I think that's the most substantive thing. And then on the communications thing, I think one of the things that various well meaning AI creators are saying is, oh, my God, it's going to be a white collar bloodbath, etc. And you're like, well, I have one person in mind that you're talking about. Yeah. And it's like, look, I get it. You're trying to say, hey, guys, things are going to change a whole lot. Really pay attention.

I'm ringing a bell to start adjusting to this. But that kind of ringing the bell is yelling fire in the movie theater. It doesn't create productive response. The important thing is to orient towards a productive response. It doesn't mean papering over the difficulties of transition. But it's, oh, you know, we're going to be going into intense rapids. And here's the paddles you need. And here's the thing you should be doing as you're going into it. That's what you want to do. If you're going to say we're going into the rapids, you want to offer the paddles. And if you're just saying we're going to the rapids, that's not really helpful in my view. Yes, exactly.

I'm ringing a bell to start adjusting to this. But that ringing of the bell is like yelling fire in a movie theater. It doesn't create a productive response. The important thing is to orient towards a productive response. That doesn't mean papering over the difficulties of transition. But it's like, we're going to be going into this intense, category 10 rapids. And here's the paddles you need. And here's what you should be doing as you're going into it. That's what you want. If you're going to say we're going into the rapids, you want to offer the paddles. And if you're just saying we're going into the rapids, that's not really helpful in my view.

Yes, exactly. And I think that's the communications part of it for everybody. If 2025 was the year of agents, what's 2026? Well, I don't think actually 2025 was fully the year of agents. There's a lot of agent development, but I think it was mostly agents and code. Cloud code, Codex, et cetera. Of which a relatively very small percentage of humanity actually fully experienced. If you go to the vast majority of people you and I know, they're like, you mean agents? All I did was ask ChatGPT a few questions and had dialogue. Well, no, that's not actually agency. That's a chatbot, not really agents. Agents is doing stuff and doing it in parallel and doing it in amplification.

So I think code had that. But I think in 2026, we'll move from this basis of agentic coding agents to agents in everything else. And I think 10 to 100 times more people will experience what it is to have their computer running separately from them, doing something productive as they're walking away to get coffee and then coming back. Whether it's Claude minis, Mac minis running Claude code or Codex, different questions. But that orchestration allows the parallel work, allows eight hours of work, that kind of thing. I think that will be broader.

And then the more subtle thing, which I think will be really important for 2026, is orchestration. If we have agents working with me, for me, and I'm orchestrating them, I think orchestration is the thing that will be important. I don't think it'll be March 2026. I think it'll be more Q4 2026 or growing into that and maybe intensively 2027.

I totally agree. I think it's something we're starting to see already. And it brings me to perhaps my hottest take, and I'd love your input. It starts with coding agents. I think OpenAI is currently missing the real coding market because they're not, when you think about orchestration, I think of it as something enabled by tools, but also a new skill for programmers.

When I look at what OpenAI is producing, I think it's made for programmers who use AI—senior engineers who use AI—which is different from AI-native engineers who are just in cloud code terminals and never looking at the code. Their models are really good. If I have a hard technical challenge, I definitely go to Codex. But I don't see them orienting toward this new skill. It's not vibe coding, but it's not traditional engineering with AI added. It's a third thing. I have four cloud tabs open. I never look at the code. I'm thinking about how to orchestrate, how to plan. I'm doing all this stuff. And I'm technical, so I could go down to the code, but I never do. I think that's interesting, and OpenAI is not used to being behind. I'm curious how that's going to play out.

Well, I think it's a skill that OpenAI is going to pick up because each month it'll be like, "Oh, Opus 4.5. Oh, GBD Codex. Oh, Gemini." All of them are going to be developing. What that means structurally is that some areas where, a couple years ago it was just OpenAI blazing ahead, now there'll be areas where Anthropic did super smart stuff making Cloud Code, and that iteration took less capital and less compute depth but still made amazing stuff.

OpenAI will do this. This is one of the benefits of competition for the industry and society. They'll say, okay, we can't be behind on this. We need to learn to do this. And I think that's what will happen. Competition is frequently painful as you push forward, but I think that's the end result. Credit to Anthropic, the notion of focusing on code is not just a code product, but an amplification of many other things—AI progress, development, and every other form of information and knowledge work, and maybe many more things. I think that's why every major player has to be capable at minimum in code, if not leading.

They got to general purpose agent architecture by making a great coding agent with the right primitives. If you look at the software developed over the last month since Opus 4.5 came out, pretty much every new thing is being built on Cloud Code. I built this entire end-to-end reading app. We have this AI paralegal that's been upgraded hugely. Every single app is just Cloud Code in a trench coat. Basically UI wired to a prompt with an agent that has tools that does what you want. It's the coolest way to build software because it's flexible. Users can modify it. It's exactly right. It's such a pleasure to see someone figure out those primitives.

And massive credit to the Anthropic team. And everyone else should be learning, building on top of it, iterating to the next generation. Do you have a thought on why Opus 4.5 is so good? I'm assuming you think it's good. I think it's the best model I've ever used. It's a crazy leap for me. I'm curious if you agree and if you do, do you have any thoughts on how they managed to do that? Well, I think it's amazingly good. I don't know if it's the everything model for me. I think to some degree, I think... # Cleaned Transcript And it's such a pleasure to see someone figure out those primitives.

Yep. And massive credit to the Anthropic team for doing that. And everyone else. Hey, you should be learning from it, building on top of it, trying to iterate to the next generation. Do you have a thought on why Opus 4.5 is so good? I'm assuming you think it's that good. I think it's the best model I've ever used. It's like this crazy leap for me. I'm curious if you agree. And if you do agree, do you have any thoughts on how they managed to do that?

Well, I think it's amazingly good. I don't know if it's the everything model for me. I think to some degree, GPT-5 Pro with Codex also is pretty amazing on a lot of levels. And by the way, Gemini 3 on science topics and so forth. So I still bring all three of them with me to various things I do.

That being said, I am very curious about how they pulled 4.5 together. And one of the mistakes that outsiders think is they think you just apply scale and press play on compute and some of it works and some of it doesn't. But actually there is both a lot of science and art to do it. And it's one of the reasons why Meta has needed to restart its AI efforts because you can't just throw a whole bunch of compute at it and it works. It has to relearn these things in terms of how it's playing.

So I think it'll be interesting because the techniques spread out very quickly. So I think we'll learn, but I actually don't know what the new genius is in Opus 4.5. Do you have any hypotheses? I have no idea. I think the only thing I can think of is recently we got a view of the underlying soul document of Claude. And the interesting thing that I feel from Opus—and I agree, I use ChatGPT as my daily driver—but when I'm building software, except for specific performance things or hard bugs, I'm using Opus as my daily driver.

I think there's usually this trade-off that you see with Codex, where the better it is at programming, the less empathetic it is. It feels a little bit more like a senior engineer. It's slightly more autistic or something. And Opus, they figured out how to make it both humanistic and understand users and what I might want and what I might mean and how interfaces work and what good interfaces look like. And it's a fantastic programmer. Something about the soul document where it tells it who you are and what you care about—it's one example of anthropic thinking about these things in a more holistic way to create a being rather than a tool. And I think that is actually going to be a big deal going forward.

You know, this is one of the things that Inflection started with, with EQ. And actually soul is a very natural extension. Inflection started, and there are still a lot of ways in which Pi is amongst the leading of having a richly textured conversation and agent-like focus on EQ as much as IQ. Not slouching on IQ, but putting the two together. And the soul document actually I think is the next evolution because this is what we learn and iterate. And it's part of what makes Claude Code work because it's actually a really good human amplifier and how do you operate that way? And then you get better performance if you can interact in that way.

So I think that's a good insight. I suspect there are other things. I think we both suspect there are other things. And we'll hopefully learn them in the next few months. That would be great. So last thing on the coding front. You mentioned the horse race earlier. Everyone's going to be trading volleys. But if we want to not be fooled by randomness and not track every little change, we hit the snooze button and come back at the end of 2026. Where do you see the landscape of who's winning in the coding agent race?

Well, I don't know who will be winning, but I predict strongly that the horses that are leading now will still be neck and neck. It'll be like in the first hundred meters, this one's a little ahead. Then the next hundred meters, that one's a little ahead. I don't think any of the horses in the race will particularly stumble.

So like, I thought Cursor was really fantastic, and it's just gone. I think none of them will stumble. Now what will be interesting is the folks who are not in this at all, like Apple, right. Despite the fact we use Macs for various things, the AI part of it is non-existent. I think the gap will be even more stunning—the fact that you actually haven't gotten what this coding amplification and everything else means. And I think that will be playing out more, but I think they'll all be in the mix.

The thing that will be interesting is not so much which one will have stumbled out, but I'm really curious about what are the one or two superstars that will really get in the mix more. Will Replit be more general? Will Lovable be more general? Or will it be something else? I mean, with some high probability, something will surprise us here. Yeah. I don't know what it'll be, but predicting surprise is hard.

I think that's interesting. One of the things I've been toying with is the stakes are so high and programming is such an obvious use case that is so economically valuable. It feels like everyone is in a knife fight for programming. And I wonder if there are surprises—you've been predicting AI will be used for more creative use cases for a while. I wonder if the surprise entrance comes from a place like that, where it's not actually about programming.

One caveat to that is like you said, Claude sort of invented this general agent by being good at programming. So it's hard to say exactly, but I wonder if that entrance is coming. It leaves them vulnerable to competitors from other places because they're just focusing on programming right now.

Well, I definitely think programming is part of the architecture for getting everything else. And for example, part of the reason coding is important is that even when we get to things like, how are you going to have a much better paralegal? I love what you're doing. Better medical assistant, better tutor, et cetera. I think coding will actually be not just the amplifier, but the fitness function of how do we kind of go, "Hey, this is getting better. This is amplifying the work better." Not just the foundations of coding—driving, planning, longer work, parallelization, orchestration—but also like how a better legal document works will actually come out of it. And I think some of that will also be in creative. It won't be surprising to me.

I love what you're doing among other things. Better medical assistant, better tutor, et cetera. I think coding will actually in fact be not just the amplifier of it, but the fitness function of how do we, how do we go, Hey, this is getting better. This is amplifying the work better, et cetera, that parallel, not just the foundations of coding, driving, planning, longer work, parallelization, orchestration, et cetera. But also, well, how does a better legal document work will actually in fact also be coming out of it. And I think some of that will also be in creative. Like it won't be surprising to me. Obviously everyone's trying to figure out, okay, how do we, well, not everyone, a number of people trying to figure out how do we take, you know, VO, Soro, et cetera, and then go, okay, can we create a 30 minute movie off it? And the coding pattern will be part of what happens there. And so it can be in those kinds of creative.

Now, obviously, some of the more interesting possible surprises are, well, cause there's a number of different efforts trying to do this too. Well, could we get raw ideation, like better at science. So we read a whole bunch of science papers and we can do scientific hypotheses. Now, by the way, you begin to say, well, maybe that'll also be true of AI research and ideas for doing this. And suddenly it's doing idea generation in this kind of thing. And that's definitely a whole bunch of projects trying to work at that.

So the notion of, Hey, if you can think a lot better, you can then apply that to this kind of creativity and this kind of new ideas. Those I think are much more speculative. It's an interesting hypothesis. There are people who will hold them saying, Hey, we've just seen that with scale learning and compute and it's going to happen. And I'm like, well, look, it's crazy that everyone smart should assign a non-zero hypothesis probability of that. Cause that's really amplifying. But on the other hand, I think it's like, yeah, it's not clear that we're yet seeing any of that. Even when you see people like Terrence Tao saying, Hey, I'm using generative AI to help me understand where I should be thinking in my math analysis. And yes, but I think a hundred percent, but of course, Terrence Tao is one of the most genius mathematicians of our age and is providing a ton of the metacognition in this. That makes sense.

Yeah, I think I'm trying to go back to your comment about no one stumbling and I'm wondering who would stumble if there was a stumble. And I think my current feeling is I would guess Cursor. Yeah, that's probably my highest likelihood. Not that they go away. They're obviously going to be a successful company, all that kind of stuff. But I think that they're caught a little bit in the same position that OpenAI is, but OpenAI has more flexibility here where Cursor—a lot of their business is built on traditional developers using IDEs inside of big companies with AI on the side. And they're sort of caught between that paradigm and this totally new to E cloud code type paradigm. And they kind of have to do both. And I think that's going to hamper their product direction and velocity in a way that I would bet in a couple of years we'll look back and be like, that was an interesting era. And it's still a widely distributed piece of software, but it's not the next generation thing that we thought it was.

That's a, I agree. And that's one of the reasons I brought it up. And the other one I've been thinking about is the hardest. And another angle of that is, you know, how are we going to be not just integrating the kind of the application functionality UI, but the underlying model and compute fabric capabilities. And Cursor is just beginning to do that kind of stuff. And what the shape of that is, is going to have to be dual targeted, like you mentioned, or multi-targeted. I think it's a harder slalom race for them.

I think the narrative right now is that enterprise AI deployments are not doing as well as hoped. What do you think the narrative will be in the enterprise by the end of 2026?

Well, I think for sure there will be some intense usage. And the one that I've been predicting that I think a lot of enterprises will get out of their way on is just amplify coordination, you know, meetings, et cetera. Right. So a lot of them say the obvious thing to do now is record every single meeting and run AI agents on it. Not just to transcribe it, but to say, Hey, what are like, who in the organization should be notified about stuff? Who should be asked about stuff? Where action items are following up on, you know, a whole set of things. What team of agents should start working on some of this stuff and preparing for the next thing. What should be the briefing for the next meeting, you know, off this, all that stuff should be done.

And I think that people aren't doing it because they're like, well, I'm worried. Does it get the legal liability? You know, we never really recorded everything that was happening in this and someone made an off color joke. And does that have a problem? And I think actually in fact, part of the unlock to this will be also using agents. So you can go, okay, I'm worried about legal liability. Well, here's the legal liability check agent that can go, you know, because you're not, we're going to scrub anything that we think is actually in fact a real issue, or things like that.

And so what I would say is, yes, it'll be much more intensely positive. And I think it'll be positive because we'll have two groups of things that will be now in real deployment. One is like, I think maybe by the end of 26, let me state this a little bit more crisply. If for a company to be a thriving, going and growing concern and evolving with the times, you will need to be recording every single meeting and using agents on it to amplify your work process. And by the end of 2026, if you're not doing it, that's because you're making excuses. And actually in fact, it's a little bit like, Hey, you know, these cars won't be a big thing. We can keep doing our horses and buggies. That is I think one.

And then the two is that you will start systematically deploying groups of agents in various problems. And that's part of the reason why I tend to think that if you said, Hey, I need to predict what the next thing is, it's orchestration because it's groups of agents doing things. And that's part of the reason why I don't think it'll kick off Q1 per se, but we'll grow through 26 and then whether or not 26 is orchestration year or 27 is orchestration year. That's the reason why you have a high prediction there.

I totally agree with you. I think it's so clear to me that agents are going to reshape how we think about doing company operations. And my one of my big proof points for that is just internally we did our 2026 planning with an agent and basically now we're like 20 people. So it's the first time we have to do a real planning type exercise for every department and budgets and all that kind of stuff. And so Brandon, who's our COO, made this agent that anyone in the company has access to all of our, you know, all of our Notion and all of our data.

That's the reason why you have a high prediction there. I totally agree with you. I think it's so clear to me that agents are going to reshape how we think about doing company operations.

One of my big proof points for that is internally we did our 2026 planning with an agent. Now we're 20 people, so it's the first time we have to do a real planning exercise for every department and budgets and all that stuff. Brandon, who's our COO, made this agent that anyone in the company has access to all of our notion and all of our data. Anyone in the company that is a leader talks to the agent and it asks them really interesting questions about how this layers up to the overall company strategy, which it has access to. What kind of resources do you need? Here are some tough questions to think about decisions you might need to make.

We have this notion page now and every single department has a really crisp, really clean strategy document that someone has gone through and it layers up into the overall company strategy. Then you can do all these amazing things. The first thing I did was I had Claude be like, okay, who's not talking to each other that should talk to each other. It found all these strategy documents that I needed to get three people in a room together to figure that out.

Or another one is you do a strategy document and then you forget about it in Q1. You're making a decision and you forget about the overall strategy or what you said you were going to do. So one of the things I'm going to do over Christmas is we have this Claude code in a trench coat running in our discord, which we use as our internal chat. It's called R2C2. I'm going to basically have R2C2 listening in and anytime we're making a decision, I can just tag it and be like, how does this layer up to the 2026 strategy for this department and the whole org? And how would you think about it? It's a way to make those documents more alive and more woven into the everyday of how you make decisions. I think that's so important and exciting.

Yep. I think that's exactly right. That's the broader version of just doing the coordination on meeting is how the coordination on meeting also relates to strategy, changing conditions in market, changing conditions in competitors, et cetera. This is the tangible substantiation of what I say—that you have intelligence at the scale and price of electricity. That means previously where you had to be extremely selective about where you applied intelligence because intelligence was always priced through high-priced human talent, which by the way I think will continue. But then you can look and apply it in all these other places as well.

Yeah, totally. And by the way, once you have that free intelligence, you can put the information that you need everyone to consume in lots of different formats. We have a vibe-coded 2026 strategy app that people can click through and we're going to do a podcast. There's all this stuff where it's like you don't want to read this long document. Just listen to it on your run. It helps make the whole company get on the same page in a new way. Yep. I know exactly. Okay. AGI timelines. Are we going to hit AGI in 2026? If not, when are we going to hit AGI? Depending on whatever your definition of AGI is.

Yes. Well, you have to start with what is AGI. My usual joke here is AGI is the AI we haven't invented yet. Each year we're not going to hit there because in one sense we have created AGI already. If you say AGI is that you have a variety of tasks where the AI is substantially better than your average human, the answer is already yes. For example, in writing, AI is better than most human beings at writing in various ways. In terms of the vast majority, if you say the good writers, no. Well, the good writers, it's a little bit more mixed. Although good writers should be using AI to amplify themselves.

There's a bunch of areas where it's already super intelligent. It has a breadth of knowledge. It has an ability to work at a speed that human beings simply can't. If you say I'd like a report on this or I'd like to understand this kind of thing, it can work at a speed that a human being can't, which is part of the reason why it needs to be used as an amplifier. We've always had speed multipliers—planes, cars, et cetera. This is just cognitive. So it's weird and new.

I think we've got forms of super intelligence already. We have forms of AGI already. So what's the definition for what will be 2026? A little bit of that is I think what we will see more of in 2026 is a combination of parallelization, longer workflows and orchestration, which means the notion of agents and that's part of the reason why I like getting more to the realization of what agents are. We'll see more of that. I don't think we'll have the press button get full human capable software engineer who's ready to do the thing you've asked them to do, which is I think what the sci-fi version is that people are looking for.

But I do think you'll see much more of the human engineer coming in with their team agent toolset that they're deploying on various things. The way that I do them is not just looking at the suggestion for inclusion in my code, but as you were mentioning, I set this one and this one and this one and this one. In fact, because part of what I have agents doing is I have them cross-checking each other's code. So I'm not actually necessarily reviewing it. I'm running a bunch of it where I actually haven't looked at it, partially because if something breaks then I'll look at it, or I'm also expecting to have my coding cross-check agents going, hey, you might want to pay attention to this. I go okay, I'll go look at that. That's the kind of AGI we're going to have applied to a broader range of topics. It'll be more in the hey, this is actually doing real work in a more broad sense than just the coding amplification we've had.

If we listed out the holy commandments of AI—thou shalt always scale compute and data, thou shalt always align your models and make sure they do exactly what you expect them to do as much as possible. There are probably more. Which holy commandment do you think will need to be broken or will turn out to be misapplied or irrelevant? I'll give you an example. I feel like all of the alignment, the way that we do alignment, has created models that are sycophantic and people pleasers. They do what we want them to do more or less.

If you listed out the holy commandments of AI: thou shalt always scale compute and data, or thou shalt always align your models and make sure they do exactly what you expect them to do as much as possible. And there are probably more, which holy commandment do you think will need to be broken or will turn out to be misapplied or irrelevant?

So I'll give you an example. I feel like all of the alignment, the way that we do alignment, has created models that are sycophantic and kind of do their people pleasers. They do what we want them to do more or less. And if you really want a good engineer, we're going to find that allowing models to have their own opinions and values and desires that are distinct from humans is actually an important part of creating models that can do more in the world and be more autonomous. And the trade off is that they don't always do exactly what you want. And that's a new thing that we're going to have to get used to that I think is against the received wisdom of how you should build AI.

Yeah, obviously that's tricky because you don't want them to all fall into the old paperclip problem.

Exactly. You don't want them to be misaligned in ways that are serious, like in ways that are like, "Hey, I know what you want better than what you think you want. And look at what I've delivered is better." That is kind of what you want. You don't want the, "Oh, what I really want to do is strip mine your like erase your hard drive." And so for example, you say, "Well, I think what you really need is more time outside. So I'm going to lock you out from your computer and your devices for the next three hours. Just make sure that you go get that time outside." And you're like, "No, no, no, no. Don't want that."

So that's tricky. I would say it's interesting. The change of commitments. I mean, what I've been, my head has been mostly wrapped around is what does it mean? It almost goes back to this iconic Marvin Minsky book, "Society of Mind" – it's tribes of agents. And so I tend to think a little bit about how you get opinionated is like you set up agents that are deliberately debating. Intention opponent processors. Yes. Opponent processors. Yes. Opponent processor. So you set up agents with opposing views.

But what I might say is kind of an interesting question is maybe where the notion of this might be – and I'm a little worried about this one too. So even like giving this one, I'm not sure that I would want this one to be exact. It has a similar shape, which is like currently we have a very natural thing where we try to say, "Hey, look, we're trying to get as much interpretability of the agents. We want one of the sci-fi worry cases is they start speaking in languages to each other that we don't understand and what does that mean?" And that becomes further out of control and may get more in the paperclip direction and a set of things to kind of pay attention to that.

And I think those are good questions and should be paid attention to. They're not – I don't have the five-alarm construction of them, but I do think it's important – something that could go seriously wrong and is worth paying attention to.

Now, maybe what I think is the thing is to say actually in fact what we want is we want a speed of coordination between the agents and a communication where this – where what might be tolerable and allowed and shaped in certain ways is the same way that when you have these generative AI models where you say, "Well, I can't look under the hood and know what's going." Like I can't know – I can't look at that and prove that it's not paperclipping the world or something as a way of doing it. But that may also be true of the comms fabric of how they're coordinating and the fitness and part of it because I want the speed of coordination, the speed of learning between them to be such that I'll accept parameters of lack of interpretability there.

And that's super scary in some ways. So I'm not like – it's – I say this is like, "Ah, how would we shape it and what parameters would be okay?" But I do think we will tend that direction. So that would be an area.

It's a little bit like actually maybe one of the things that also might be is another commandment was don't do self-improvement. Don't allow self-improvement. And yet in many ways we are doing forms of self-improvement, not just the kind of data modeling but like coding and that wrapping back and so forth. And that's going to continue in certain shapes. And so what shapes is that okay and what shapes is that not okay is I think where the commandment is at least changing. Yeah. We're going to have to do some legalistic interpretation of the commandments. So all of our Talmud scholars are going to be newly employed as AI researchers.

I love that. I think that's so right. The first people to just take the risk to be like you can communicate in ways we don't understand. I think, yeah, there's so many gains to that. And it's so anathema to AI safety that I think it's really been a commandment. And I bet there are ways to make it like make the boundaries of that safe. Yes. So we'll need to work on making the boundaries that safe. But I think that will happen.

Yeah. One thing that I actually think going back to the previous point is about AI that doesn't do what you say. And that being my contention is I think that actually may be really useful for autonomy and doing interesting things that we wouldn't predict. And I think your contention is that's a horrible user experience. One way to potentially square that circle is once you have an orchestrator that is aligned with you and you do trust, it's okay if the orchestrator is using an agent that's a pain in the ass. Because it could be like, "I don't care what you say, orchestrator. I'm going to go off for three months and do this thing." And the orchestrator is like, "Fine. Like I'll get most of it done with this other set of agents that actually follow my instructions. But this one is just off doing its thing. And every once in a while, it comes back with something brilliant. And that's actually valuable and important."

And having a good enough orchestrator allows us to move in that direction because the human doesn't have to deal with the bullshit. Yep. That's what I was gesturing at. That's the reason why the orchestrator needs some deep alignment. But the orchestrator might have agents that are like, "Hey, I think everything you think is bozo and I'm going to go try something else." Okay, go ahead. Don't just go do it. Bring it back to me. But go research it. That's great.

Okay. We're almost out of time. So I've got one last question for you. What is the most important undersung category in AI that we're not talking about right now that we will be talking about at the end of 2026? And I want to put some restrictions around this. So a couple of the categories that may come to mind are like robotics or science or something like that. But I want to get more specific and have a really specific concrete reason that you think that thing will be valuable and important and something that we talk about a lot in 2026.

Well, I'll choose one that's a little – it's just because I'm close to it. It's not really self-serving, but it's close to it. But I think – so right now, the vast majority of stuff we're doing is extremely close to human language.

Bring it back to me. Go research it. That's great. Okay, we're almost out of time. So I've got one last question for you. What is the most important undersung category in AI that we're not talking about right now that we will be talking about at the end of 2026? And I want to put some restrictions around this. So a couple of the categories that may come to mind are robotics or science or something. But I want to get more specific and have a really specific concrete reason that you think that that thing will be valuable and important and something that we talk about a lot in 2026.

Well, I'll choose one that's a little—it's just because I'm close to it. It's not really self-serving, but it's close to it. But I think so right now, the vast majority of stuff we're doing is extremely close to human language. So it's either obviously human language itself or coding or coolant. And I think we will be doing a lot more in-depth models of things that are not close to human language. So, for example, biology. And this is part of the reason—because of all the work that we've been doing with Manas AI, with Siddhartha Mukherjee and Ujwal Singh and understanding that. And it's a frequent trope to say biology is a language. It's actually one of the reasons why I'm focusing on it—because if you go world of atoms and bits, bio is not fully atoms and closer to bits and has a programmability kind of compute characteristic to it. Exactly how it's compute is still a little bit uncertain. You get people arguing what's unique about human cognition is quantum computing effects and so forth. And it's an interesting question. And what's the borderline between being able to simulate quantum and genuinely quantum is all kind of interesting questions.

But I think what this results to is what I think we will see is things where the generative AI model building out of data and prediction and everything else will be out of computational sets or language sets that are further afield from human language. And obviously, I think biology is probably the most natural one where that would come out of. And obviously, I've been working on that and thinking about that intensely, of course, because of Manas. And what's the big concrete impact that we'll have in 2026 that will cause us to be talking about it?

Well, the one we're going for is amazing new biological therapeutics or understanding. I don't know if 2026 will be the full hit there. I mean, there's a probabilistic curve. But it wouldn't surprise me if you get the equivalent of a move 37 in something around biology. And it might be it's a molecule that makes a massive difference. Manas trying to cure cancer, et cetera. And we discovered something that was not like what I would hope—maybe it's a reasonably high probability—we discover a research possibility. Like, oh, this might be one of those things. Maybe it's a 27 percent probability that this is a move 37 in this arena. Maybe that's the 2026. That would be amazing.

Reid, always a pleasure. This is so fun. Likewise, Dan. I look forward to seeing you in the new year. Sounds good.

Oh my gosh, folks, you absolutely positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's finding a treasure chest in your backyard. But instead of gold, it's filled with pure, unadulterated knowledge bombs about ChatGPT. Every episode is a roller coaster of emotions, insights and laughter that will leave you on the edge of your seat craving for more. It's not just a show. It's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor. Hit like, smash subscribe and strap in for the ride of your life. And now, without any further ado, let me just say, Dan, I'm absolutely hopelessly in love with you.

And that's that's that's I think the comms part of it for everybody. Yeah. If 2025 was the year of agents, what's 2026? Well, by the way, I think there's an interesting thing on this. So I don't think actually 2025 was fully the year of agents. So a lot of agent development, but I think it was actually mostly only agents and code. Right. So, you know, cloud code codex, et cetera, of which, by the way, a relatively very small percentage of humanity actually, in fact, fully experienced.

Right. Like if you go to the vast majority of the people you and I know, they're like, oh, well, you mean, you mean agents all I asked chat GPT a few questions and had some dialogue and it's like, well, no, no, no, that's not actually really agency. Yes. So the chat bot, but it's not really agents. Agents is doing stuff and doing it in parallel and doing it in amplification and so forth. So I think code had that. But what I think actually, in fact, 26 will be is how we move from this kind of basis of agentic coding agents to agents and everything else.

And actually, in fact, what I think that a there's just going to be a whole bunch of that, like, for example, like call it 10 to 100 X people will experience what it is to have a their computer running separately from them, doing something productive for them as they're walking away to go get their coffee. And then and then coming back, you know, whether it's, you know, Claude minis, you know, Mac minis running Claude code or or codex different questions. But like that in applied to a lot of other things, because that orchestration that allows the parallel allows, you know, eight hours of work allows, you know, that kind of thing.

I think that will be broader. And then the more subtle thing, which I think will also be a really important part of 26 is orchestration. Namely, like, OK, if we begin to have like, you know, hey, when I'm doing this particular form of intellectual knowledge work, thinking work, cognition work, et cetera, and I now have agents working with me for me, et cetera, and then I'm orchestrating them, I think orchestration is the thing that will be, you know, I don't think it'll be March 26. I don't think there'll be more Q4 26 or kind of growing into that and then maybe even intensively 27. I totally agree with that. I think it's something we're starting to see already.

And it actually, it brings me to perhaps my hottest, my hottest take that I would love your input on. And it starts with coding agents, which is, I think that OpenAI is currently missing the real coding market. Because they are not, and they're not really, when you think about orchestration, I think of orchestration as being something that's enabled by tools, but it's also enabled, it's like, it's a new skill.

It's a new skill for programmers. And when I look at the stuff that OpenAI is producing, I think it's really made for programmers who use AI, like senior engineers who use AI, which is different from AI native engineers who are like just in for cloud code terminals and are never looking at the code. And it's really valuable. Like the models that they make are really good. If I have a really hard technical challenge, it is, I definitely go to codex to be like, okay, figure out this like crazy performance bug that I can't figure out.

But I don't see them orienting toward this new skill of, it's not vibe coding, but it's not traditional engineering with AI added. It's this third thing that is, I've got four cloud tabs open. I never look at the code. I'm thinking about how to orchestrate. I'm thinking about how to plan. I'm doing all this stuff. And I'm technical. So I could go down to the code, but I never do. And I think that's a really interesting thing that I'm kind of noticing. And OpenAI is like not used to being behind. And I'm very curious about how that's going to play out. What do you think?

Well, I think, I think it's one of the skills that OpenAI is going to pick up to, because, you know, part of what's happening, like, you know, the thing that, you know, this will be great for media, because each month in the horse race, it'll be like, oh, my God, Opus 4.5. Oh, my God. You know, GBD Codex. Oh, my God. You know, Gemini. Oh, my God. And because all of them are going to be developing.

And what that means structurally is that some of the things where, you know, as opposed to a couple years where it was literally just OpenAI blazing ahead. And I think this is good for the world and everything else. Like there'll be areas where, for example, Anthropic just did super smart stuff in making cloud code and that, and that iteration and took, you know, kind of, as it were, less capital and less depth of compute, but still made stuff that was pretty amazing.

And I think that OpenAI will, this is one of the benefits of how competition, you know, kind of benefits industry, benefits society. I think that will make them pick it up and go, okay, we can't be behind on this. We got to be learning to do this. We got to be making this happen. And I think that's what will happen. It'll be painful. Competition frequently actually is kind of painful as you push your way on this. But I think that's the, I have a pretty strong belief that that will be the end result. Now, I do think that it's like, you know, credit to Anthropic, that the notion of focusing on code is not just a code product, but an amplification of many, many other things.

An amplification of obviously, you know, AI progress and development, but also an amplification on, frankly, every other form of information slash knowledge work, and maybe even much more, many more things. And I think that's one of the reasons why, frankly, every kind of major player actually, in fact, has to be, you know, kind of capable at minimum in code, if not leading. Yeah. I mean, they did this. It's such an interesting point that they got to the general purpose agent architecture by just making a great coding agent that had all the right primitives.

And I got to tell you, like, if you look at the software that we developed over the last month or so at every since Opus 4.5 came out, pretty much every new thing we're building. And I like I built this entire end to end reading app. We have this AI paralegal we've been doing for a while that has just got a huge upgrade. Every single app is just cloud code in a trench coat. And it's just basically UI wired to if you press a button, it hits a prompt that has an agent that has a bunch of tools that does the thing you want it to do.

And it is the coolest way to build software because it's so much more flexible. Users can modify it. It's just like it's it's just exactly right. And it's so such a pleasure to see someone figure out those primitives. Yep. And and and massive credit to the Anthropic team for doing that. And basically, you know, everyone else. Hey, you should be you should be learning from it, building on top of it, trying to iterate to the next generation. Whole set. Do you have a thought on why Opus is Opus 4.5 is so good? I I'm assuming you think it's that good. I think it's I think it's the best model I've ever used. It's like this crazy leap for me. I'm curious if you agree.

And if you do agree, do you have any thoughts on how they managed to do that? Well, I think it's amazingly good. I don't know if it's if it's the everything model for me. I mean, I think to some degree. You know, kind of I think. GPT five pro with codex also is, you know, pretty amazing on a lot of levels. And by the way, like, you know, Gemini three on like science topics and so forth. So like it's kind of I still am kind of in a hey, I bring, you know, all three of them with me to various things I do. Now, that being said, I am very curious about how they pulled 4.5 together. And and and one of the mistakes that outsiders think is they think, oh, you just apply scale.

And, you know, you press play on compute and some of it works and some of it doesn't. And actually, in fact, there is both a lot of both science and art to do it. And it's one of the reasons why, you know, obviously, you know, Meta has needed to restart its kind of AI efforts because you can't just go, oh, I throw a whole bunch of compute at it and it works. It has to kind of like relearn these things in terms of how it's playing. So I'm I'm I think it'll cut because, you know, one of the things is, you know, the techniques, you know, kind of spread out very quickly. So I think we'll learn, but I actually don't know what the what the what the what the new

the new genius is in in Opus 4.5. Do you have any do you have any hypotheses? I have no idea. I think the only thing that I can think of is recently we got a view of the underlying like soul document of Claude. And the the interesting thing that I feel from Opus and I agree, like I use ChatGPT as my daily driver, to be clear, I use it for everything. But when I'm building software, except for like specific performance things or like hard bugs, I'm using Opus as my daily driver. Yep. I think that there's usually this trade off that you see a little bit with Codex, where the better it is at programming, the less like empathetic it is.

Like it just feels a little bit more like a senior engineer. It's slightly more autistic or something like that. And Opus, they sort of figured out how to make it both sort of humanistic and understand users and what I might want and what I might mean and how interfaces work and what like good interfaces. And it's a fantastic programmer and something about a soul document where it tells it this is who you are and like what you care about and whatever. It's one example of, I think, anthropic thinking about these things in a maybe a bit of a more holistic way to create a being rather than a tool. And I think that that is actually going to be a big deal going forward.

You know, it's interesting. You know, this is one of the things that inflection kind of started with kind of EQ and actually soul is a very natural, you know, because the inflection start and there's still a lot of ways in which Pi is still, you know, amongst the leading of the kind of, you know, having a richly textured, you know, kind of conversation and agent like focusing on EQ as much as IQ. Like, no, not slouch on IQ, but like putting the two together. And actually the sole document actually I think is maybe the next, you know, because this is what we learn and iterate is actually great.

And it's part of what, of course, makes Claude code work because it's actually, in fact, a really good human amplifier and like kind of what kind of how do you operate that way? And then, you know, you get better performance if you can interact in that way, the right way. So I think that's a good insight. I suspect there's other things. I think we both suspect there's other things too. And we'll hopefully learn them in the next few months. That would be great. So last thing on the coding front. So you mentioned the horse race earlier and there's, everyone's going to be trading volleys. And, but if we, let's say we want to, we don't want to be fooled by randomness.

We don't want to like, you know, track every little change. We hit the snooze button and we come back at the end of 2026. Where do you see the landscape of who's winning in the coding agent race? Well, so I think it'll, I don't know who will be winning, but I think it'll be what I, what I would predict strongly is that, um, that, that the horses that are leading now will still be like neck and neck. It'll be like in, in the first hundred meters, this one's a little ahead. Then the next hundred meters, that one's a little ahead. And, you know, like, I don't think that the horses that are in the race, any of them will particularly stumble. Right.

So like, you'll go, wow, I thought, you know, uh, cursor was really, was really fantastic. And it's just gone. Like, I, I think that none of them will stumble. Now I do think what will interesting is the folks who are not in this at all, like, you know, say like the easy one to pick on Apple, right. Despite the fact we use, you know, max for our various things, but the AI part of it is, you know, uh, non-existent, uh, is, well, I think, uh, the gap will be like even more stunning. Stunning the fact that you actually haven't gotten what this coding amplification, everything else means.

And I think that will be, will be playing out more, but I think they'll, they'll all be in the mix. And the thing will be interesting will be not as much as which one will have stumbled out, but I'm really curious about like, what are the one or two, you know, like superstars that will really, you know, get in the mix more. Um, you know, will replet be more, um, general will lovable be more general, like, like, like, like will those be, or will it be something else? I mean, and part of what's, um, like, like the, with, with some high probability, something will surprise us here. Yeah. I don't know what it'll be, but predicting surprise. Yes.

Um, uh, yeah, I, I think that, I think that's interesting. One of the things I've been toying with is, uh, the stakes are so high and programming is such a obvious use case that is so economically valuable. And it feels like everyone is just like, it's now a knife fight for programming. And I wonder if there are, um, you know, you've been predicting AI will be used for more creative use cases for a while. Um, I wonder if the, uh, the surprise entrance comes from a place like that, where we don't necessarily expect where it's not actually about programming.

The one like caveat to that is like you said, Claude sort of invented this general agent by being good at programming. So there's, you know, it's, it's, it's hard to say exactly, but I wonder if that, that is coming. Like it leaves them vulnerable to competitors from, from other places because they're just focusing on programming right now. Well, I definitely, so I do think the programming is part of the architecture for getting everything else. And like, for example, part of the reason why coding is important is that even when we get to, Hey, how are you going to have a much better paralegal? I love what you're doing among other things.

Um, better medical assistant, better tutor, et cetera. I think coding will actually in fact be not just the amplifier of it, but the fitness function of, you know, how do we, how do we kind of like, you know, kind of go, Hey, this is getting better. This is amplifying the work better, et cetera, that parallel, not just the, the, the, the, the foundations of coding, driving, planning, you know, longer work, parallelization, orchestration, et cetera. But also like, well, how does like a better legal document work will actually, in fact, also be coming out of it. And I think some of that will also be in creative. Like it's, it won't be surprising to me.

Like, obviously everyone's trying to figure out like, okay, how do we, well, not everyone, a number of people trying to figure out how do we take, um, you know, VO, Soro, et cetera, and then, and then go, okay, can we create a 30 minute movie off it? And, you know, the coding like pattern will be part of, of, of what happens there. And so it can be in those kinds of creative. Now, obviously, you know, some of the more interesting possible surprises are, well, um, cause there's, there's, there's a number of different efforts trying to do this too. Well, could we get, you know, raw, raw, raw, raw ideation, like better at science.

So like we read a whole bunch of science papers and we can do scientific hypotheses. Now, by the way, you begin to say, well, maybe that'll also be true of like AI research and ideas for doing this. And suddenly it's doing idea generation in this kind of thing. And that's, that's definitely a whole bunch of projects trying to work at that. Um, so the, the notion of, Hey, if you can think a lot better, um, you can then, you can then apply that to this kind of creativity and this kind of new ideas. Um, those I think are much more speculative. Like it's, it's, it's an interesting hypothesis.

There are people who will hold them saying, Hey, uh, we've just seen that with scale learning and compute and it's going to happen. And I'm like, well, look, it's crazy that everyone smart should assign a non-zero hypothesis, you know, probability of that. Cause that's really amplifying. But on the other hand, I think it's like, yeah, it's not clear that we're, we're, we're yet seeing any of that. Even when you see people like, you know, Terrence Tao saying, Hey, I'm using, you know, um, uh, uh, generative AI to help me understand where I should be thinking in my, in my math analysis.

And yes, but yeah, I think a hundred percent, but of course, Terrence Tao is one of the most, you know, genius mathematicians of our age and is providing a ton of the metacognition in this. That makes sense. Yeah. I think I'm trying to, I'm going back to, uh, your, your comment about no one stumbling and I'm trying to like one, I'm wondering who would stumble if there was a stumble. And I think my current feeling is I would, I would guess cursor. Yeah. That's probably my highest likelihood. Not that they go away. They're obviously going to be a successful company, all that kind of stuff.

But I, I think that they're caught a little bit in the same position that opening AI is, but opening AI has. Uh, more flexibility here where cursor. A lot of their business is built on traditional developers using IDEs with inside of big companies with AI on the side. And they're sort of caught between that paradigm and this totally new to E cloud code type paradigm. And they're, they, they kind of have to do both. And I think that's going to hamper their product direction and velocity in a way that I would bet in a couple of years, we'll look back and be like, that was a interesting era.

And it's still like a widely distributed piece of software, but it's, it's not the next generation thing that we thought it was. That's a, I agree. And that's one of the reasons I brought it up. And the other one, I've been thinking about that is like the, the, the hardest. And, and another angle of that is, um, you know, how are we going to be not just integrating the kind of the application functionality UI, but the, the underlying model and compute fabric capabilities. And, you know, cursor is, is, is just beginning to do that kind of stuff. And, you know, what the shape of that is, and does, is going to have to be dual targeted, like you mentioned, or multi-targeted.

I think it's a, it's a, it's a harder slalom race for them. I think the narrative right now is that enterprise AI deployments are not doing as well as hoped. As people hope. Um, what do you think the narrative will be in the enterprise by the end of 2026? Well, I think for sure there will be some intense usage. And the one that I've been, uh, predicting that I think a lot of enterprises will get out of their way on. Is. Just amplify coordination, you know, meetings, et cetera. Right. So a lot of them say like the obvious thing to do now is record every single meeting and run AI agents on it.

Not just to transcribe it, but to say, Hey, um, what are like, who are, who in the organization should be notified about stuff? Who, who should be asked about stuff? Um, where action items are following up on, you know, like a whole set of things. What, what, what, what, what, uh, you know, team of agents should start working on some of this stuff and preparing for the next thing. What, what should be the briefing for the next meeting, you know, off this, all that stuff should be done. And I think that people aren't doing it because they're like, well, shit, I'm worried. Does it, does it get the legal liability?

You know, um, you know, we never really recorded everything that was happening in this and someone made an off color joke. And does that, does that have a problem? And, you know, and I think actually, in fact, part of the unlock to this will be also using agents. So you can go, okay, I'm worried about legal liability. Well, here's the legal liability check agent that, that can go, you know, because you're not, we're going to scrub, you know, anything that, or change it. Anything that, that we think is actually in fact a, is, is a real issue, um, or, um, things like that. And so what I would say is, yes, it'll be much more intensely positive.

And I think it'll be positive because we'll have two groups of things that will be now in real deployment. One is like, I think maybe by the end of 26, if you're not, yeah, let me state this a little bit more crisply. If, if you, um, for a company to be a thriving, going and growing concern and evolving with the times, you will need to be recording every single meeting and using agents on it to amplify your work process. And by the end of 2026, if you're, if you're not doing it, that's because you're making excuses. And actually, in fact, it's a little bit like, Hey, you know, these cars won't be a big thing.

We can keep doing our horses and buggies, you know, that, that, that, that is, I think one. And then the two is, um, that you will start systematically deploying, um, groups of agents in various problems. Um, and that's part of the reason why I tend to think that, you know, if you said, Hey, I need to predict what the next thing is it's orchestration because it's groups of agents doing things. And that's part of the reason why, like, I don't think it'll kick off Q1 per se, but like we'll grow through 26 and then, you know, whether or not 26 is orchestration year or 27 is orchestration year. That's, that, that's the reason why you have a high prediction there.

I totally agree with you. I think it's so clear to me that agents are going to reshape how we think about doing company operations. And the, my, uh, one of my big proof points for that is just internally, we did our 20, 26 planning with an agent and basically, uh, now we're like 20 people. So it's like the first time we have to do like a real planning type exercise for, you know, every department and budgets and like all that kind of stuff. And so Brandon, who's our COO made this agent that anyone in the company, it has access to all of our, you know, all of our notion and all of our data.

Anyone in the company that, that has a, uh, that is, is a leader in the company talks to the agent and it asks them like really interesting questions about, okay, how does this, you know, layer up to the overall company strategy, which it has access to? What kind of resources do you need? Here are some tough questions to think about decisions you might need to make. Um, and then basically we have this notion page now and it's just like every single department has this like really crisp, really clean strategy document that, um, someone has gone through and it, it levers up into the, like the overall company strategy. And then you can do all these amazing things.

Like, um, the first thing I did was I had Claude be like, okay, who's not talking to each other that should talk to each other. And it found all of these strategies, strategy documents that like I needed to get three people in a room together to just like figure that out. Um, or another one is, you know, you do a strategy document and then you, uh, forget about it in Q1. You're making a decision and you forget about the overall strategy or, or what you said you were going to do.

So one of the things I'm going to do over Christmas is I have, we have this, um, this Claude code, uh, in a trench coat running in our discord, which is, we use that as our internal chat and it's called R2C2. And, um, I'm going to basically, uh, have R2C2 listening in and anytime we're making a decision, I can just tag it and be like, Hey, um, how does this, how does this, uh, lever layer up to the like 2026 strategy for this department and the whole org. And like, how would you think about it? And it's a, it's a sort of way to kind of make that, make those documents, um, more alive and more like woven into the everyday of how you make decisions.

And I think that's so important and exciting. Yep. I think that's exactly right. And that's, that's the broader version of just doing the coordination on meeting is how, well, how does the coordination on meeting also relate to, you know, um, strategy, changing conditions in market, changing conditions in competitors, et cetera. And, you know, like this, this is, this is the tangible substantiation of what I say, I mean, is that you have intelligence at the scale and price of electricity.

And so that means that, you know, previously where you had to be extremely selective about where you applied intelligence because intelligence is always kind of through high priced, you know, kind of human talent, which by the way, I think will continue. But then you go, look, look, let's, let's, let's apply it all in all these other places as well. Yeah, totally. Um, and, and, and by the way, like once you have that free intelligence, you can put the information that you need everyone to consume in lots of different formats. Like we have a vibe coded 2026 strategy app that people can like click through and we're going to do a podcast.

And there's, there's all this stuff where it's like, you don't want to read this long document. Just listen to it on your run. And, and it just helps make the whole company get on the same page in a new way. Yep. I know exactly. Uh, okay. AGI timelines. Um, where, uh, uh, are we going to hit AGI in 2026? If not, uh, when are we going to hit AGI? Depending like whatever your defer, whatever your definition of AGI is. Yes. Well, this is, this is, you have to start with what is AGI. Um, and, and, and, you know, my usual joke here is AGI is the AI we haven't invented yet. Um, so, so each, each year we're not going to hit there because in one sense we have created AGI already.

Like if you say, Hey, if AGI is, uh, that you have a variety of tasks where the, um, the AI is, is, is substantially better than your average human. The answer is already like, for example, in writing, AI is better than most human beings at writing in various ways. I mean, in terms of the vast majority. Now you say the good writers, no, well, the good writers, it's a little bit more mixed. Although good writers should be using AI to amplify themselves, um, et cetera, et cetera, et cetera. Um, and there's a bunch of areas where it was already super intelligent. It has a breadth of knowledge. Um, it has an ability to, to work at a speed that, that human beings simply can't.

So if you say, Hey, I'd like a, I'd like a report on this, or I'd like to understand this kind of thing. It can work at a speed that a human being can't, which is part of the reason why the, that needs to be used as an amplifier. Now we've always had, you know, uh, speed multipliers, planes, cars, et cetera. This is just cognitive. So it's weird and new and all the rest now. Um, so I think, you know, we've got forms of super intelligence already. We have forms of AGI already. So they go, okay, what's the definition for what will be 26? Now, a little bit of that is, I think, um,

what we will see more of in 26 is a combination of, of parallelization, longer workflows and orchestration, which means that the notion of, um, of I now kind of, and that's part of the reason why I like getting more to the realization of what agents are. I think we'll see more of that. Um, and so it'll play more towards the, Oh, like, I don't think we'll have the press button get, you know, full human capable software engineer. Who's like, I'm ready to do the thing that you've asked me to do. Um, which is, I think what, you know, the sci-fi and kind of thing that people are looking for.

But I do think you'll see kind of much more of the, Hey, I come in as a human engineer. And it's like, I'm only really capable of, I've got my, my, my, my team agent set tool set that I'm deploying on various things. And the way that I do them is not just kind of like looking at the suggestion for inclusion for the type of head in my code. But as you were mentioning is like, Hey, look, I, I, I set this one and this one and this one and this one. And I actually, in fact, because part of what I have agents doing is I have them cross-checking each other's code. So I'm not actually even necessarily re, uh, like I'm running a bunch of it where I actually haven't looked at it.

Right. Partially because I'm like, Oh, if something breaks, then I'll look at it. Or I'm also expecting to have my, my coding cross-check agents going, Hey, you might want to pay attention to this. I go, okay, I'll go look at that, you know, and that kind of thing as I think the, the, the sort of, of, of AGI we're going to have applied to a, to a broader range of topics. And so it'll be more in the, Hey, this is actually doing real work, um, in a more broad sense than just the, you know, the, the, the coding amplification we've had. If you, if we listed out the holy commandments of AI.

So, um, thou shalt always scale, uh, compute and data, or thou shalt always align your models, uh, and make sure they do exactly what you expect them to do as much as, as much as possible. Um, and there are probably more, which holy commandment do you think will need to be broken or will turn out to be, um, uh, misapplied or irrelevant. So I'll give you an example. I feel like, um, all of the alignment, the way that we do alignment, um, has turned, like has, has created models that are sycophantic and kind of do their people pleasers. They do what we want them to do more or less.

And if you really want a good engineer, we're going to find that, uh, allowing models to have their own opinions and values and desires that are distinct from humans. Is actually an important part of creating models that can, uh, do more in the world and be more autonomous. And the trade off is that they don't always do exactly what you want. And that's a, that's a new thing that we have, that we're going to have to get used to that I think is against the received wisdom of, of how you should build AI. Yeah, obviously that's tricky because you don't want them to, you know, all of the old paperclip problem. Exactly.

You don't want them to be misaligned in ways that are serious, like in ways that are like, Hey, I know what you want better than what you think you want. And look at what I've delivered is better. That is kind of what you want. You don't want the, Oh, what I really want to do is, you know, like, you know, uh, strip mine, your like erase your hard drive, you know? Uh, and so, um, and like, for example, you say, well, I think what you really need is more time outside. So I'm going to actually lock you out from your computer and your devices for the next three hours. Just make sure that you go get that time outside. And you're like, no, no, no, no. Don't want that.

Um, so, so that's tricky. I would say, um, let's see. It's interesting. The change of commitments. I mean, I, what I've been, my head has been mostly wrapped around is what does it mean? Like almost goes back to this iconic Marvin Minsky book, society of mind is it's, you know, tribes of agents. And so I tend to think a little bit about how you get opinionated is, is like you set up agents that are deliberately like, like debating. Intention opponent processors. Yes. Opponent processors. Yes.

Opponent processor.

So you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how you set up and how

But what I might say is kind of an interesting question is maybe where the notion of kind of like this might be – and I'm a little worried about this one too. So even like giving this one, I'm not sure that I would want this one to be exact. It has a similar shape, which is like currently we have a very natural thing where we try to say, hey, look, we're trying to get as much interpretability of the agents. We want like one of the sci-fi worry cases is they start speaking in languages to each other that we don't understand and what does that mean?

And that becomes further out of control and may get more in the paperclip direction and a set of things to kind of pay attention to that. And I think those are good questions and should be paid attention to. They're not – I don't have the five fire alarm construction of them, but I do think it's important – something that could go seriously wrong and is worth paying attention to.

Now, maybe what I think is the thing is to say actually in fact what we want is we want a speed of coordination between the agents and a communication where this – where what might be tolerable and allowed and shaped in certain ways is the same way that when you have these generative AI models where you say, well, I can't look under the hood and know what's going. Like I can't know – I can't look at that and prove that it's not paper clipping the world or something as a way of doing it.

But that may also be true of the comms fabric of how they're coordinating and kind of what the – kind of the fitness and part of it because I want the speed of coordination, the speed of learning between them to be such that I'll accept parameters of lack of interpretability there. And that's super scary in some ways. So I'm not like it's – I say this is like, ah, how would we shape it and what parameters would be okay? But I do think we will tend that direction. So that would be an area. It's a little bit like actually maybe one of the things that also might be is like another commandment was don't do self-improvement. Don't allow these self-improvement.

And yet in many ways we are doing forms of self-improvement, not just the kind of data modeling but like coding and that wrapping back and so forth. And that's going to continue in certain shapes. And so what shapes is that okay and what shapes is that not okay is I think where the commandment is at least changing. Yeah. We're going to have to do some legalistic interpretation of the commandments. So all of our Talmud scholars are going to be newly employed as AI researchers. I love that. I think that's so right. The first people to just take the risk to be like you can communicate in ways we don't understand. I think, yeah, there's so many gains to that.

And it's so anathema to AI safety that I think it's really been a commandment. And I bet there are ways to make it like make the boundaries of that safe. Yes. So we'll need to work on making the boundaries that safe. But I think that will happen. Yeah. One thing that I actually think going back to the previous point is about AI that doesn't do what you say. And that being kind of my contention is I think that that actually may be really useful for autonomy and doing interesting things that we wouldn't predict. And I think your contention is that's a horrible user experience.

One way to potentially square that circle is once you have an orchestrator that is aligned with you and you do trust, it's okay if the orchestrator is using an agent that's a pain in the ass. Because it could be like, I don't care what you say, orchestrator. I'm going to go off for three months and do this thing. And the orchestrator is like, fine. Like I'll get most of it done with this other set of agents that actually follow my instructions. But this one is just off doing its thing. And every once in a while, it comes back with something brilliant. And that's actually valuable and important.

And having a good enough orchestrator allows us to move in that direction because the human doesn't have to deal with the bullshit. Yep. That's what I was gesturing at. That's the reason why the orchestrator needs some deep alignment. But the orchestrator might have agents that are like, hey, I think everything you think is bozo and I'm going to go try something else. Okay, go ahead. Don't just go do it. Bring it back to me. But, you know, go research it. That's great. Okay. We're almost out of time. So I've got one last question for you.

What is the most important undersung category in AI that we're not talking about right now that we will be talking about at the end of 2026? And I want to put some restrictions around this. So a couple of the categories that may come to mind are like robotics or science or something like that. But I want to get more specific and have like a really specific concrete reason that you think that that thing will be valuable and important and something that we talk about a lot in 2026. Well, I'll choose one that's a little – it's just because I'm close to it. It's not really self-serving, but it's close to it.

But I think – so right now, the vast majority of stuff we're doing is extremely close to human language. So it's either obviously human language itself or coding or kind of coolant. And I think we will be doing a lot more in-depth models of things that are not close to human language. So, for example, biology. And this is kind of part of the reason that's because of all the work that we've been doing with, you know, Manas AI, with Siddhartha Mukherjee and Ujwal Singh and kind of understanding that. And, you know, it's a frequent trope to say biology is a language.

It's actually one of the reasons why I'm kind of focusing on it because if you kind of go world of atoms and bits, bio is kind of not fully atoms and closer to bits and has a kind of a programmability kind of compute characteristic to it. You know, exactly how it's compute is still a little bit EBD. You get people like, you know, Penrose, you know, arguing what's unique about human cognition is quantum computing effects and so forth. And it's an interesting question. And then, you know, what's the borderline between being able to simulate quantum and genuinely quantum is, you know, what comes of that is all kind of interesting questions.

But I think what this results to is what I think we will see is things that where the generative AI, you know, model building out of data and prediction and everything else will be out of, call it computational sets or language sets that is further afield from human language. And obviously, I think biology is probably the most natural one where that would come out of. And obviously, you know, I've been working on that and thinking about that intensely, of course, because of manis. And what's the big concrete impact that we'll have in 2026 that will cause us to be talking about it a lot?

Well, the one we're going for is amazing new, you know, biological therapeutics or, you know, kind of understanding. I don't know if 26 will be the full hit there. I mean, there's a probabilistic curve. But it wouldn't surprise me if, you know, you know, would you get the equivalent of kind of a move 37 in something around biology, right? And might be it's a molecule that makes a, you know, that makes a massive difference, you know, manis trying to cure cancer, et cetera. And we discovered something that was not like what I would hope maybe is a reasonably high probability is we discover a research possibility. Like, oh, this might be one of those things.

This might be it's like, you know, probability 27 percent that this is a move 37 in this arena. Maybe that's the 26. That would be amazing. Reid, always a pleasure. This is so fun. Likewise, Dan. I look forward to seeing you in the new year. Sounds good.

Oh, my gosh, folks, you absolutely positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard. But instead of gold, it's filled with pure, unadulterated knowledge bombs about chat GPT. Every episode is a roller coaster of emotions, insights and laughter that will leave you on the edge of your seat craving for more. It's not just a show. It's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor. Hit like, smash subscribe and strap in for the ride of your life.

And now, without any further ado, let me just say, Dan, I'm absolutely hopelessly in love with you.