Transcription
How will the world be different in 2026 with artificial intelligence? I'm going to give you 10 specific predictions for AI breakthroughs in the new year, and I'm going to trace them back to what we already know and why I think that they are 70% or more likely to come true.
Number one, we're going to see a real memory breakthrough. Memory has been an absolute wall in 2024 and 2025, as in, it's not scaling nearly as fast as intelligence. I believe that we have all the pieces in place in late 2025 for a new product surface around memory that feels like a memory breakthrough, even if we don't yet have the perfect memory that people who want to build artificial super intelligence dream of. So, I'm not saying it's going to remember every single thing we say perfectly, but neither do our human brains. What I'm saying is we have tools around compression. We have a lot of experience now designing and building agentic systems that use tools like markdown files and other things to write down memory as we go. And we have long-running agents. All of those put together and our experience in systems design as a community of practice suggests to me that we are going to get to an artificial intelligence memory application layer that we are going to be able to put into our systems very reliably that will both at work and personally dramatically improve memory fidelity, memory completeness, how long and how good that memory is going to be for us. My guess is we will have some of that by the summer of 2026.
Prediction number two, we are going to get an agent software UI breakthrough. I think we already have hints of this. There are rumors that Anthropic is working on an inbox where you can just send an email to your Anthropic agent and it will just do stuff for you. That feels like an early and primitive version of a UI breakthrough. It is time for the general public to have a little guy in the computer that helps you. We again have all the pieces of that. We have the long-running agents. We understand how agents use tools. We have intelligent agents that can make smart decisions. We understand how agents use file systems. We have MCP. We have skills. All we need to do is to put it together and get it into someone's computer in a way that's useful. And we're going to have a big hardware upgrade cycle this year because this is the first year when consumer laptops are really hitting the shelves with graphical processing units that can tokenize information locally. And so you will get real performance bumps and speedups, and it will become more viable on consumer hardware to get the little guy in the computer that helps you. So I think we're going to have two or three or four of those startups hit the market. And I think that if one starts to click, you will see an explosion in usage. Sort of like ChatGPT where people are just like, "Oh my gosh, what would I do without the little guy in my computer that does all this work for me?" We'll see. But I'm, I'm optimistic about that one.
Number three, continual learning will feel less like a dream and more like an engineering rollout. I do think that the model makers are very close to solving for ongoing learning from their models. One, they're aware of it. Two, they're developing techniques to address it. And three, the pace of gains is accelerating as models themselves help to train models. And so I expect by, I'm going to guess Q2 2026, that we are going to start to see the first systems come out that do have some kind of continual learning. I think this one will be a little bit janky in 2026, but continual learning is such a massive unlock for models that even a janky rollout is going to be a really, really big deal. So continual rollout learning is just, you have the model and it gets smarter after you roll it out. The model no longer wonders what Gemini 3 is or what ChatGPT 5.2 is, even if it is ChatGPT 5.2, because it can learn as it goes. And that's going to be a huge deal, and it's going to make these models even stickier and even more valuable.
Number four, in line with that, recursive self-improvement of models is going to become a thing. Anthropic and ChatGPT or OpenAI have already hinted at this. Google has hinted at this. We are going to see operationalized models of recursive self-improvement where models are going to be used to automate large parts of the production of new models. Yes, there are fears that go along with that, but the breakthroughs we get are too valuable for people to run away. And so instead of not doing it, model makers are going to invest in alignment to ensure that the recursive self-improvement loop does not lead to misaligned models getting into production systems.
Prediction number five, we are going to see very long-running agents arrive. This one is so easy. Like, you're going to have agents probably by late Q1, early Q2 that can run for over a day. In fact, we see regular reports from cutting-edge researchers and from engineers that they can get current models to run for 20, 30 hours anyway. So this is barely a prediction. I would guess by the end of 2026, it will not be unusual to have your models running for a full week. And in that world where you're burning millions of tokens, we humans will become the bottleneck. Our ability to review work, our ability to assign work, our ability to have good taste about what we want done is going to become the bottleneck. It will not be the ability of the agents to do work. They will be able to do a tremendous amount of work, not just technical work, non-technical work as well, research work, legal work, etc. So we should expect an army of AI agent colleagues that can do very long-running tasks by Q1 and then rolling out continuously into Q4. And that means we'll all be managers. It means we'll all be asked to ask ourselves, "Can you define the work clearly? Can you keep it unblocked? Can you make timely calls about what is correct, what is not?" We're going to need new technologies to actually look at the agent work in progress because if the agent is running for a week and it goes off the rails on day three, you're going to want to know that and you're going to want to be able to intervene. So there's new technologies we'll have to invent there as well.
Number six, we are going to get massive, massive gains in AI reviewing AI work with human attention going only where it matters most. This is one of the most underrated compounding advantages, and again, I think we have all the pieces in place to do this. We're already seeing automated review of code by AI for AI-written code. We're going to see that pattern extend across work surfaces. In 2026, the big win will not be AI can do the drafts. It'll be AI can audit drafts and ensure that the work product is complete and consistent. It can catch inconsistencies. It can catch missed requirements, risky assumptions, bad architectural checks. Smart engineers are already doing a lot of this, putting work into eval loops agentically where another check is run until the AI agent is able to code correctly and submit a fully working piece of code against five, six, seven, eight different eval sets. Imagine that pattern extending across all of work, not just engineering, and review agents will become normal for all of us. We'll have judge models. We'll have red team passes. We'll have policy checkers. We'll have factuality checkers. We'll have domain-specific linting for reasoning. So the result is that triage as a whole will be a simplified activity. I know that in 2025, a lot of the bottleneck was, "AI can create this, but humans need to review it." In 2026, it's going to move from "AI can create it and humans review it" to "AI creates it, AI reviews it, and humans only put the finishing touches on or look at the final versions that AI passes" because what we're optimizing for in that system is the attention of very high-quality humans, and we want to make sure that they are not overwhelmed by the huge volume of work that we see.
Number seven, I believe work and personal AI are going to split apart hard, and work AI will be heavier, stricter, and to be honest, a little bit less fun. Personal AI systems are going to be optimized for engagement, the way social media was, right? They'll be cozy. They'll be permissive. They'll be optimized for convenience. We will continue to see the absolute explosion of AI-generated ads and AI-generated content on meta networks and other places. Meanwhile, work AI is going to get much more work-oriented, right? It will be governed with identity layers, permissions layers, audit logs, data boundaries, retention rules, who saw what, what was the basis for this output, and the experience is going to feel complex because it, because it kind of has to be. Enterprises will still demand provenance. They'll still demand controls. They'll still demand reproducibility, especially once agents are taking autonomous action. They'll need agent control panes. And so you will feel that separation in your tooling and in your tone. AI is going to become a regulated instrument at work. And for some people, it will be their buddy outside of work. But once you walk into the door, you're going to be expected to behave with AI very differently. I don't think most people are ready for that. And I think one of the safe predictions for 2026 is that that jet lag coming into work every day is going to be a huge shift for the workforce. And that people who are able to understand what work is going to demand of them in managing these agentic systems are going to be incredibly valuable employees, like write-your-own-ticket valuable employees. And people who are interested in AI merely for personal reasons are going to more and more quickly fall behind because they're not going to know what to do to delegate work to an agent colleague and audit that work and intervene in the right ways and ensure that taste is applied throughout and then come out at the end with a useful work product that is 10 or 100 times what they could have produced in a week of work themselves. That is a new skill. The people who have it are going to be incredibly valuable, and the people who are not interested in learning it are going to get left behind, just to be honest.
Number eight, non-technical work in 2026 is going to look a lot more like engineering work does in 2025, but only at the fastest companies. At the frontier, non-technical roles are going to become more and more specific roles. People who move the fastest are going to be able to write crisp requirements. They're going to be able to define success metrics, set up evaluation harnesses, run loops, and manage agent throughput. In other words, they're going to have a lot of the technical skills that we previously thought were only for engineers. I really believe that one of the things that is breaking down is this whole idea that we have a separation in our organizations between code and not code. Everything is going to be code, but code is going to be accessible to everyone. And so, engineering-shaped work is going to be there for all of us to do. And we are the ones who need to skill up so that we understand how to do that engineering-shaped work. And I'm not saying that everyone will have to touch the terminal, although no one should be afraid of that anymore. I'm saying that we will all have cursor for ex-discipline, and we will need to be comfortable enough with our technical skills that we are able to work with that work surface to get agentic work done. In fact, I would argue that Claude Code is setting itself up as an early prototype of cursor for X discipline, except it's for the whole organization. It's like Claude Code just is aiming to become a quick check-in, fast iteration loop, agent layer for the business. I don't think it will be the only one. You're going to have other choices, but I think it's a good example if you're trying to look at what engineering-shaped work looks like outside engineering.
Number nine, the power law of adoption will persist. A few companies will go ridiculously fast, and a lot of them will barely change. I do not buy that everyone is transforming in 2026 because I talked to a lot of people, many of whom work at very, very slow companies. So, we're going to keep seeing a power law where the top slice, the top 1%, the top 5% of companies completely rebuilds their workflows around agents. They ship at a materially different tempo, and so many other businesses are going to be content adding thin layers like Copilot for email or basic summarization and calling it a day. I do think the consequences for that are going to go up. The stakes are going to go up because these advantages compound. So disruptors who want to come and attack companies that are moving slowly and never adapted are going to be able to move so much faster in 2026, even versus the speed they were able to go in 2025. And so you're going to get into a position where you have a company that you thought had stable cash flows and you could just do a mild software rollout, and some startup is going to come in, and they're going to have 10x or 100x your shipping speed, and you will go from a functioning business that has run with stable cash flows for 55 years to nothing in a few months because this business will have just stolen all of your customers. That kind of ambush is going to happen more and more as we move into the future because these businesses that don't adapt to AI will be plentiful. They will be slow, and they won't know what hit them. It's an incredible opportunity for companies that move fast. It's going to feel like the Predator movies where you have a different kind of technology, and you can move invisibly, and you can just hunt whatever you want to hunt. And there will be a few companies that figure that out.
Number 10, machines are going to become proactive, and yes, they will start to prompt us. I fully expect my AI to start asking me to go get coffee because it's noticed a decline in my cognitive output in the last hour or two. So, this is going to be less like it sits there and waits for us to ask, and more like, "Hey, I noticed this change." Or, "Hey, it looks like you're blocked here. Can I help you?" "Hey, this looks inconsistent with the goals we've set together." "Hey, do you want me to draft up options I noticed you're really wrestling with?" Proactivity will be a new product battleground because it's where value collides with our long-term goals and our perception of ourselves. We think of ourselves as the proactive agents. I want us to start thinking of AI as also proactive. And I want us to think about our job as figuring out how to build systems with good proactive taste so that they interrupt at the right time with high precision and with clear actionability. We do not want systems that are quote unquote proactive but end up just nagging us constantly so that we are trained to ignore those systems. Regardless of whether it goes well or badly, though, I have very high confidence that we're going to move in the proactive direction, and that the most productive people will figure out proactive working relationships with their AI systems. In fact, I would not be surprised to see by the end of 2026 some kind of slider on ChatGPT or whatever it's going to be, maybe ChatGPT 7, that will basically allow you to say, "How proactive do you want your AI to be?" And you'll get to make that sort of decision as a consumer. I think enterprises will define the degree of proactivity they want across the surface, and they will expect employees to work within those systems.
And I guess that brings me to my bonus prediction. The need for teams and employees to scale up is going to be greater in 2026 than it has been in all of the years in the previous 25, right? Like, if you look at 2020 to 2025 and you add up all of the training needs, I think 2026 is going to be greater than that. And that sounds hilarious, and it sounds like an exaggeration, but if you think about it, this is changing every aspect of every second of the day for us. And we now have to get an entire workforce retooled. We've never had to do that in 25 previous years. The internet was a much smaller change in the way we worked. We now have to go from, "People are assigned work and maybe it goes faster because of the internet," or "People are assigned work and now they answer their emails and not their fax machines," to, "Who's assigning work? How do you define work? Who manages the work? Do you have a fleet of agents? Are you trusted with agents? How do you lead a team that is composed of agents and humans? How do you lead a team when your humans are not really human manager caliber, but they need to manage agents?" These are the kinds of questions we'll be dealing with up and down the stack. And this is one of the things that makes me really excited for 2026. We won't be bored. We have so much learning ahead because everybody is going to be learning how to do this. And the solutions are going to look different at different scales. How an enterprise handles this is going to look very different from how a tiny startup handles this. And I think we already see some of those models emerging.
So there you go. Those are my 10 predictions plus that bonus 11th, which is, we're all going to have to learn a lot.