This is AGI: Sequoia AI Ascent 2026 Keynote

5,076 words · about 20 min to read

0:02 Good morning. How's everybody doing?

0:05 All right. All right, a little bit better. Hey, thank you all for being here. We really appreciate it. We do this as a service to the community because we are living through important times. And it's an honor for us to be able to serve as a bit of a gathering place for people to come together. And this is by far the best agenda we put together and by far the best set of attendees. And so we just want to start out by saying thank you. We know you are all crazy busy. Thank you for being here today.

0:35 So, we do have a pretty exceptional agenda put together as per usual. Sonia in the impeccable outfit Constantine and [snorts] I are going to say a couple of words to start. You know, we have we have the privilege of being in the midst of a lot of conversations with a lot of interesting people. And so once a year we like to try to synthesize that a little bit and share back with you what it is we've been hearing. And so I'll say a few

AI is a computation revolution

0:59 words of overall calibration and then Sonia will say a bit about what we see today and then Constantine will say a bit about what we think might be coming tomorrow. So, for calibration we're going to start by zooming out. Going back to the silicon [snorts] based transistors which gave this area its name. They got built into systems connected by networks that went public in the form of the internet, supported applications like social media and the cloud, eventually showed up on our pockets in mobile devices that today are capable of doing something indistinguishable from magic which is AI.

1:36 The reason we like to show this slide and those of you who've been here in the past have seen this before is because it reminds us that all of these waves are additive. And we sort of needed all of these decades of evolution to have the compute, the bandwidth, the data, the talent to make the most of this moment. Now this AI wave is a little bit different in three ways. First it's the biggest wave yet.

2:04 And that's generally true, but there is something more specifically true about this wave. Which is it is the first one that is both software and services. The top row shows the first 15 years of the cloud transition where the TAM for software went from about 350 billion to 650 billion and cloud grew to be about 400 billion of that. The bottom row is what is brand new. This is the services revenue that seems to also be available now.

2:34 $10 trillion is a conveniently round number. We don't know if it's 10 trillion or 5 trillion or 50 trillion. We do know that legal services in the US alone is a $400 market. That is one vertical and one geo. And it's the same as all of software. So this opportunity is immense. Point number two fastest wave yet. I think we can all feel this.

3:00 What it means is that this white space, and I direct your attention to the AI side of this page this white space is getting filled pretty fast. These logos are the companies that got to a billion plus of revenue as a result of the cloud mobile and now AI tectonic shifts. And at current course and speed there are more coming soon.

3:25 Point number three which is probably the most interesting one and I borrowed this from my partner Constantine

1 Highlight · AI is a revolution in computation

3:33 is that there are two basic kinds of revolutions in technology. There are revolutions of communication which are about the way information is distributed. Most of the people in this room have only lived through revolutions in communication. The internet the cloud mobile those are all about information distribution. Those were revolutions in communication.

4:00 AI is different. AI is this one. AI is a revolution in computation. It's about how information is processed. And that might sound like semantics, but these are fundamentally different shapes of waves. And maybe the most visceral way to feel this is to think about the fact that the floor keeps moving underfoot. The technology foundation on which everybody builds changes every day when new capabilities come out.

4:28 And we've had three major inflection

Inflection points driving toward practical AGI

4:29 points over the last handful of years. First one chat GPT chat GPT moment, November 2022, the world saw the power of pre-training. Second one, couple years later 01 model reasoning. All of a sudden a second scaling law emerges around inference time compute. Third one just recently cloud code no plus 45 now 47.

4:54 The world saw the power of long horizon agents. And while these look like three points on a continuum it's kind of a heartbreak between two and three. It's a little bit of a discontinuous change. And if we may may be so bold we would say that this is AGI. And look I'm an econ major we're venture capitalists not about to propose a technical definition for AGI. Okay, we study founders and markets and the collision thereof which is businesses.

5:26 But we do study businesses. And so from a commercial standpoint, from a practical standpoint from a functional standpoint if you can dispatch an agent to do a job and it can recover from failure and persist until that job is done I don't know. That feels pretty much like AGI. Even if you don't think it's AGI, which is fine. By the way, Sonia will talk a lot more about this in her part.

5:53 Even if you don't think it's AGI I think we can all see that the cars have arrived. Last few years we've had a lot of faster horses. Applications that made you 10 or 40% more productive, but didn't fundamentally change the way you work. Now we're starting to see cars. Applications that make you 10 or 40X more productive. And absolutely change the way that you work. Change the nature of your work. Change the nature of your organization.

The MAD strategy for application builders

6:20 Cars have arrived. This is the founder of Sequoia, Don Valentine. He was famous for asking one question. So what? Why does all this stuff matter?

6:33 [clears throat]

6:34 Well, it matters cuz just in the last few months the race has begun. And it's a different kind of race than what we're used to. The way you drive a car is different than the way you ride a horse. The way you build a car is different than the way you take care of a horse. So it's a very different sort of race. And one of the reasons that we wanted to gather everybody here today is because nobody has all the answers. And the more time we can spend together, the more we can learn and hopefully figure out where all this stuff is headed.

7:02 And it's important that we do so as soon as possible cuz there's a lot at stake. Just from a commercial perspective there's $10 trillion up for grabs. We got labs coming at it from a tech out approach. We got startups building on top coming at it from more of a customer back approach. We do have all of the labs represented in this room, but most of you are building on top. And so we'll take a minute and uh talk about that customer back approach.

7:29 So our advice for those of you who are building on top of the labs is free advice and so it's worth every

2 Highlight · The MAD framework for building on top

7:35 penny you paid for it. Our our advice would be to get mad. And we don't we don't actually need you to be angry. You can be angry if you want. If that's what drives you, that's cool. Go ahead and be angry. But this is just a convenient acronym for motes, affordance, and diffusion which are three characteristics or three pillars of a strategy for building on top of the models. First on motes just for fun. Anybody remember this slide from last year?

8:03 One person and he's my partner. Okay, that's cool. Um we're we're not going to Okay, as a reminder this slide shows the merchandising cycle which is the links in the value chain required to take something from an idea to a happy customer. We're not actually going to go through the links in the chain. The point that I want to make here if you approach things from a tech out point of view, each link in the chain gets approached a little bit different. If you approach things from a customer back point of view, each link in the chain you approach a little bit different.

8:32 Now here's the part that's counterintuitive. In a revolution of computation which is about information processing what you want to do is look down here because there's cool new stuff coming out all the time. What you should actually do for the sake of building motes is look up here because your customers are not changing nearly as fast as the capabilities are changing.

9:01 The things that you built might be irrelevant tomorrow. The degree to which you wrap yourself around your customers is going to be a bit more durable. That's not to say the product and technology is not important. It is insanely important. And generally speaking best product wins. But in a world where product changes so fast because capabilities change so fast in thinking about motes, we would encourage you to go as customer back as possible and think about all the ways you can wrap yourself around those

9:31 customers. Okay, the A in MAD stands for affordance. This is a term that we borrow from the design world. A hammer is an object that has affordance. I have a two-year-old son. If I give him a hammer he would know what to do with it. He would grab it and start hitting stuff. That's why we don't give him hammers. Okay? An object with affordance is one that doesn't need to be explained. People just know what to do with it.

9:58 Claude code is insanely powerful. Go open up a terminal for the average Fortune 500 employee and see how far they get. While it is powerful, it does not offer that much affordance. That's not a knock in on Anthropic, but it is an opportunity for anybody who wants to build on top. And to create paths of least resistance for your specific customers and their specific problems so that it's just brain dead simple for them to figure out how to get to the outcome that they need for their business. That's the concept of affordance.

10:33 And then finally, the D in MAD is diffusion. And the diffusion gap is the opportunity for companies building at the application layer. The rate at which capabilities are diffusing out into the market is far shy of the rate at which those capabilities are being created. And every day that the foundation models move faster than your average Fortune 500 enterprise, that gap gets bigger and that opportunity gets bigger.

11:03 So, for motes, try to think customer back. For affordance, try to think about creating those paths of least resistance for your customers. And that diffusion gap, that represents your opportunity. Unless that slide from earlier with the white space starting to fill up was discouraging for anybody, may we remind you that no lead is safe. There's [clears throat] this expression in racing. You cannot pass 15 cars in the sun, but you can pass 15 cars in the rain.

11:33 And right now there is a torrential downpour of new capabilities coming out of the foundation models. Which means that no lead is safe, but it also means that anybody can win. What a time to be alive. And with that, I'll hand it off to Sonya.

11:56 Thank you, Pat. And can I just say it's so nice to see so many friendly faces in

Agents evolve from toys to autonomous workers

12:00 the audience. There is an exceptional group of people here today and I'm just really happy to be part of this ecosystem with all of you. And so, the purpose of my section is to talk about what's happening in AI right now, which for 2026 is agents. Okay, flashback to 2022. Show of hands, does anybody here remember AutoGPT or BabyAGI? Okay. Okay. So, these projects were overnight hits on GitHub and what they did was they took GPT-3, gave it some tools, wrapped in a loop, and let it run towards a goal.

12:30 And it was promising until you watched those agents just fail over and over and over again. Kind of cute, kind of endearing, but completely useless. And I put this slide here to remind us that you know, we all knew agents were coming. Uh we could have seen it years ago, but back in 2022, the models just weren't ready yet. Fast forward to today, something around the turn of the year really changed. Suddenly we have agents everywhere around us and they seem to actually be working.

13:02 Two agents in particular have been home runs. Claude code for the technical crowd and open claw and all of its lobster brethren, uh which democratized agents to anybody with a phone. And so, whether you are a hardcore engineer or a normie, the punchline is that anybody can create agents now. And so, what we're seeing is people are building agents for everything. There is silly stuff like an open claw agent that will literally snitch on your neighbors for tax fraud. Please don't do this. Or actually, maybe please do this. Uh there's entrepreneurial stuff. Agents running generative media campaigns to sell construction services.

13:37 And then there's the professional layer. I can tell you there's a huge race internally at Sequoia for who can build the best agents to do our jobs better. So, what does it mean to be an agent? Here is one possible definition. An agent is a system that perceives its environment, chooses actions, and progresses autonomously towards a goal. By the way, guys, I made this in C dance by myself. I'm I'm very proud of it.

14:00 The video models have come a long way. Uh and more specifically, I view agents as having three functional components. First is the ability to reason and plan. Uh this is the baseline level of intuition and the ability ability to think on the fly. Second is the ability to take actions. This is tools, search, write, compile. And then finally, the ability to iterate towards a goal. This is the persistence that gives agents the ability to accomplish things over long time horizons.

3 Highlight · Anatomy of an agent: models, tools, and harnesses

14:34 And so, agency combines these three things. It is simply the ability to get done. If we boil the agents down into their constituent components, the models, the tools, the harnesses, each component has progressed rapidly over the last year. First, the models are the brain. This is the most important thing that's happened. The meter chart measures uh how long a model can sustain progress on a complex task without going off the rails. And we've gone from the order of tens of minutes a year ago to the order of hours today.

15:07 And so, this is this is the most important thing that's happened. The models are finally getting capable enough to sustain per- performance on long horizon tasks. Second, the tools are the arms and the legs. These give models access to things that make us productive on the computer. The terminal for file systems and dev tools, iMessage Slack uh web search, computer use, you name it.

15:33 And the last two decades that we spent building tools for humans have ended up being uh able to port over to be incredibly useful for for agents as well. And there's a common refrain that SaaS is dead. I think to the contrary, the value of these tools is going to explode as the number of agents using them increases. Models and tools give agents capability. The harness is what gives them persistence, the ability to stay on task, adapt, and keep going. And that feedback loop is now really starting to crank.

16:05 Especially now with reinforcement learning, we're giving these agents uh we're taking them to driving school, giving them uh you know, training them uh in RL gyms, and we're pushing performance in different settings from mechanical engineering to design to finance. We're also seeing the early glimmers of self-improvement or the machine building the machine.

16:26 Uh for example, Andrej's other research project uh improves research autonomously towards a GPT-2 level model in just 2 hours. So, what does the world of agents everywhere look like? Agents exist on a sliding scale of agenticness. Um and so, let's take coding as an example. In 2023, we had tab auto complete, one AI assisting a human in line. This was incrementally useful, fundamentally not transformative.

16:57 We now have agentic development, one human talking to an agent, instructing it what to do, maybe managing a team of agents. But this paradigm is getting pushed further. We're now seeing background agents, async agents, agents spawning sub agents. We think that async agents in this whole paradigm is likely to overtake the current paradigm in volume just because of the amount of leverage in the system. And then finally, pushing the bleeding edge of the frontier, what I call dark factories, um taking human review out of the system completely. This sounds crazy, but I've seen it happen in production including with cybersecurity companies. It is possible with good enough guardrails and good enough engineering.

17:38 So, we're progressing up a scale of agenticness. Agents are going from little helpers that do a little amount by your side to interns that need to be managed to interns that manage themselves. And eventually to interns that can be trusted enough to to push to prod without oversight. And so, that's the evolution that's happening not just in coding, but across all of agents. Uh the most important takeaway uh for the founders in this room is that services is the new software. Pat's been saying this for as long as I've known him. Uh and our partner Julian, who's in the audience today as well, uh published a great article on this. And we've known this for a long time, but I think it's actually happening. So, in medicine, you're able to hire an agent that inspects your genome, gives you personalized recommendations, can prescribe you medication, uh recommend you clinical trials. And law, you'll be able to hire agents that can negotiate contracts on your behalf, even perform litigation and settle for you.

18:31 In math and the sciences, we're seeing agents that can solve air those problems or discover new superconductors. Like how thrilling is that? Or in uh in the consumer world, personal agents that can manage your inbox for you, your calendar, your finances, uh file your taxes. And we expect there's going to be agents everywhere and that's in part because hiring agents is so much easier than hiring employees. Humans are hard to scale. Infinites are infinitely scalable with compute.

18:59 Humans are hard to keep happy, except for me, I'm always happy. Uh agents are low maintenance. Uh humans are uh humans are expensive. You pay them salaries. Uh you pay agents tokens. Generally, it costs less to accomplish a task with tokens than the equivalent in salary. Today, humans are still generally smarter, but the bitter lesson presses on and soon agents will be smarter at many things. And so, the point of this slide is not that we humans are out of a job. Uh I think a uniquely human trait is adaptability. But we do expect the deployment of agents across the application layer to be swift and at an unprecedented rate and scale because the economics are so clear uh and because of the inherent scalability of bits.

19:43 So, if you add all this up, the number of agents is ballooning on some sort of exponential, maybe super exponential. And I think we're about to hit the point where things get genuinely [clears throat] strange. Uh what happens when commerce happens between agents? Can they pay each other? What happens when agents can actually negotiate the terms of a transaction with each other? Are we going to have swarms of agents policing us, preventing things like cybersecurity Armageddon? All we know is the world is getting weird extremely quickly.

20:11 And so I'll close by channeling my inner my inner Bene Gesserit. Uh long horizon agents are here. The curves that they're on is very clear. And for founders, I think everybody has examples of people that are accomplishing insanely hard timelines thanks to AI. So Nathan from Zed accomplished a three-year moonshot project over the holidays by himself with cloud code. Uh Bret Taylor rebuilt Sierra over a weekend. Um the Notion team rewrote 8 million lines of code in just 6 weeks. And so everybody has these examples of compressed timelines. But I think very few people outside of the AGI labs have seen what happens when you take these compressed timelines and you stack them on top of each other.

20:52 And that's what's possible now. And so whatever you could imagine building over the next 100 years, we think is now possible in 100 days thanks to agents. I will pass it over to Konstantin. Thank you, Sonia.

21:07 All right. Thank you so much, Sonia, Pat, for the brilliant overview, analysis. In this section, we're going

Cognitive work mirrors the industrial revolution

21:13 to talk a little bit about what's next. Uh so the goal here is we all know we're in an AI age. What's it going to look like? What's it going to feel like? How's it characterized? Earlier in the presentation, Pat bifurcated technological revolutions between compute and communication. We're going to do another bifurcation here for types of work. There is physical work. This is a package on the Pony Express. This is a satellite on a Falcon 9.

21:44 Work equals force times distance, physical movement. And then there's cognitive work. This is Pythagoras coming up with his theorem. This is DeepMind solving the pro- protein folding problem. Conscious thinking. These are very different types of work. But we believe that they're going to follow a very similar pattern in revolution. So let's talk about physical work because we've been through this revolution with the Industrial Revolution.

22:13 For the vast majority of human history, all the work or virtually all the work for serving humans was done by some sort of muscle. People or animals. People moving something or an animal pulling the human along. This starts at 1700, but it goes back millennia. Then things started to change.

22:40 Water and wind. Steam engines. And then things accelerated. Steam engines, combustion, electric motors. Today 2026 you could estimate that 99 plus percent of all the physical work done on planet Earth for humans is done by machine. The plane that brought you here,

4 Highlight · Cognitive work mirrors the industrial revolution

23:06 the manufacturing of all the goods in this room, all the transportation that sets up for the pinnacle of the human experience you're having right now. Well, we think a similar pattern's going to happen in cognition. We're just a little earlier on.

23:23 [snorts]

23:23 So for most of human history, all the thinking on planet Earth for humans was done primarily by humans. Maybe a little bit for animals. The the sheepdog chasing the sheep. Right? And there was this sliver on top of mechanical work, the astrolabe or the clock. Now, over the past couple hundred years, there was not a lot of progress until electronic computation.

23:52 And in the past 100 years, think about all the trillions of calculations that are happening at any given moment to serve you the human. All of that work, all of that cognitive work that's happening to serve us at any given moment. Trillions of calculations. We believe that the neural network is the next big wave. And that in the near future, 99.9% of cognition on planet Earth will be done by machines.

24:27 Well, the parallel is pretty stark. And the good news is we've been through a revolution like this. The cognitive revolution is going to be a lot like the Industrial Revolution.

Navigating a future of alien design and intelligence

24:40 Just much, much bigger and much faster. So what's it going to be like living in this future? I'd like to share some motivations for this future in the form of four [snorts] short stories. The first story. In the mid-1800s, America wanted to build a grand monument to our first president and our greatest war hero, George Washington.

25:08 So we designed the tallest building in the world at the time, the Washington National Monument, and we wanted to cap it with the most precious metal in the world, 100 oz of the most precious metal in the world. So precious, in fact, that we put it on display at Tiffany's in Manhattan. That metal was aluminum. Within decades of the completion of the Washington National Monument, a young inventor came up with electrolysis, the process of separating aluminum from dirt.

25:44 And within decades, aluminum was used to wrap our candies and our sandwiches, and then tossed into the trash. Aluminum is intelligence. Electrolysis is artificial intelligence. We're about to enter a world where some of the most precious skills that took decades to earn, PhD-level skills, are so instantly invoked that right after using them, you can crumple them up and throw them right in the trash.

26:26 Story number two. We are entering a world of alien design. The world as we see it today is all about design for humans. You know, it's it's been optimized in a way that makes sense to our brains because we are doing almost all the cognition in the world. Well, when machines do the cognition, it's going to be a little different. In 2006, NASA was optimizing an antenna for a large space mission, satellite space mission.

26:54 And traditionally, their antennas looked like this. It was a beautiful geometric, symmetrical pattern that optimized surface area for some power constraints. This time around, they said, "We're going to hand it over to computer and we're going to have an evolutionary algorithm." A lot like reinforcement learning. The result, this antenna right here. Dramatically more productive. Not intuitive to the human mind.

27:21 In this AI era, when we hand over cognition to machines, we're going to get results that are not intuitive to us. When AI's designing chips, cars, buildings, they might look dramatically different. The world that we enter into, we have to be open-minded because the AI is not going to think like us. It's going to have alien design. The third motivation story is on emerging sciences. Not emerging science.

27:53 We all know there's emerging science. I'm talking about emerging sciences. In the early Industrial Revolution, you had great engineers like Newcomen and Watt. And they perfected combustion engines. Basically, put a petrochemical into a piston, ignite it on fire, millions, billions of particles explode, move the piston, work. For almost 100 years, all of that was tinkering.

28:21 It was an engineer saying, "Ah, that works a little bit better." Maybe something you could see see like a scaling law, but it was engineers playing with the product and seeing how they could improve it a little bit. Over 120 years after Sadi Carnot came around and formalized this in a new science, thermodynamics. He said, "Wait a second.

28:46 There are millions or billions of particles. We can actually formalize what that all looks like." In this case, there are billions of neurons, trillions of tokens. Right now, we're in the tinkering phase of AI. Even if we think it's an understood science, it's not. In the future, we will have a science as fundamental as thermodynamics introduced in the next couple decades. Someone in this room might come up with that science. And that science will be taught in high schools. It will be that fundamental.

29:15 And it will help us master AI. It will even help us master consciousness. Fourth story. The art of unreason. So for the vast majority of human history, tens of thousands of years, art has been a progression towards realism. This is a cave painting from about 25,000 years ago.

29:41 Egyptian hieroglyphs. Greek pottery. Uh Renaissance paintings. uh a grand transformation toward realistic art. Just look at the difference over tens of thousands of years, the triumph of humanity. And then, engineering came along. The daguerreotype, early photography.

30:10 And all of a sudden, what was spent decades of life to perfect the skill of getting every brushstroke perfect, done. So, how did the world react? They thought that painting was over. Oh, that's it. The machine can do it better than any human. Art is Art is ended. Well, what happened? How did humans respond? Humans responded by saying, "Was the purpose of this art to capture the moment in the way the eye sees it?

30:43 Or was it to capture the moment in the way the heart and the soul sees it?" Impressionism expressionism [snorts] cubism, neo-expressionism. All these new forms of art are how humanity responded to this dramatic change in science. 2,500 years ago, Greek philosopher Protagoras wrote, "Man is the measure of all things."

31:17 What he meant is that nothing in a vacuum has value to humans. Not aluminum, not art, not intelligence.

31:30 [snorts]

31:30 It only has value because of the experience. AI can do the work. AI will do the work. But, only the human connection can give you a reason to care. That's why we're all in this room today. A decade from now, work is going to be dramatically different. Things are going to change so much, but the one thing that will be constant is the relationships that you form today with the person right next to you will endure.

32:06 That's what you're going to look back on. That's what's going to be valuable from today. So, I encourage you to form those relationships with the people next to you. Enjoy your time together at this AI Ascent, and really lean into what makes us most human.