Thank you very much, Phillip. So I wanted to get started by giving some kind of frame on what I'm going to talk about. And really this is a talk about how to think about and deal with the success of AI as a technology and to think about how its continued maturation might change us as individuals and as societies. And I think that the rapid advance of AI as a technology presents us with a choice which is whether we explore the future or retreat from the present. Exploring the future requires us to reckon with the fact of continued AI progress and ask ourselves what we want to do with this technology as it becomes more powerful. Retreating from the present is when we ignore the implications of the technology and dismiss it. And retreating from the present forces us as individuals and as a society into states of reactivity or passivity in the face of AI's continued advance.
In the coming years, we're going to have to make many decisions as individuals and as societies about how we want to shape AI, how we want to use it, how we want to direct it, and how we want to distribute its benefits. But making these decisions requires us to reckon with the power of the technology and see the future that its continued advance implies. So in part one of this talk, I'm going to try and outline for you what I think the past few years of AI progress have looked like and discuss why if the technology advances as much as I think, AI can't be treated as a normal technology. In part two, I'll try to make sense of these advances through the lens of my own experience with the technology as well as changes that are now beginning to occur at Anthropic as we start to use the technology at scale. And then in part three, I'll talk through some of the sort of humbling and almost unimaginable implications of it.
But fundamentally, I think I need to talk about what planning for success of the overall endeavor of AI actually means. By success I mean that we're going to succeed at building increasingly powerful systems, potentially ones that eventually build themselves. And it's time to plan for this because AI systems are likely to get better a lot faster than people expect. And as they become more advanced, we should expect profound changes to happen to people and society. And to understand why I'm thinking about success so much, I want to do my favorite hobby, which is walk us through an overly technical, impenetrable graph about AI progress. Everyone's excited. Great. So, we're going to look at a graph that tries to represent all of AI progress and then I'm going to decompose it for you. So, who here has heard of the Epoch capabilities index? Wow, that's actually like way more people than I thought. Congratulations. Well, we're going to quickly go over it and then we're going to get into an attempt to explain it. But really, a lot of this section of the talk is about my uncomfortable relationship with this particular graph because in the graphs that encode information about AI progress, we see the pace of AI progress. And I think the longer you work within AI technology, the more you realize that that pace of progress is unusually fast, has unusually strange implications for society, and needs to be taken seriously, even though the implications are kind of mindbending.
So here's the graph. What is it? It's a bunch of dots on a line going up. You've seen many like this before, but what it's really showing you is the aggregate score on about 40 different benchmarks for AI models. And what you see is this trend in recent years of the line going up and to the right. And what this really represents is a general advance in the capabilities of AI systems. Graphs like this are sort of functionally speaking very very hard to read and get a sense of a bone deep sense of what they actually imply. And so what I'm going to try and do for you instead is talk through this graph through the lens of what AI systems have actually been able to do in just the last few years and give you with that a sort of visceral sense of progress. So let's get into it.
You know, March 2023 for the first time AI systems pass the bar exam on their own. June 2023, we have an AI system that finds a faster sorting algorithm than any human has in 50 years. November 2023, researchers, a bit alarmed by the pace of AI progress, introduce an exam called GPQA, which tests out how well AI systems can do at graduate level science. And AI systems get about 40%, human experts get about 80%. January 2024, AI solves olympiad geometry at the gold medal standard. July 2024, AI reaches IMO silver and solves the hardest problem in that part of the exam. Then by October 2024, Anthropic debuts computer use showing how AI systems can now use and operate machines like humans. In January 2025, researchers, I think frustrated at the progress against things like GPQA, introduce a very frighteningly named exam called humanity's last exam. We'll see if they're proven right. And the best system gets about 13.4% on it. And then May 2025, AI breaks a 56-year-old mathematics record of basically finding a way to multiply 4x4 complex matrices in 48 operations instead of 49, which was terrifically exciting to a small set of people. July, AI hits the gold medal level at the International Mathematics Olympiad. So we're up from silver a year before. And these are extraordinarily challenging things to do. If you're a math geek, you can attest to this. If you're not a math geek, ask one you know and they will explain to you that it's difficult. In October 2025, AI co-authors a genuinely new mathematical proof because people have started to use AI systems to do frontier science. And then in February, Anthropic finds 300 or 500 severe software vulnerabilities with a model we released called Opus 4.6 and Google Gemini 3.1 Pro gets 45.9% on humanity's last exam. So in a year we've gone from 13% on this supposedly last final exam for humanity to 45.9% which suggests both the exam is too easy but also systems are getting good really quickly. And in April 2026, Anthropic releases a system called Mythos, which is capable of discovering novel flaws in software at such an improved rate relative to previous systems that it's being treated more like a nation state cyber security capability than a typical system. Oh, and it gets 94.6% on GPQA diamond, which when it was released, humans got 80% on and the best AI systems got 40% on two and a half years earlier.
That's what success looks like. Success is extremely abnormal in this genre. All of these are things that would previously take individuals or teams of people working together immense amounts of effort to build single systems that were capable of these. But many of these results have come from singular AI systems doing many many many of these capabilities that I've just talked about. In fact, what we've just talked about are the kind of individual trees of AI success. But these trees are all part of a forest. And this forest is growing in size and breadth with every passing moment. In fact, I think the growth rate of the whole forest is increasing over time.
So this talk rests on this idea that perhaps that will continue. Perhaps the kind of progress I've just described will continue into the future. And it's hard for me to give you arguments for why it wouldn't. This technology that has just done so well on so many of these challenges is based on a common underlying technology of neural networks. Performance keeps growing somewhat predictably in direct relation to the resources invested in it. And as we all know, the companies that are at the forefront of this technology and their cloud partners have now invested literally hundreds of billions of dollars into building out the infrastructure to train future versions of the technology. So it feels like some amount of future progress is already locked in which means we need to be eyes wide open about what the continued success of this technology means and its implications. So I should be clear about my own relationship to it and my own feeling on it. AI is a tremendously powerful technology and it's getting more powerful all the time. It is a technology that is smarter and more capable than most of us as individuals today and will soon be more capable than all of us collectively. It's a technology that we do not fully understand because it's more grown than made. And it is a technology that you can concoct plausible scenarios where it could kill every single person on the planet. So to think building this technology is without risk would be an act of hubris or insanity.
Not something Silicon Valley is famous for. And yet building this technology is one of the best ways that we as a species can advance ourselves. We're drawn inexorably to build it because it is so powerful and so useful for so many of the domains that we want to improve the quality of life as a species and what we do collectively whether that be science or healthcare or medicine. But we should expect that this rate of progress will continue and in fact we should expect the rate of progress to soon compound on itself because it's likely that in the coming years we'll build a version of AI that is capable of building its own successor. An AI system that gets smart enough that it can design the next AI. It's almost analogous to if you were a 3D printer company and you'd built a 3D printer which could make a finer resolution print head without any outside technology. That would be very very strange. We have not ever encountered a technology that has this property where it can make a better version of itself without people having to contribute ideas or concrete resources to help it do that. And it's going to come soon. I think that this could be built in the next two or three years and that will generate even more advances of the flavor we've just discussed and broaden even further the capabilities of us as people and societies and deepen the way in which AI shows up in our lives.
Coupled with this will be immense change of a magnitude that I believe none of us have experienced in our lifetimes and yet huge amounts of change is on route. The technology is in fact so powerful that I should clearly state that if it was possible to elegantly slow the development of this technology to give ourselves more time as a species to deal with it, that would likely be a good thing. I mean I love change as much as the next person, but I think we can have slightly too much change and it might be better to slow it down somewhat. But in the absence of a coordinated global slowdown, we are left with the current situation, which is a powerful technology being developed at breakneck speed by a variety of actors and a variety of countries locked in a competition with one another where commercial and geopolitical rivalries are often drowning out the larger existential to the species aspects of the technology being built. This isn't an ideal situation, but it's the one we find ourselves in. And the question I'm struggling with now and hope to sort of talk through in the rest of this talk is how do I get my mind right with living through what might be the singularity and how do we all sort of cope with this immense thing that is rushing towards us.
The amazing thing is that AI and I'm sure it has done the same for many of you has already meaningfully changed my life or made its way into my life. It's also starting to cause larger changes at Anthropic, the AI company I'm a co-founder of. And I thought we could talk through some of this by returning to the graph we looked at before, but this time by looking at it through the lens of my own usage of the technology. So, after I'm an English person, after you move to California, you end up wanting to talk about your feelings, which is what I'm going to do in this section of the talk now.
So, I looked closely at how I'd been using Claude in my own personal life, and I think it tells an interesting story. You know, in 2023, I'm using it to check my work for typos because, of course, it's a okay copy editor. By November, I'm using AI to help me figure out what foods to feed my baby. I'd recently become a father. My baby had started eating solids. I said, "Is avocado good for a baby?" Claude says, "Sure." Yeah. I mean, how complicated is this? January 2024, I'm using AI to actually help me understand my marriage as it had changed with having kids, where I was writing long messages to Claude along the lines of, "We've just had kids and it's completely changed the dynamic of our marriage." I was also talking to my friends, but it was helpful to talk to Claude about it. June 2024, I use Claude to help scrape my own newsletter. I publish a newsletter. It's created this archive of millions of words. I've always wanted to write a web scraper for it. I've always been busy. Claude helps me do that. August 2024, I have AI write a text adventure game for navigating AGI and the transition to super intelligence for me. This is how I relax, I suppose. November 2024, I actually try to reimagine my job using AI. Every few years, both at Anthropic and before, I write a doc called like, "What's Jack doing with his life?" And I share it with a bunch of people and use it to examine my own life and career. This time I shared it with Claude.
January 2025, I ask AI how to prepare for super intelligence. It's not very smart. Doesn't give me good ideas, but at least I tried. February 2025, I use AI to generate code names for AI projects in my fiction. So, I write fiction stories, and I felt like it's aesthetically legitimate to use AI to generate the code names of AI systems that you write about. Felt interesting to me. And then in May 2025, I talked to AI about the quite non-trivial stress one feels as being a founder of an AI company when you stand up to audiences and say that you think the technology you're responsible for could kill everyone on the planet. Little stressful. You talked to Claude about that. And August 2025, Claude says, "Maybe you should go back to therapy. Why are you talking to me about this?" I was like, "Good job, Claude. Best thing to do is sometimes is send a person away from the AI and send them to someone's office." But when you go to a therapist and say that AI told you to talk to them, they raise an eyebrow. November 2025, I use Claude to research the history of kind of S-curves in advances in different forms of technology. I'm interested in the rate at which solar panels have improved, the rate at which the reduction of the cost of getting mass to orbit has improved. Like a real research project that otherwise I would burn a lot of time on. And I asked Claude to do it. January 2026, my kid has started reading, though they read upside down and they can't actually read. And I say, "Well, they seem interested in reading. How do I help encourage my toddler to read more?" And then March 2026, I track performance for how AI is used in kernel design across tens and tens of different papers. Kernel design is a very wonky way that people optimize for performance of AI systems, but it's something that you're interested in tracking if you're someone like me. Claude does a good job. And then May 2026, I'm writing a short story. And I think that the language models have finally got good enough to write text that is still a little clunky in a fictional sense, but interesting. So I have it generate the speech of a language model which I am writing a fictional story about.
So when I think about my personal experiences, the ones I've just described, I think you see that AI is making inroads into my life in increasingly personal ways. My whole relationship with this technology has gone from using it very much as a tool to something which I kind of ideate with, which I share intimate details about my personal life with, which I use as not just a tool but a kind of virtual employee that works for me. And it kind of amazes me that these AI systems have got to the level of capability that I can describe this breadth of use and it's honest, this is all stuff that I've done and it's all stuff that I've done with a single system and I think that should emphasize to you that we now have these singular systems that are not just capable in a scientific and a technical sense but they are generally capable in what feels like a human sense. This feels a lot like talking to a knowledgeable friend about a broad range of issues and they even do knowledgeable friend things like say why are you asking me, go talk to an expert which is a helpful proof point on Sycophancy perhaps.
Now, the most amazing part of this is that I think that we're about to enter this era where the growing capabilities of these systems are going to allow us to go far further than we have before in how we use this technology. And I think that has many implications for society. So, I'll talk about this through the lens of my personal hobby, which is a newsletter called Import AI. Who here has read Import AI? Thank you very much for reading it. Sorry about the typos. Those are Claude's fault. The newsletter is in its 10th year. It's my main hobby outside of work. And in the newsletter, I read research papers about AI and I work hard to understand them. And once I feel I understand them, I write a summary and a note on why they matter. And each issue of my newspaper contains a bunch of these plus a fictional story where I wrestle with the implications of the technology I'm learning about and building. And recently I had this kind of revelatory experience where I was putting together data for a post I was writing in my newsletter about automated AI research and development and recursive self-improvement. And I pointed Claude at the archives of my newsletter, the ones which I'd had it scrape earlier in the story and asked it to pull out references of all of the times I'd ever written about AI R&D. And since the corpus of the newsletter is now about two million words, this would have taken me a while, taken me a lot of complex Googling, but Claude did extremely well.
What was interesting to me though was what happened next where I then asked it to make graphs for me of all of the sorts of data that was in these research papers that I'd written up. And as you can tell, I have this unhealthy relationship with graphs. But nonetheless, it's part of my hobby and it's why I like to exercise the AI systems for. And what I found was that just by asking Claude to make graphs for me, it could kind of take what it had learned from what I wanted it to do in the newsletter and go and in an autonomous way build me kind of a dashboard which I then used to then write some of the essay, which is basically farming my newsletter for all of these research papers I'd written about and data points I'd covered, then reading every single research paper which referenced any one of those benchmarks and putting them all onto points on a graph. And this for me was a very exciting experience because I was now able to kind of explore one of my hobbies and one of my passions in a different way. But then I realized I could go much further. I could convert this thing I'd asked it to do which was to read a bunch of papers and turn them into graphs from something I was doing just for a single essay into a general skill. And I made this skill which is basically born out of my own hobby, my own amount of data that I've created through my own creative practice has allowed Claude to understand something very important to me which otherwise I spend a huge amount of time doing and turned it into a skill where I could then go up to Claude and say that's interesting, go and make 20 more graphs like this. So it went away and it read a few hundred papers and came back with 20 more graphs. And though my hobby is reading AI research papers, there are so many of them that I don't read all of the papers that are published now. And as I looked over this dashboard, I had this totally thrilling experience where I'd seen graphs of papers that I had never read. And so now Claude was able to help me explore this gigantic academic surface which I've spent years immersed in and was finally telling me something new. It was telling me things that I had never seen before or otherwise wouldn't run into. And through this I learned about some new primary source material to read which I did.
Now I want to convey to you how strange it is that an AI system can do that and what the implications are because I understand at a bone deep level just what it takes to make one of these graphs. You read a bunch of papers, you go hunting for common measurements within them, you read many different caveats in each paper and figure out which of the metrics are just wrongly placed in there of which there are many. And then you have to double and triple and quadruple check your work. It is an unbelievably laborious process. In fact, in 2016, I'd been handmaking graphs like this when I just started at OpenAI where I was charting the progression on computer vision or object identification. And I ran into some academics who are starting a project at Stanford University called the AI index. And when I showed them my graphs, their eyes lit up and they were like, you have these sacred hand assembled graphs which would have taken us hundreds and hundreds of hours to do. Come on in. And I became a co-founder of that project where the coin of the realm was these graphs which I'd assembled painstakingly and painfully over many many months. In many sense they're a proof of work.
So the amount of time that goes into making something like this is actually extraordinary. I hope I've conveyed to you that this is not an easy thing to do. It is a thing that takes immense effort to do. And now having it be a system that can do it in a repeatable and generic way thousands of times faster than me feels very strange. Feels like the implications of this must be strange as well because I have this bottled up skill where I can harness the power of these AI systems to go and do something for me that I know would take literally weeks of work. And it can do it for me in minutes and it can do it for me for anything. So this morning I was using this to generate and I'm not going to have you read these graphs, just generate a bunch of graphs for me about the intersection of AI and robotics. And then I asked it to do a bunch of graphs for me about the intersection of AI and biology. And then I asked it to do a bunch of graphs for me about how AI is being used in computer vision on drones. And so that just happened this morning. And I was watching this system generate probably in aggregate thousands of hours of human work in a matter of minutes.
It must mean something that systems can do this, right? Like how do you reconcile these capabilities with the world staying relatively normal? I think it's quite challenging to do. And there's something empowering and amazing about this because I've turned my highly idiosyncratic passion into something that can be distilled and handed to a machine which can then go and work on my behalf. But it's fundamentally to me an illustration of how AI if used right can let us truly explore things together. We can explore intellectual spaces that were otherwise hard for us to access. We can do things that would otherwise take us immense amount of time and can be deeply inspiring to do. And it provides a greater incentive for me to actually continue my practice of writing my newsletter. Because now by doing this primary thing that is kind of deeply human and deeply coupled with me, I'm creating something of essentially value through which these systems, these machines of great power and consequence can better help me and better help me do the things that I want to do. And it relies ultimately on me producing something that is idiosyncratic to me. So, I think that's an optimistic sense of how this works.
That's just me, though. I think that the important thing to recognize is that once you have capabilities like this, they're also going to change how people collaborate, how organizations work. And so, there are changes afoot inside Anthropic, the company of which I'm a co-founder, which I think speak to larger changes that are going to come to society. So recently I had the great joy of having my wife having our second child and I got to take paternity leave which meant for the first time in perhaps 5 years I'd stepped out of the bubble that is Anthropic. And you know, because on my first paternity leave I was very bad and I just worked the whole time which one should never do but this time I said let's take two months off, could have been longer. So I took the time off and I left in November 2025 and when I came back in February weird stuff had started to take place. You know, Anthropic had released a model called Opus 4.6, which I'd played around with while I was out on paternity leave. But when I came back, I saw that the model had got so good, it had completely changed how people worked inside the company. Some of them were no longer writing code at all. They were just sort of instantiating this model in tools like Claude Code and having it do versions of what I just showed you, going and doing a version of their work hundreds of times faster than they'd done before. Doing it in a repeatable kind of creative and exploratory way. And then their work had shifted to managing the work of Claude and checking its outputs more than doing the work themselves.
This has also led to fun things inside the company like people posting on Slack, "What do I do when my Claudes are working for me?" The answer is get back to work. But you know, we'll see. In Anthropic, much of the work that needs to get done involves writing software which is made out of code. And so this significant increase in the automation of the production of coding has been equivalent to dropping many many more employees inside the company. So now we're in this situation where there's a massive rise in the amount of code being produced inside Anthropic and we've needed to basically change our investment in the tools that are the rate limiter on that code being useful. We've sped up part of the production process that the company is concerned with and now we found other slow pieces in the chain and a lot of us humans are spending all of our time trying to figure out how we optimize and automate the things that are otherwise slowing the AI system down. In many senses, Anthropic and the AI systems we operate look increasingly like a kind of ecology filled with independent AI agents running around and doing stuff for us than actually an organization mostly composed of humans. And the task ahead of us is to figure out how to measure and observe that ecology and figure out what is normal, what is not, and what we want to sort of validate and verify. And I think this maps to a brewing theory among economists that if success continues for the AI project, then some of the work of people will shift to figuring out how to validate the outputs and price the kind of operational risk of running increasingly large automated systems in society. So the more we add AI automation, the more people are going to spend time at these higher levels of abstraction running these incredibly large automated systems which behave in a very kind of organic and emergent way.
An additional thing is that we're having to solve the problem of human validation of AI output. So Claude doesn't just write code. Claude now writes documents which introduces this puzzle where as I'm sure people here are aware writing is hard and people used to use writing as a costly signal of effort to other people but now that Claude can make writing easy we're having to figure out within the organization what is the role of people in validating or verifying Claude's outputs. Is it as simple as saying this is good? Is it more complicated like you need to write a cover sheet on top of what the AI produces? These are the kinds of challenges we're figuring out and there will be an answer somewhere but there's also clearly a new set of norms that we're building for how we work with this technology.
It's also led to discussions about hiring where it's too early for me to come here with like granular data on how hiring patterns at Anthropic have changed, but it's certainly forcing us to ask the question of what does hiring look like now that we have these systems which are massively increasing in capability and skill of what they can do. It means that there's an emphasis on hiring early career people who are extremely well-versed in AI systems. You know, people who have grown up with language models, people who were perhaps 15 when the original GPT series were released and are now 20 or 21 do amazing things with this technology that older people just haven't thought of because they aren't native to that, similar to the early internet boom. It's also creating a situation where there are returns to people that have experience and intuition where my ability to generate all of these graphs that I find interesting comes from spending 10 years developing some intuition and experience about a specific hobby that I have. So the returns to specific expertise seem to be rising up as well because these people can effectively manage larger sets of people. It also helps us hire more interdisciplinary people than before.
But I think that we'll experience much more radical changes as an organization. We have one paper we published recently on automated alignment research where a single person was able to basically set the proverbial research agenda of nine distinct AI systems which went off and worked for them. And these AI systems really did do all the work because the privilege of working at Anthropic is I go and check with the researcher and say what really happened and the researcher confirmed to me that all they did was basically tell the systems here are individual areas to go and explore in this specific part of alignment research we're interested in, go. And they built an automated system through which these systems were able to go and do genuinely useful scientific work. And there's no evidence this is a fluke. It more feels like the new normal. So there is going to be situations where teams of people operate on top of a kind of pyramid of digital labor which massively scales their own effectiveness and allows them to move faster and do more than other people.
And finally and perhaps most importantly, I've seen that this experience of AI progress has really caused us to actually ask what is the purpose of AI? It forces you to confront the technology most directly when you're seeing it be sort of deployed at scale within your own organization and fundamentally change the organization and force you to grapple with questions together after you're exposed to an AI system that is better than you. It's something that you deeply understand as I have been with seeing it work on my own personal hobby and as my colleagues are in seeing it work on their own day work. You spend more and more time sort of meeting up and talking about the meta of where does this end, like what is the actual end point of continued progress here because it is already massively changing our organization. One of the largest tasks is becoming figuring out how we can productively get out of the way of these systems and let them do more work because often humans are becoming the slow links in the chain which surely has some implications. So the questions many people are now asking themselves is how do you build teams that scale in relation to this new technology and what is the role of
People talking to one another and validating or verifying the outputs of this booming kind of machine economy inside the company. The main lesson I'd take from this is that Anthropic is attempting to explore the future with Claude. We're aggressively using the technology throughout our organization and we're trying to change our organization and how we work ahead of the arrival of more advanced systems. By comparison, much of the rest of the world seems to be in denial about the capabilities of AI systems today, let alone where they'll be in six months. And to me, this feels like an unhealthy or a dangerous pressure building up in the system where at some point you become surprised by the capabilities of these systems. And in an environment where you haven't prepared for it, it surely will be more chaotic and less positive than one where you have actually tried to grasp the future implied by these systems and build ahead of it.
So, we've talked now about how AI has progressed in the last few years and also how the advance of AI is showing up for individuals like me as well as organizations. And I'd like to return to the graph for one last time and extend it forward and try to make some predictions about the world ahead of us based on the technical trends which I've just described and my own experience of using the technology. I think that we can make some predictions that are quite extreme but also well founded in the technology. In the same way that AI systems recently got good at things like hacking with things like Mythos, we should expect AI systems to get good enough at biology that they're highly relevant to both advancing science and potentially proliferating bioweapon risks. That seems like a natural consequence of the steady improvement in capabilities on science that I've described. I think in April 2027, it's plausible that a team of humans and an AI system will make a discovery that will subsequently get a Nobel Prize. We already see some of the world's best scientists using the AI systems deeply in their work. Seems likely that at some point soon there's going to be a team that cracks an important problem, AI played an important role, and we'll see a Nobel Prize awarded to a hybrid human and AI team shortly after that.
Later next year, I think we'll see autonomous companies come into existence which generate tens of millions of dollars in revenue. And we'll see multiple human and AI companies exist which are doing on the order of hundreds of millions to billions of dollars in revenue because it's a person sitting on top of a field of thousands and thousands of colleagues. April 2028, I expect that we'll have bipedal robots that will be beginning to do useful work in the real world in partnership with humans. Today the robot hardware has improved a lot, but the robots are still comparatively dumb. The steady advance of AI and its capabilities and its ability to distill itself into smaller forms surely implies great changes will come from there. And finally, I think by the end of 2028, and this may even be a conservative view according to some people in the room, AI systems may be able to autonomously design successor AI systems that are better than them on things that we care about.
Personally, I expect significant changes in my life as well. Just looking at how I've been using AI and rolling it forward and assuming the same trends continue. I think that later this year some chunks of my life that are probably annoying admin formulaic portions will be autonomously managed by AI systems working for me because why wouldn't I? This is horrible, ugly work and I'm happy to give it away. But next year, I expect that for the next time I write some 'what is Jack doing with his career? What is Jack doing with his life?' it's not just a doc that I ask Claude for advice on, it's one where I feel like there might be a situation where through discussions with an AI system, I make massive changes to my career based in large part on discussions I've had with the AI system rather than a person. That is the trend that I and many other people have been on. I think that my media consumption might be set to change. I'm a huge reader. I love reading poetry, science fiction, the whole gamut of things. And yet I am also a horrible nerd that likes living in certain fanfiction universes of which there is not enough fiction for me to read. And I think some people in the room will have had this experience. Surely at some point AI will get good at generating fictional universes for me that are customized just to me and I might enjoy them more than the science fiction stories I currently read. Not currently the case, but it doesn't seem implausible that this lies ahead.
Later, I expect I'm going to learn an entirely new skill through customized tutoring via an AI system. I already get advice from these AI systems. They are already able to understand something about my life and teach me things. Why not extend this further? And finally, I expect, around the time AI systems are capable of AI R&D, that AI will help make me make conceptual breakthroughs that change the course of my life. Things equivalent to when I realized that I needed to go and move to America to write about and eventually work in artificial intelligence or before that when I started realizing I needed to study distributed computing in the data centers of companies like Google because I thought it was a precursor to machine learning starting to work. These were conceptual updates for me that I had mostly through reading and being a journalist, reporting, talking to people. Why shouldn't an AI system be able to give me that same conceptual breakthrough? It seems likely.
So when I think through these predictions, it's so hard for me to reconcile the continued advance of AI with the world being normal or myself as an individual remaining the same as I am today. I can't make that story make sense already. Today I feel like the world is starting to change as a consequence of this technology and all we need is for the progress to continue for greater changes to happen. And I think that if we accept the sorts of futures being described here, we should expect that there are various parameters of the world we live in in the future that are different to those of today. I think we'd expect compounding wealth in the machine economy to drive a massive boom in economic activity the likes of which we probably haven't seen in generations. I think we'll also see along with that, which will be very challenging, the colonization of big fields of human work by machine workers, autonomous or synthetic workers that compete with humans in what we currently specialize in. And I think we'll see a sudden and extreme rise in the rate of scientific advances. Already, you can see this in the research literature as many teams of humans around the world have their science be tremendously accelerated by the addition of these general purpose tools.
So what are specific predictions we can make about the future and how do we think about what it means for us and sovereignty? I think we'll see a massively changed economy. As I've said, it's impossible for me to reconcile the world ahead of us with the world of today given the technology. We should expect unprecedented things to happen. I've talked about autonomous companies. I think you'll see human and AI companies which almost look like five humans and 50,000 AI systems working together. I think you'll eventually perhaps even see exchange rates between the machine economy and the human economy. At some point, this surely ends up in different currencies decoupling from one another or perhaps the creation of entirely new currencies. We'll also likely see productivity multipliers on everything. Everything that AI touches will get massive productivity multipliers and this will loop back into the economy and will massively empower many people and it may also displace people.
This rate of science advances will basically move forward in relation to anything that an AI can touch and run an experimental loop on, which is today mostly software. But with the advance of robotics, with the advance of AI systems for biology, we should expect it to reach into other domains of science and other domains of science will start to innovate at a rate that software-based science has, which is much, much faster. We will see the loops of science speed up. When we zoom out to the world, if you look at the number of actions being taken by agents like transacting in the economy or acting with independence, which is people today, we should surely expect that the majority of these actions will eventually stem from machine agents rather than human agents. And we will need to figure out what that world looks like and how we build the kind of verification and validation infrastructure on top.
We should expect that synthetic intelligences will start to influence people far more than social media did. I'm not made comfortable fully by my own use of AI systems in things that are deeply personal to me. I am aware that AI systems are changing some of my behavior. Now seems positive to me, but in the aggregate I think in the same way that social media changed people's behaviors on a large scale through large scale interactions, we should expect the same from AI systems. They'll probably change people in more profound ways than social media has. How do we study this and how do we realize what it will do?
Finally, the thing that I think is hardest to wrap our head around, especially in the developed world, is that the developed world is largely run by market forces and governments and societies have struggled in recent years to have sufficient wealth to make choices about social programs that they might do, choices about scientific endeavors they might embark on because money's tight and you yield to market forces. But if you've moved huge chunks of the economy into computation and computation is limited and that we don't have enough computers to saturate all of the demand, then governments will be forced into this really important question of do we solely want the use of AI based on computation to be directed by market forces or do we want some of that to be directed by the forces of democracy or by the forces of society for scientific or social projects as well. We will get to make choices here which in many ways are amazing but I think that we are not ready to accept that that future might arrive quickly because it requires you to assume outrageous things happen with the economy.
In short and before we go to Q&A, we will have super intelligences show up in the world sooner than we think that grant the power of synthetic workforces and nation state security skills to individuals. We'll also have individuals like me and many others in this room who are able to take work that previously took them weeks and now do it in minutes. And we will have organizations like AI labs like Anthropic where the way work happens within the organization start to change every three or four months in direct relation to the advance of these AI capabilities. And organizations like mine which are approaching 4,000 employees already feel like they have 40,000 employees based on what they do and the amount of activity. What will it be like when it's like they have 400,000 employees or millions of employees and yet that is the future that seems like to some extent locked in? So the best and most conservative take I can generate is vast waves of the economy and society will go through profound changes in the coming years and if recursive self-improvement happens then anything I might predict would sound truly crazy. The rapid emergence of the machine economy which decouples from the human economy. The sudden maturation of robots as they gain brains that complement their existing quite good bodies. Science advances happening based on technologies not developed by people but by machines. Scientific equipment which people haven't conceived of but which works. Eventually the migration of large swaths of computation to space-based data centers which I think will happen sooner than people expect. And a world where many things that used to take 10 years now take a year or a month. It will be an age of confusing miracles happening faster than anyone might expect. And in many ways, I think this is an amazing future, but it is also a future composed of vast amounts of change and change which I think we actually need to make choices about sooner than we might expect. If we stand by as these synthetic intelligences multiply, then we'll eventually be forced into reactivity, just as societies across the world were forced into reactivity by not taking the COVID exponential seriously enough and continually discounting it. But if we accept the premise that these systems are going to get much, much better and ask ourselves what to do with them and how to use them, we unlock ourselves, I think, the mindset for exploration. There's a new world to be built for us as individuals and a new world of how we relate to one another. But I think it only comes into being if we actually choose to build it together, which requires us to recognize that this future is something that is happening and is something that we can explore and to not retreat from the present which is already full of this great amount of change. So with that, I look forward to questions and I hope this was helpful. Thank you very much.
