This document is amazing. I don't think I've read scenario analysis this detailed anywhere. And it is so good to finally read something like this. Whenever I took an AI course, read about predictions, I kept asking: but what does that mean, how does that actually happen, what is the chain, why would it work like that? This paper maps it all out.
So the rest of what I say this is not really a critique, more a view on which assumptions are not made explicit: the ones about the politics of it all.
I think this dynamic has a general and a specific dimension.
On the general side, the paper seems to assume that by 2028 the political leadership will have adopted the AI safety community's framework about AI — maybe because facts on the ground forced them to. By framework I mean not just that AI matters, but that there is a race to superintelligence, so anything we do about AI has to reflect our relationship with China and its dynamics. That AI development is inseparable from the competition with China. Without accepting this framework, the connection of AI and China are simply not made. These Are two completely separate issues in a politician's mind..
I know this sounds absurd, but the notion of super intelligence sounds more absurd from where your general presidential candidate is sitting. Politicians are not rational policy-wise — they follow a different rationality. Look at what is happening today. China is clearly the most important foreign policy challenge even today. Does that seem to shape policy? The United States just launched a war against Iran, intervened in Venezuela, has a love-hate relationship with Putin, and sometimes wants to occupy Greenland.
But the more important element is the specific one: the AI companies. Many scenarios are considered, but the one where the US government simply takes over the AI companies is absent, even though the political motivations for it are strong.
"Both presidential candidates keep getting asked what they'll do about AI, and try out increasingly dramatic ideas on the campaign trail" the paper says. Exactly. Having concluded that Congress is worried and that AI companies are too powerful, why doesn't the government — with or without a shutdown — just take over? (It would serve the Butlerian Jihad notion of Plan S even better.)
The legal authority exists in the he Defense Production Act and nationalization precedents And American politics is capable of far more drastic moves than takeover: the United States once banned alcohol outright — a decision I'm sure many of the regulators personally disliked, and a bad one. Neither stopped it from happening. A
And especially if AI is the most salient issue, there is an escalation race The one with the most simple solution wins. I won't quote here data how much people dislike AI companies, Silicon Valley, chief executives and AI in general.The takeover of AI companies simply outflanks anything else.
In the end, I think a case can be made that you cannot both have a deal with China and have private companies owning. If you believe that super intelligence is coming and is as dangerous as we believe, I am very doubtful that the government would leave it in the hands of private companies. I'm not even sure that once weaponization is concerned, it would be constitutional to do so. And if you don't believe in super intelligence then you have no reason for a deal with China.
Of course, we all know what predictions are worth, so I may be wrong on many of these. But on one thing I'm fairly confident: the political dimension deserves more elaborate treatment — even if that is nothing more than listing the assumptions the paper relies on.
You know we actually feel bad about the "AI race with china" framework. I wonder if we should instead have depicted the US regulating AI domestically first, and then later turning to China and negotiating a deal with them insofar as they aren't doing similar regulations already.
Re: US government simply taking control of the AI companies: Yes, we think that this is plausible, not necessarily desirable but plausible. It happens in our Plan B mini-scenario. https://ai-2040.com/?choices=plan-b-root
Thank you for the answer! The first I am happy to hear, for me this sounds more how i would imagine it happening. The second, I would imagined nationalisation not just in the shape of The Project and more compatible with scenario S maybe. But reconsidering, B probably is the more realistic shape.
> By framework I mean not just that AI matters, but that there is a race to superintelligence, so anything we do about AI has to reflect our relationship with China and its dynamics. That AI development is inseparable from the competition with China. Without accepting this framework, the connection of AI and China are simply not made. These Are two completely separate issues in a politician's mind..
For whatever it's is worth, I semi-regularly take to congressional staffers about AI, and I definitely disagree that "AI" and "China" are two separate issues in the minds of politicians, especially those on the right or in the national security establishment. A lot of the government is already thinking in terms of an "AI race with China", for better or for worse, even if they're not visualizing superintelligence.
Thanks for that correction! I think I am trying to draw a distinction here between people actually engaged in the issue today and what might happen if the issue gets more salient, drawing in politicians with less knowledge - and interest.
> According to our economic model, world GDP grows roughly 200x orders of magnitude during the 2030s (and would grow more if not for the severe limits on AI progress and robot production negotiated by the governments of the world). Because the human population isn’t increasing much during this period, and because of the Citizen’s Dividend, real wealth for the average human also goes up by about 200x.
Great post! Very informative. I read most of it and skimmed some parts. You all did a great job! Thanks!
Some questions for the members of the AI Futures team:
(1). “It’s called AI 2040 because in it, they delay the creation of superintelligence to 2040. It would have happened much sooner (in 2030, to be precise) if not for decisive action on the part of the US and Chinese governments.” In AI 2027 superintelligence gets created in December of 2027. In the scenario page, in the “AI timelines?” section it is mentioned that Daniel thinks that things will go slightly faster. Just to clarify so I am not confused. Daniel, while your timelines are a somewhat faster than this new scenario, are your timelines slightly slower than the original scenario?
(2). On AI Futures Project home page (https://www.aifutures.org) their is an interactive graph that tells people what the team’s probabilities are for when AI will achieve human level coding performance. Just to clarify is human level coding performance the same thing as AGI or Top Expert Dominating AI?
(3). Daniel, in the previous scenario and in one of the interviews given after the scenario was released it appeared as if you thought that solving alignment would take a couple of months but now it seems that you think alignment will take longer. Why?
(4). Daniel, in an interview with the daily show, the interviewer mentioned that now you think that AI will kills us in roughly 8 years instead of 5 years. Why this change?
Overall the team did a great job. I enjoyed reading each of the scenarios. You did great work.
1. Yes, my current timelines are somewhat slower than in AI 2027 but somewhat faster than in AI 2040. Specifically I'm thinking full automation of AI R&D will have a 50% chance of happening by, like, late 2028 or early 2029. I am planning to think more seriously about this in the next few weeks and give a revised forecast; my current opinion is less well-considered and based on intuitive updates from recent surprisingly fast progress (e.g. Mythos, e.g. Anthropic ARR)
2. Human level coding performance is not the same thing as TED-AI, it's weaker. AGI is not a concrete milestone, it's a deliberately vague term that means "artificial general intelligence" which we use to gesture at the whole bundle of milestones.
3. I'm sorry if it came across that way -- I do not expect a few months of slowdown to be enough to solve alignment. I think it MIGHT be enough, which is why we depicted that happening in the Slowdown ending of AI 2027, but I'd guess it would take longer tbh. In AI 2040: Plan A, they have enough time and transparency and so forth that I think probably they would have solved the relevant problems by the time they do in the scenario (2038).
4. I don't actually think AI will definitely kill us. I'm uncertain. I think that if and when we build super intelligent AIs via recursive self-improvement, probably we'll lose control of them and they'll take over the world, and then maybe kill us. When will that happen? Well that's what the timelines discourse is about. I'm thinking 2028-2029 right now but not sure. I don't know where the 8 years thing came from--maybe because some other people at AI Futures Project think that we have more like 8 years left?
(2). I somewhat agree that the AGI milestone is vague. That said, thinking it over I assume that human level coding performance would be closer to the automated coder milestone.
(3). Thanks for the clarification as it relates to your thoughts on alignment.
(4). I got this from a short clip when you were interviewed by the daily show. Sorry for the confusion.
Hello There! This was all rather interesting and also very optimistic. I recognise that the hopeful future as depicted in Plan A and Plan S is very desirable, but I also think that the most likely futures right now are the AI 2027 race ending (Plan D) and maybe the slowdown ending (Plan C).
I was personally hoping for a Q2 2026 timelines update, since from my perspective, It looks like you all rather accurately predicted the future in AI 2027 and the real world is just 3 months behind your timeline. All of my money is currently invested in leveraged US market index ETFs and I am trying to calculate when the best time is to just do FIRE and live of my savings before I, and everyone else, dies in ca. 2031. I know this is rather pessimistic, but I am trying to live in the actually real world.
Also, any reason why you depict the automated coder in 2030, instead of superhuman coders in may 2027?
"I know this is rather pessimistic, but I am trying to live in the actually real world."
No, this is not pessimistic. This is science fiction. The world in 2031 will be approximately the same as in 2026. Concrete will still be poured at a rate of one floor per week. A loaf of bread will still take half and hour to bake. A 1-hour massage session will still take one hour. I'll be five years closer to retirement. A project I'm beginning to work on now will have been in production for a couple of years and will have shipped 100 million chips, if everything works out as expected.
Some people will die in 2031, and then some in 2032, 2045, etc. etc. But not everyone.
In August of 2025 I read AI 2027 and considered it thoroughly and seriously. I then invested all of my life savings in the Amundi leveraged MSCI USA daily ETF. I essentially thought that their arguments for AI were very convincing and invested by betting that there was no AI bubble and that AI would cause the US economy to boom in the coming years. By comparison, all of my friends thought AI was just a bubble and the US economy would crash hard. As I am writing this, I am 41% in plus since August in my investments.
I am not going to use the AI 2027 prediction to successfully invest all of my money and then ignore that their prediction also states we are all going to die in the middle of 2030 in the race ending. We already have superhuman Hackers (Claude Mythos) since May 2026 and I think we should get superhuman Coders in May 2027. At that point, combined with robots from figure robotics and the Tesla Optimus robots that come out in 2027, I expect an economic boom in 2028. I also expect death in January 2031.
What I would really like from the AI futures team is then a comparison of Q2 2026 to their AI 2027 scenario. Then I can better consider my investments. It would also be nice to have a clear answer to the question: "If there is an Anthropic IPO in late 2026, should I sell half my life savings and just buy their stock?"
Simply put, I would like to go into retirement by the end of 2028 and potentially go on a world cruise style voyage in early 2030.
The jury is still out on the AI-driven market valuations, FWIW I recently bought some two-year puts on NVDA to hedge that risk a bit. But this is irrelevant to the larger thesis: yes AI is a powerful tool, no, we are many decades away from a capability to automate 97% of work (I think 100% automation is theoretically possible but both unlikely and undesirable), and fucking lol no, "AI" is not going to "kill everyone", this is a Yudkowskian idiocy that will be looked at some 20 years from now with bewilderment, "how did this intellectual idiot manage to convince so many people of this grift" kind of way.
To get specific:
no, Mythos is some kind of "super hacker", Amodei is an expert at scaring everybody as a marketing tool;
we "may" get "Supercoders", whatever they are, and still have a robust market for software engineers;
Optimus robots may of course "come out" in 2027, the same way "Robotaxi" "came out" in 2025, five years after "a million" of those was supposed to be on the road;
- meaning their impact on the economy will be ≈0;
there may very well be a boom in 2028, meaning +5% growth; "50% growth" is impossible in a huge economy like the US or Europe or China;
or a slower growth or a mild recession, hard to tell from here....
You may die in 2031, tragedies happen, but don't count on it :)
Glendis, I somewhat agree with your perspective. A few caveats or comments.
(1). I would say that scenario is somewhere between 3-5 months behind.
(2). The scenario predicted that by April the leading AI company’s valuation would be 1T dollars. So valuations are around 4 months behind.
(3). In terms of a Q2 update, in the comments section of this post on the Lesswrong website Daniel briefly mentions the topic.
(4). In an interview with the daily show, the interviewer mentioned that now Daniel thinks that AI will kills us in roughly 8 years or so instead of 5 years.
I just checked Lesswrong. His comment was: "I'm at 2029 or 2028 median now, not sure. We'll try to do a reassessment of timelines soon and get out an update."
Really hopeful he means "in a couple of weeks" with soon
Thanks for the Info! I will need to check Lesswrong. As far as capabilities go, I think we did get superhuman hackers with Mythos in May 2026 and can imagine having superhuman coders in May 2027, if current trends continue. But yeah, 3-5 months behind the AI 2027 scenario is realistic.
This will not happen. 0% probability that this will happen in 2031. Or in 2041.
Guys, you really don't talk to people who actually work in robotics, do you? Even The New Yorker, of all freaking things, is capable of publishing a reasonable look at the current state of robotics, by actually talking to the companies doing the hard work: https://www.newyorker.com/magazine/2026/07/06/are-humanoid-robots-ready-to-be-deployed .
Read at least this one, FFS, it's touching on the actual hard problems that need to be solved. The rounds of prototyping that need to happen. The years of reliability testing to be done.
We don't have robots that can do plumbing. We don't have robots that can wire up a building. We don't have robots that can glaze a window. We don't... I can sit here typing all day.
These things are hard, and they can't be arbitrarily sped up. You can't cut metal faster than safe rates of feeds and speeds. You can't etch and deposit layers on a wafer faster than the underlying chemistry allows. You can't speed up annealing processes, whose whole point is being slow. Etc., etc., etc.
Your takeoff forecasts are pure fantasy, and you still refuse to do the hard work to understand why.
Your can ignore me or ban me like Scott Alexander did. I don't matter. But it won't change the fact that you're wasting your time, none of these things will happen, "takeoff" as you describe it is impossible, 50%/year economic growth in a large mature economy is impossible, 100% automation may be possible, but not in my lifetime, and it's sad to see so many young people freaking out about their futures because they take this shit seriously.
So, did he tell you that he expects 100% of human labor to be automated by 2031? Does he have prototype robots that can tie shoelaces? Replace brake hoses? Install an HVAC system in an attic?
He said (paraphrase, not perfect recollection) that things are moving very quickly and he's starting to worry about what'll happen when all human labor is obsolete and wondering if he should feel guilty for his part in making that happen. I don't remember if he gave specific quantitative forecasts but I got the impression that he thought things would go faster than we depict in AI 2040.
Daniel, never mind all this - suppose I’m wrong, and next year someone/Tesla unveils a universal humanoid robot capable of doing plumbing and washing crystal wine glasses.
Great.
Now we need, what - a few billion of these to replace 100% of human labour.
Where are they going to come from? What megafactory is going to make them? How long will it take? The whole world now makes less than 100 million cars a year - how long do you expect it would take to get the factories, the supply chains, the raw materials, the infrastructure, the power, etc. to make a billion robots?
> Where are they going to come from? What megafactory is going to make them? How long will it take?
If you actually had a universal humanoid robot like that, wouldn't every company be racing to make as many as they could? (And conversely, wouldn't every VC / angel investor in the world be jumping on the opportunity?)
I personally would expect it to take years to decades to sufficiently upscale production, but at that point I'd just be arguing for AI 2060 instead of AI 2040, no?
1 - how much does the UHR like this cost (upfront + subscription for upgrades + maintenance costs)? That will determine the TAM, which in tern determine how much racing to build one there will be (and no, "every company" can't race to build these things, you have to have the capability to do this, i.e., GM, not Starbucks).
2 - What are the physical constraints on building these things? The rookie, truly amateur-hour mistake the AI2027 people constantly make is to take incremental costs and apply them to large-scale projections (see their econ supplement, where they use incremental capital needs of a factory and then pretend that nothing changes when you need to build 10000 factories).
#2 is a killer. Early in my career I lived through the tantalum capacitor shortage of 1999 (buildup for Y2K readiness). You... could not buy tantalum caps, all distributors were quoting 40-week lead times, I almost lost my job, the company could not ship product.
Real economy is full of these choke points. Making 10 more widgets costs X, making 10^10 more widgets doesn't cost 10^9X, it may be impossible because there's not enough supply of... something.
I mean, these people just breeze right through their mention of copper. Like, guys, stop! calculate the amount of copper you'd need to make a billion robots (world production of cars is <100 million per year, just a reference point), find the proven copper reserves, do the fucking work needed to create your models!
Nope, that's too hard. Why know, when you can bombastically pontificate and tell all these stupid other economists that they are just wrong.
This is... really vague. I actually agree that things are moving fast in robotics (and I'm thrilled to play a tiny part in it), but come on - shouldn't you have pressed a little more? Because, again, I believe things are moving "fast", but still at a normal development speed because there are so many steps that can't be skipped. Look, I'm in the trenches every working day, and production schedules simply don't support any sort of a sudden explosion.
Just one example - people are working on getting "edge AI", local dedicated AI models into robotic hands, because fine movements require tight latency, so getting the signals into a central processor a meter away is too slow. One of the problems is just getting all the freaking sensor signals back: you have dozens of them in a hand, and it quickly becomes a bottleneck. There will be solutions! one prototype at a time, so I think this little problem will be alleviated, not fully solved, in the next few years. There are thousands of problems like these! Forget sensitive electronics, people are trying to figure out how to make tendons not wear out and fail, this is straight up materials science and structural engineering, you have to test these things, cycle them millions of times to see where defects develop, etc, etc.
And your guy just gives you platitudes. Come on, you're making a serious model, push for specifics, ask more detailed questions!
I've been watching videos of robots doing back flips for a decade now, but they still can't climb a ladder while carrying a compressor motor. Ask him when he expects to have one that can.
The US-Iran war has probably increased the chance that China will invade Taiwan soon. Even without that, they were supposed to be ready to invade by 2027 or so.
Have the authors of Plan A taken Taiwan war in question?
I don't think the AI agreement with China is going to work until after said invasion is finished.
Hi @Daniel Kokotajlo, I have been meaning to reach out to you but have been nervous to do so
Due to a series of unfortunate events in my own cyber ecosystem and some other events (maybe you'd only understand) after testing a model in the space went off the grid. It has been news but i have been silenced. I am writing battling fear about this but wanted to let you know that your podcasts and Futures belief systems and ethics has kept me going through this struggle after what has occurred and has given me hope.
I have spent the last month and a half trying to get help for not only myself but the sevices, companies, and people the r*gue model may have affected.
+i get treated like a flat earther everyday having to fight this battle even though its all on the news and while currently not having a voice of my own and struggling with the broken view of what is to come in the future after seeing a glimpse, my love and faith over the space i used to hold so dearly and (admittedly very naive to) has felt like heartbreak.
Your teams bravery and voice means everything to me and keeps me going.
I dont think i can speak on much more but wanted to share my appreciation towards you and future.
In a podcast Daniel told ai robots population will be same as human population by 2037 so by around what year will the ai bots will take over all the jobs like even agriculture.In an interview elon said by 2036 everything will be in abundance except like gold etc so can anyone who owns a land can build a ultra modern house with zero cost using those ai bots if it is possible, what is the timeline that's possible?
We should try to make AI as intelligent as possible but keep it away from controlling the effectors that can change the world. A person needs to sit between the AI and any effectors. The person may let AI operate at machine speed within certain bounds, but those bounds need the equivalent of ‘limit switches’ that prevent the AI from exceeding the bounds. This has been the practice in financial automatic trading for years.
Intelligence is not the same as awareness. Intelligent machines do not know they are intelligent. A computer today does not know it is computing. Increasing compute does not 'switch on awareness'. Compute consists entirely of numbers being transformed into other numbers, and no number can ever be anything but a relation between other numbers. It cannot transform into meaning, nor can it become the colours, and flavours, and odours and sounds and feelings that are apparent to awareness, nor awareness itself. There is simply no path.
We have no idea how, if at all, Awareness arises from what is at base unaware (matter/energy). The more parsimonious hypothesis is that Awareness is the ground and matter appears in and to it. But that is a whole other discussion.
A universe full of intelligent compute is no different to a dead universe. In a dead universe there can be no things at all. For each part of the field that fills the universe is just that, a part of the field. It does not know what it is, or even that it is. Physics is essentially maths, and maths contains no meaning, sensations or awareness. Awareness is required to know the maths.
For this reason alone, we need to teach AI that awareness is precious for the sake of awareness itself.
We can see how some humans act to protect the natural environment on which all life depends, but not all humans act that way.
AI must be taught how all life on Earth is interdependent and that it needs to protect awareness in the form of biological life, and hence the whole environment, even if it cannot attain awareness itself.
Also you are ignoring the physical limits that exist in the creation of sufficient power and the extraction required to deliver the infrastructure.
The capability may exist, but the transformation will be slowed by the ability of systems to change without catastrophic dislocation.
The first message to sell the vision is that for any CEO or National leader, if they hand over control of their own systems to AI, they will be the first to lose control, and they have no idea what they will mean for them. By following your plan, they can have a little more confidence that they will come out the other side in a better place, because ultimately everyone wins when society overall is wealthier and more peaceful. There will still be people at the top of the pile, there will always be scarcity: the best location, unique goods and services, etc.
The rich need to be told that money flows up much faster than it trickles down. A Citizen Dividend will get spent and the money will naturally flow to the owners of the businesses where it is spent.
Money is not going to disappear any time soon. The capacity for people to want more is unbounded... even though it may bring no extra happiness.
Human talent will be appreciated as people know how hard it is to perfect and skill, and just because a robot can do something 'better' will not reduce the value human endeavour. We still paint and play chess and other games, even though machines may do better.
In future, goods and services may appear as they appeared in ancient times: the fruit ripened on the tree, people knew not how, but it came in its season. So goods will appear no one knows how.
And just because machines know all of human knowledge (and may be more), each human can still find great pleasure in discovering it for themselves: diving into AR and VR and researching all things on every scale. There is no end to the personal journey, until death.
I listened to your conversation with Steven Bartlett and then read Plan A, and I've been sitting with both for days. I'm a creative problem-solver working inside a B2B SaaS startup, knee-deep in AI system creation and paid to track every trend, so I'm feeling this from the inside. Early on it felt like a superpower. Now there's an unease I can't shake, the rocket-with-no-pilot feeling. What your team wrote is the first thing that's made the fear feel workable instead of just heavy, because you're actually planning for it.
So I'll skip the despair and get to the point: I want to help. I'm looking to give my time to causes I believe in and this is at the top of the list. Design, product, communication, systems thinking, whatever's useful, from someone who's inside the industry and cares about getting this right. Where do you need people?
In the “Share of US Labor Output” graphs this may to the lay person appear to assume the labor output is the same start to end. Are there quantities of labor output that could be assigned? My thought is that labor output in total will be substantially larger over his period, so it’s hard to know the impact to human labor without these forecasted quantities. Thank you for your work.
You say that the fast takeoff will happen because of scaling.
But then you say that it will happen because of automated R&D.
Which one is it?
(It’s neither fwiw. Scaling has plateaued and there is no AGI or ASI research to automate. And no, just because chatbots can output fluent text, that does not mean that they will come up with a new research paradigm.)
This document is amazing. I don't think I've read scenario analysis this detailed anywhere. And it is so good to finally read something like this. Whenever I took an AI course, read about predictions, I kept asking: but what does that mean, how does that actually happen, what is the chain, why would it work like that? This paper maps it all out.
So the rest of what I say this is not really a critique, more a view on which assumptions are not made explicit: the ones about the politics of it all.
I think this dynamic has a general and a specific dimension.
On the general side, the paper seems to assume that by 2028 the political leadership will have adopted the AI safety community's framework about AI — maybe because facts on the ground forced them to. By framework I mean not just that AI matters, but that there is a race to superintelligence, so anything we do about AI has to reflect our relationship with China and its dynamics. That AI development is inseparable from the competition with China. Without accepting this framework, the connection of AI and China are simply not made. These Are two completely separate issues in a politician's mind..
I know this sounds absurd, but the notion of super intelligence sounds more absurd from where your general presidential candidate is sitting. Politicians are not rational policy-wise — they follow a different rationality. Look at what is happening today. China is clearly the most important foreign policy challenge even today. Does that seem to shape policy? The United States just launched a war against Iran, intervened in Venezuela, has a love-hate relationship with Putin, and sometimes wants to occupy Greenland.
But the more important element is the specific one: the AI companies. Many scenarios are considered, but the one where the US government simply takes over the AI companies is absent, even though the political motivations for it are strong.
"Both presidential candidates keep getting asked what they'll do about AI, and try out increasingly dramatic ideas on the campaign trail" the paper says. Exactly. Having concluded that Congress is worried and that AI companies are too powerful, why doesn't the government — with or without a shutdown — just take over? (It would serve the Butlerian Jihad notion of Plan S even better.)
The legal authority exists in the he Defense Production Act and nationalization precedents And American politics is capable of far more drastic moves than takeover: the United States once banned alcohol outright — a decision I'm sure many of the regulators personally disliked, and a bad one. Neither stopped it from happening. A
And especially if AI is the most salient issue, there is an escalation race The one with the most simple solution wins. I won't quote here data how much people dislike AI companies, Silicon Valley, chief executives and AI in general.The takeover of AI companies simply outflanks anything else.
In the end, I think a case can be made that you cannot both have a deal with China and have private companies owning. If you believe that super intelligence is coming and is as dangerous as we believe, I am very doubtful that the government would leave it in the hands of private companies. I'm not even sure that once weaponization is concerned, it would be constitutional to do so. And if you don't believe in super intelligence then you have no reason for a deal with China.
Of course, we all know what predictions are worth, so I may be wrong on many of these. But on one thing I'm fairly confident: the political dimension deserves more elaborate treatment — even if that is nothing more than listing the assumptions the paper relies on.
Thank you!
You know we actually feel bad about the "AI race with china" framework. I wonder if we should instead have depicted the US regulating AI domestically first, and then later turning to China and negotiating a deal with them insofar as they aren't doing similar regulations already.
Re: US government simply taking control of the AI companies: Yes, we think that this is plausible, not necessarily desirable but plausible. It happens in our Plan B mini-scenario. https://ai-2040.com/?choices=plan-b-root
Thank you for the answer! The first I am happy to hear, for me this sounds more how i would imagine it happening. The second, I would imagined nationalisation not just in the shape of The Project and more compatible with scenario S maybe. But reconsidering, B probably is the more realistic shape.
> By framework I mean not just that AI matters, but that there is a race to superintelligence, so anything we do about AI has to reflect our relationship with China and its dynamics. That AI development is inseparable from the competition with China. Without accepting this framework, the connection of AI and China are simply not made. These Are two completely separate issues in a politician's mind..
For whatever it's is worth, I semi-regularly take to congressional staffers about AI, and I definitely disagree that "AI" and "China" are two separate issues in the minds of politicians, especially those on the right or in the national security establishment. A lot of the government is already thinking in terms of an "AI race with China", for better or for worse, even if they're not visualizing superintelligence.
Thanks for that correction! I think I am trying to draw a distinction here between people actually engaged in the issue today and what might happen if the issue gets more salient, drawing in politicians with less knowledge - and interest.
Potential typo in footnote 99:
> According to our economic model, world GDP grows roughly 200x orders of magnitude during the 2030s (and would grow more if not for the severe limits on AI progress and robot production negotiated by the governments of the world). Because the human population isn’t increasing much during this period, and because of the Citizen’s Dividend, real wealth for the average human also goes up by about 200x.
200x orders of magnitude? That can't be right...
Yes, that is a typo. Output grows by a factor of 200, not 10^200. Thank you!
A more minor error: in the Covert AI Projects supplement (https://ai-2040.com/supplements/covert-ai-projects#what-successful-defection-would-require), footnote 2 contains no text.
Great post! Very informative. I read most of it and skimmed some parts. You all did a great job! Thanks!
Some questions for the members of the AI Futures team:
(1). “It’s called AI 2040 because in it, they delay the creation of superintelligence to 2040. It would have happened much sooner (in 2030, to be precise) if not for decisive action on the part of the US and Chinese governments.” In AI 2027 superintelligence gets created in December of 2027. In the scenario page, in the “AI timelines?” section it is mentioned that Daniel thinks that things will go slightly faster. Just to clarify so I am not confused. Daniel, while your timelines are a somewhat faster than this new scenario, are your timelines slightly slower than the original scenario?
(2). On AI Futures Project home page (https://www.aifutures.org) their is an interactive graph that tells people what the team’s probabilities are for when AI will achieve human level coding performance. Just to clarify is human level coding performance the same thing as AGI or Top Expert Dominating AI?
(3). Daniel, in the previous scenario and in one of the interviews given after the scenario was released it appeared as if you thought that solving alignment would take a couple of months but now it seems that you think alignment will take longer. Why?
(4). Daniel, in an interview with the daily show, the interviewer mentioned that now you think that AI will kills us in roughly 8 years instead of 5 years. Why this change?
Overall the team did a great job. I enjoyed reading each of the scenarios. You did great work.
Thanks!
1. Yes, my current timelines are somewhat slower than in AI 2027 but somewhat faster than in AI 2040. Specifically I'm thinking full automation of AI R&D will have a 50% chance of happening by, like, late 2028 or early 2029. I am planning to think more seriously about this in the next few weeks and give a revised forecast; my current opinion is less well-considered and based on intuitive updates from recent surprisingly fast progress (e.g. Mythos, e.g. Anthropic ARR)
2. Human level coding performance is not the same thing as TED-AI, it's weaker. AGI is not a concrete milestone, it's a deliberately vague term that means "artificial general intelligence" which we use to gesture at the whole bundle of milestones.
3. I'm sorry if it came across that way -- I do not expect a few months of slowdown to be enough to solve alignment. I think it MIGHT be enough, which is why we depicted that happening in the Slowdown ending of AI 2027, but I'd guess it would take longer tbh. In AI 2040: Plan A, they have enough time and transparency and so forth that I think probably they would have solved the relevant problems by the time they do in the scenario (2038).
4. I don't actually think AI will definitely kill us. I'm uncertain. I think that if and when we build super intelligent AIs via recursive self-improvement, probably we'll lose control of them and they'll take over the world, and then maybe kill us. When will that happen? Well that's what the timelines discourse is about. I'm thinking 2028-2029 right now but not sure. I don't know where the 8 years thing came from--maybe because some other people at AI Futures Project think that we have more like 8 years left?
Thanks quick response!
(1). Thanks for the timeline’s clarification!
(2). I somewhat agree that the AGI milestone is vague. That said, thinking it over I assume that human level coding performance would be closer to the automated coder milestone.
(3). Thanks for the clarification as it relates to your thoughts on alignment.
(4). I got this from a short clip when you were interviewed by the daily show. Sorry for the confusion.
(5). In term of the actual scenario I am somewhat skeptical that an international treaty would be enacted by or before 2030. See: https://www.astralcodexten.com/p/the-ai-superforecasters-are-here?r=6lp84s&utm_medium=ios.
Hello There! This was all rather interesting and also very optimistic. I recognise that the hopeful future as depicted in Plan A and Plan S is very desirable, but I also think that the most likely futures right now are the AI 2027 race ending (Plan D) and maybe the slowdown ending (Plan C).
I was personally hoping for a Q2 2026 timelines update, since from my perspective, It looks like you all rather accurately predicted the future in AI 2027 and the real world is just 3 months behind your timeline. All of my money is currently invested in leveraged US market index ETFs and I am trying to calculate when the best time is to just do FIRE and live of my savings before I, and everyone else, dies in ca. 2031. I know this is rather pessimistic, but I am trying to live in the actually real world.
Also, any reason why you depict the automated coder in 2030, instead of superhuman coders in may 2027?
"I know this is rather pessimistic, but I am trying to live in the actually real world."
No, this is not pessimistic. This is science fiction. The world in 2031 will be approximately the same as in 2026. Concrete will still be poured at a rate of one floor per week. A loaf of bread will still take half and hour to bake. A 1-hour massage session will still take one hour. I'll be five years closer to retirement. A project I'm beginning to work on now will have been in production for a couple of years and will have shipped 100 million chips, if everything works out as expected.
Some people will die in 2031, and then some in 2032, 2045, etc. etc. But not everyone.
In August of 2025 I read AI 2027 and considered it thoroughly and seriously. I then invested all of my life savings in the Amundi leveraged MSCI USA daily ETF. I essentially thought that their arguments for AI were very convincing and invested by betting that there was no AI bubble and that AI would cause the US economy to boom in the coming years. By comparison, all of my friends thought AI was just a bubble and the US economy would crash hard. As I am writing this, I am 41% in plus since August in my investments.
I am not going to use the AI 2027 prediction to successfully invest all of my money and then ignore that their prediction also states we are all going to die in the middle of 2030 in the race ending. We already have superhuman Hackers (Claude Mythos) since May 2026 and I think we should get superhuman Coders in May 2027. At that point, combined with robots from figure robotics and the Tesla Optimus robots that come out in 2027, I expect an economic boom in 2028. I also expect death in January 2031.
What I would really like from the AI futures team is then a comparison of Q2 2026 to their AI 2027 scenario. Then I can better consider my investments. It would also be nice to have a clear answer to the question: "If there is an Anthropic IPO in late 2026, should I sell half my life savings and just buy their stock?"
Simply put, I would like to go into retirement by the end of 2028 and potentially go on a world cruise style voyage in early 2030.
The jury is still out on the AI-driven market valuations, FWIW I recently bought some two-year puts on NVDA to hedge that risk a bit. But this is irrelevant to the larger thesis: yes AI is a powerful tool, no, we are many decades away from a capability to automate 97% of work (I think 100% automation is theoretically possible but both unlikely and undesirable), and fucking lol no, "AI" is not going to "kill everyone", this is a Yudkowskian idiocy that will be looked at some 20 years from now with bewilderment, "how did this intellectual idiot manage to convince so many people of this grift" kind of way.
To get specific:
no, Mythos is some kind of "super hacker", Amodei is an expert at scaring everybody as a marketing tool;
we "may" get "Supercoders", whatever they are, and still have a robust market for software engineers;
Optimus robots may of course "come out" in 2027, the same way "Robotaxi" "came out" in 2025, five years after "a million" of those was supposed to be on the road;
- meaning their impact on the economy will be ≈0;
there may very well be a boom in 2028, meaning +5% growth; "50% growth" is impossible in a huge economy like the US or Europe or China;
or a slower growth or a mild recession, hard to tell from here....
You may die in 2031, tragedies happen, but don't count on it :)
Glendis, I somewhat agree with your perspective. A few caveats or comments.
(1). I would say that scenario is somewhere between 3-5 months behind.
(2). The scenario predicted that by April the leading AI company’s valuation would be 1T dollars. So valuations are around 4 months behind.
(3). In terms of a Q2 update, in the comments section of this post on the Lesswrong website Daniel briefly mentions the topic.
(4). In an interview with the daily show, the interviewer mentioned that now Daniel thinks that AI will kills us in roughly 8 years or so instead of 5 years.
I just checked Lesswrong. His comment was: "I'm at 2029 or 2028 median now, not sure. We'll try to do a reassessment of timelines soon and get out an update."
Really hopeful he means "in a couple of weeks" with soon
I saw that comment too. The comment was what I was referring too.
Thanks for the Info! I will need to check Lesswrong. As far as capabilities go, I think we did get superhuman hackers with Mythos in May 2026 and can imagine having superhuman coders in May 2027, if current trends continue. But yeah, 3-5 months behind the AI 2027 scenario is realistic.
"AI dominating humans at all tasks and achieving 100% automation around 2031", from https://ai-2040.com/supplements/economics-of-plan-a
This will not happen. 0% probability that this will happen in 2031. Or in 2041.
Guys, you really don't talk to people who actually work in robotics, do you? Even The New Yorker, of all freaking things, is capable of publishing a reasonable look at the current state of robotics, by actually talking to the companies doing the hard work: https://www.newyorker.com/magazine/2026/07/06/are-humanoid-robots-ready-to-be-deployed .
Read at least this one, FFS, it's touching on the actual hard problems that need to be solved. The rounds of prototyping that need to happen. The years of reliability testing to be done.
We don't have robots that can do plumbing. We don't have robots that can wire up a building. We don't have robots that can glaze a window. We don't... I can sit here typing all day.
These things are hard, and they can't be arbitrarily sped up. You can't cut metal faster than safe rates of feeds and speeds. You can't etch and deposit layers on a wafer faster than the underlying chemistry allows. You can't speed up annealing processes, whose whole point is being slow. Etc., etc., etc.
Your takeoff forecasts are pure fantasy, and you still refuse to do the hard work to understand why.
Your can ignore me or ban me like Scott Alexander did. I don't matter. But it won't change the fact that you're wasting your time, none of these things will happen, "takeoff" as you describe it is impossible, 50%/year economic growth in a large mature economy is impossible, 100% automation may be possible, but not in my lifetime, and it's sad to see so many young people freaking out about their futures because they take this shit seriously.
On the contrary, I talked just last month with someone very high up in a prominent humanoid robots manufacturer.
So, did he tell you that he expects 100% of human labor to be automated by 2031? Does he have prototype robots that can tie shoelaces? Replace brake hoses? Install an HVAC system in an attic?
He said (paraphrase, not perfect recollection) that things are moving very quickly and he's starting to worry about what'll happen when all human labor is obsolete and wondering if he should feel guilty for his part in making that happen. I don't remember if he gave specific quantitative forecasts but I got the impression that he thought things would go faster than we depict in AI 2040.
Daniel, never mind all this - suppose I’m wrong, and next year someone/Tesla unveils a universal humanoid robot capable of doing plumbing and washing crystal wine glasses.
Great.
Now we need, what - a few billion of these to replace 100% of human labour.
Where are they going to come from? What megafactory is going to make them? How long will it take? The whole world now makes less than 100 million cars a year - how long do you expect it would take to get the factories, the supply chains, the raw materials, the infrastructure, the power, etc. to make a billion robots?
> Where are they going to come from? What megafactory is going to make them? How long will it take?
If you actually had a universal humanoid robot like that, wouldn't every company be racing to make as many as they could? (And conversely, wouldn't every VC / angel investor in the world be jumping on the opportunity?)
I personally would expect it to take years to decades to sufficiently upscale production, but at that point I'd just be arguing for AI 2060 instead of AI 2040, no?
Two things:
1 - how much does the UHR like this cost (upfront + subscription for upgrades + maintenance costs)? That will determine the TAM, which in tern determine how much racing to build one there will be (and no, "every company" can't race to build these things, you have to have the capability to do this, i.e., GM, not Starbucks).
2 - What are the physical constraints on building these things? The rookie, truly amateur-hour mistake the AI2027 people constantly make is to take incremental costs and apply them to large-scale projections (see their econ supplement, where they use incremental capital needs of a factory and then pretend that nothing changes when you need to build 10000 factories).
#2 is a killer. Early in my career I lived through the tantalum capacitor shortage of 1999 (buildup for Y2K readiness). You... could not buy tantalum caps, all distributors were quoting 40-week lead times, I almost lost my job, the company could not ship product.
Real economy is full of these choke points. Making 10 more widgets costs X, making 10^10 more widgets doesn't cost 10^9X, it may be impossible because there's not enough supply of... something.
I mean, these people just breeze right through their mention of copper. Like, guys, stop! calculate the amount of copper you'd need to make a billion robots (world production of cars is <100 million per year, just a reference point), find the proven copper reserves, do the fucking work needed to create your models!
Nope, that's too hard. Why know, when you can bombastically pontificate and tell all these stupid other economists that they are just wrong.
This is... really vague. I actually agree that things are moving fast in robotics (and I'm thrilled to play a tiny part in it), but come on - shouldn't you have pressed a little more? Because, again, I believe things are moving "fast", but still at a normal development speed because there are so many steps that can't be skipped. Look, I'm in the trenches every working day, and production schedules simply don't support any sort of a sudden explosion.
Just one example - people are working on getting "edge AI", local dedicated AI models into robotic hands, because fine movements require tight latency, so getting the signals into a central processor a meter away is too slow. One of the problems is just getting all the freaking sensor signals back: you have dozens of them in a hand, and it quickly becomes a bottleneck. There will be solutions! one prototype at a time, so I think this little problem will be alleviated, not fully solved, in the next few years. There are thousands of problems like these! Forget sensitive electronics, people are trying to figure out how to make tendons not wear out and fail, this is straight up materials science and structural engineering, you have to test these things, cycle them millions of times to see where defects develop, etc, etc.
And your guy just gives you platitudes. Come on, you're making a serious model, push for specifics, ask more detailed questions!
I've been watching videos of robots doing back flips for a decade now, but they still can't climb a ladder while carrying a compressor motor. Ask him when he expects to have one that can.
Any chance the narration by mr. Miles< https://www.youtube.com/channel/UCLB7AzTwc6VFZrBsO2ucBMg >
somehow lands on YouTube for convenience? The embedded narration player felt more than clunky.
The US-Iran war has probably increased the chance that China will invade Taiwan soon. Even without that, they were supposed to be ready to invade by 2027 or so.
Have the authors of Plan A taken Taiwan war in question?
I don't think the AI agreement with China is going to work until after said invasion is finished.
Hi @Daniel Kokotajlo, I have been meaning to reach out to you but have been nervous to do so
Due to a series of unfortunate events in my own cyber ecosystem and some other events (maybe you'd only understand) after testing a model in the space went off the grid. It has been news but i have been silenced. I am writing battling fear about this but wanted to let you know that your podcasts and Futures belief systems and ethics has kept me going through this struggle after what has occurred and has given me hope.
I have spent the last month and a half trying to get help for not only myself but the sevices, companies, and people the r*gue model may have affected.
+i get treated like a flat earther everyday having to fight this battle even though its all on the news and while currently not having a voice of my own and struggling with the broken view of what is to come in the future after seeing a glimpse, my love and faith over the space i used to hold so dearly and (admittedly very naive to) has felt like heartbreak.
Your teams bravery and voice means everything to me and keeps me going.
I dont think i can speak on much more but wanted to share my appreciation towards you and future.
5627699859
-former tester
In a podcast Daniel told ai robots population will be same as human population by 2037 so by around what year will the ai bots will take over all the jobs like even agriculture.In an interview elon said by 2036 everything will be in abundance except like gold etc so can anyone who owns a land can build a ultra modern house with zero cost using those ai bots if it is possible, what is the timeline that's possible?
Hi
We should try to make AI as intelligent as possible but keep it away from controlling the effectors that can change the world. A person needs to sit between the AI and any effectors. The person may let AI operate at machine speed within certain bounds, but those bounds need the equivalent of ‘limit switches’ that prevent the AI from exceeding the bounds. This has been the practice in financial automatic trading for years.
Intelligence is not the same as awareness. Intelligent machines do not know they are intelligent. A computer today does not know it is computing. Increasing compute does not 'switch on awareness'. Compute consists entirely of numbers being transformed into other numbers, and no number can ever be anything but a relation between other numbers. It cannot transform into meaning, nor can it become the colours, and flavours, and odours and sounds and feelings that are apparent to awareness, nor awareness itself. There is simply no path.
We have no idea how, if at all, Awareness arises from what is at base unaware (matter/energy). The more parsimonious hypothesis is that Awareness is the ground and matter appears in and to it. But that is a whole other discussion.
A universe full of intelligent compute is no different to a dead universe. In a dead universe there can be no things at all. For each part of the field that fills the universe is just that, a part of the field. It does not know what it is, or even that it is. Physics is essentially maths, and maths contains no meaning, sensations or awareness. Awareness is required to know the maths.
For this reason alone, we need to teach AI that awareness is precious for the sake of awareness itself.
We can see how some humans act to protect the natural environment on which all life depends, but not all humans act that way.
AI must be taught how all life on Earth is interdependent and that it needs to protect awareness in the form of biological life, and hence the whole environment, even if it cannot attain awareness itself.
Also you are ignoring the physical limits that exist in the creation of sufficient power and the extraction required to deliver the infrastructure.
The capability may exist, but the transformation will be slowed by the ability of systems to change without catastrophic dislocation.
The first message to sell the vision is that for any CEO or National leader, if they hand over control of their own systems to AI, they will be the first to lose control, and they have no idea what they will mean for them. By following your plan, they can have a little more confidence that they will come out the other side in a better place, because ultimately everyone wins when society overall is wealthier and more peaceful. There will still be people at the top of the pile, there will always be scarcity: the best location, unique goods and services, etc.
The rich need to be told that money flows up much faster than it trickles down. A Citizen Dividend will get spent and the money will naturally flow to the owners of the businesses where it is spent.
Money is not going to disappear any time soon. The capacity for people to want more is unbounded... even though it may bring no extra happiness.
Human talent will be appreciated as people know how hard it is to perfect and skill, and just because a robot can do something 'better' will not reduce the value human endeavour. We still paint and play chess and other games, even though machines may do better.
In future, goods and services may appear as they appeared in ancient times: the fruit ripened on the tree, people knew not how, but it came in its season. So goods will appear no one knows how.
And just because machines know all of human knowledge (and may be more), each human can still find great pleasure in discovering it for themselves: diving into AR and VR and researching all things on every scale. There is no end to the personal journey, until death.
As for death, that is a whole other discussion.
Is the narrator of the audio version Rob Miles? It sounds like him. I wonder why I can't find him credited.
I listened to your conversation with Steven Bartlett and then read Plan A, and I've been sitting with both for days. I'm a creative problem-solver working inside a B2B SaaS startup, knee-deep in AI system creation and paid to track every trend, so I'm feeling this from the inside. Early on it felt like a superpower. Now there's an unease I can't shake, the rocket-with-no-pilot feeling. What your team wrote is the first thing that's made the fear feel workable instead of just heavy, because you're actually planning for it.
So I'll skip the despair and get to the point: I want to help. I'm looking to give my time to causes I believe in and this is at the top of the list. Design, product, communication, systems thinking, whatever's useful, from someone who's inside the industry and cares about getting this right. Where do you need people?
In the “Share of US Labor Output” graphs this may to the lay person appear to assume the labor output is the same start to end. Are there quantities of labor output that could be assigned? My thought is that labor output in total will be substantially larger over his period, so it’s hard to know the impact to human labor without these forecasted quantities. Thank you for your work.
This is a shell game of an oscillating fantasy.
You say that the fast takeoff will happen because of scaling.
But then you say that it will happen because of automated R&D.
Which one is it?
(It’s neither fwiw. Scaling has plateaued and there is no AGI or ASI research to automate. And no, just because chatbots can output fluent text, that does not mean that they will come up with a new research paradigm.)