- Relatedly, I’m also interested in the simple method of extrapolating AI revenue growth trends until AI revenue is most of the world economy. That seems like a decent proxy for when AGI will be achieved. I trust this method less than our model for obvious reasons, but I still put some weight on it. What does it say? Well, it says “Early 2030s.” OK.
I'm curious why you trust revenue extrapolation less than the model. Intuitively revenue seems like a better thing to extrapolate to me than benchmarks or flops or whatever because it's much less gameable and there's a much more clear threshold for AGI (revenue is similar size to GDP).
1I think revenue extrapolations seem like a useful exercise. But I think they provide much less evidence than our model.
Which revenues would you extrapolate? You get different results for e.g. doing OpenAI vs. Nvidia.
Also (most importantly) are you saying we should assume that log(revenue) is a straight line?
- If so, that seems like a really bad assumption given that usually startup revenue growth rates slow down a lot as revenue increases, so that should be the baseline assumption.
- If not, how else do we predict how the revenue trend will change without thinking about AI capabilities? We could look at base rates for startups that have this level of revenue growth early on, but then obviously none of those revenue trends have ever grown until world GDP, so that would say AGI never.
edited to add: relevant graph from https://epoch.ai/gradient-updates/openai-is-projecting-unprecedented-revenue-growth:
much more clear threshold for AGI
Also I disagree with this, I think time horizon is about as good as revenue on this dimension, maybe a bit better. Both are hugely uncertain though of course.
1I basically agree with Eli, though I'll say that I don't think the gap between extrapolating METR specifically and AI revenue is huge. I think ideally I'd do some sort of weighted mix of both, which is sorta what I'm doing in my ATC.
Thanks! Letting us play with the assumptions is a great way to develop an intuitive sensitivity analysis.
We’ve significantly upgraded our timelines and takeoff models! It predicts when AIs will reach key capability milestones: for example, Automated Coder / AC (full automation of coding) and superintelligence / ASI (much better than the best humans at virtually all cognitive tasks). This post will briefly explain how the model works, present our timelines and takeoff forecasts, and compare it to our previous (AI 2027) models (spoiler: the AI Futures Model predicts about 3 years longer timelines to full coding automation than our previous model, mostly due to being less bullish on pre-full-automation AI R&D speedups).
If you’re interested in playing with the model yourself, the best way to do so is via this interactive website: aifuturesmodel.com
If you’d like to skip the motivation for our model to an explanation for how it works, go here, The website has a more in-depth explanation of the model (starts here; use the diagram on the right as a table of contents), as well as our forecasts.
Why do timelines and takeoff modeling?
The future is very hard to predict. We don't think this model, or any other model, should be trusted completely. The model takes into account what we think are the most important dynamics and factors, but it doesn't take into account everything. Also, only some of the parameter values in the model are grounded in empirical data; the rest are intuitive guesses. If you disagree with our guesses, you can change them above.
Nevertheless, we think that modeling work is important. Our overall view is the result of weighing many considerations, factors, arguments, etc.; a model is a way to do this transparently and explicitly, as opposed to implicitly and all in our head. By reading about our model, you can come to understand why we have the views we do, what arguments and trends seem most important to us, etc.
The future is uncertain, but we shouldn’t just wait for it to arrive. If we try to predict what will happen, if we pay attention to the trends and extrapolate them, if we build models of the underlying dynamics, then we'll have a better sense of what is likely, and we'll be less unprepared for what happens. We’ll also be able to better incorporate future empirical data into our forecasts.
In fact, the improvements we’ve made to this model as compared to our timelines model at the time we published AI 2027 (Apr 2025), have resulted in a roughly 2-4 year shift in our median for full coding automation. This has primarily come from improving our modeling of AI R&D automation. These modeling improvements have resulted in a larger change in our views than the new empirical evidence that we’ve observed. You can read more about the shift below.
Why our approach to modeling? Comparing to other approaches
AGI[1] timelines forecasting methods
Trust the experts
Unfortunately, there is nothing close to an expert consensus, and it doesn’t seem like most experts have thought much about AGI forecasting (e.g. a 2023 survey observed huge framing effects depending on whether they asked for probabilities of milestones being achieved by certain years, or instead asked for years that correspond to percentiles). That 2023 survey of AI academics got an AGI median of 2047 or 2116, depending on the definition.[2] There’s also this aggregation of Metaculus and Manifold markets which estimates 50% by 2030. As for the people building the technology, they tend to be more bullish; the most extreme among them (Anthropic and OpenAI) say things like 2027 and 2028. For a survey of older predictions and how they’ve fared, see this.
Given that experts disagree with each other and mostly seem to have not thought deeply about AGI forecasting, we think it’s important to work to form our own forecast.
Intuition informed by arguments
Can the current paradigm scale to AGI? Does it lack something important, like common sense, true original thinking, or online/continual learning (etc.)? Questions like these are very important and there are very many of them, far too many to canvas here. The way this method works is that everyone ingests the pile of arguments and considerations and makes up their own minds about which arguments are good and how they weigh against each other. This process inherently involves intuition/subjective-judgment, which is why we label it as “intuition.”
Which is not to denigrate it! We think that any AI forecaster worth their salt must engage in this kind of argumentation, and that generally speaking the more facts you know, the more arguments you’ve considered and evaluated, the more accurate your intuitions/vibes/judgments will become. Also, relatedly, your judgment about which models to use, and how much to trust them, will get better too. Our own all-things-considered views are only partially based on the modelling we’ve done; they are also informed by intuitions.
But we think that there are large benefits to incorporating quantitative models into our forecasts: it’s hard to aggregate so many considerations into an overall view without using a quantitative framework. We’ve also found that quantitative models help prioritize which arguments are most important to pay attention to. And our best guess is that overall, forecasts by quantitative trend extrapolation have a better historical track record than intuitions alone.
Revenue extrapolation
Simple idea: extrapolate AI revenue until it’s the majority of world GDP. Of course, there’s something silly about this; every previous fast-growing tech sector has eventually plateaued… That said, AI seems like it could be the exception, because in principle AI can do everything. Now that AI is a major industry, we think this method provides nonzero evidence. According to this Epoch dataset, frontier AI company revenue is something like $20B now and growing around 4.1x/yr. This simple extrapolation gets to $100T annualized revenue around the end of 2031.[3]
We give weight to revenue extrapolation in our all-things-considered views, but on the other hand revenue trends change all the time and we’d like to predict the underlying drivers of how it might change. Also, it’s unclear what revenue threshold counts as AGI. Therefore, we want to specifically extrapolate AI capabilities.
Compute extrapolation anchored by the brain
The basic idea is to estimate how much compute it would take to get AGI, anchored by the human brain. Then predict that AGI will happen when we have that much compute. This approach has gone through a few iterations:
Davidson’s Full Takeoff Model and Epoch’s GATE used the same method as bio anchors to determine the AGI training compute requirement, but they also modeled how AI R&D automation would shorten timelines. They modeled automation by splitting up AI software and hardware R&D into many tasks, then forecasting the effective compute gap between 20% task automation and 100% automation. The percentage of tasks automated, along with experiment compute and automation compute, determine the magnitude of inputs to AI R&D. These inputs are converted to progress in software efficiency using a semi-endogeneous growth model. Software efficiency is then multiplied by training compute to get effective compute.
At the time the FTM was created it predicted AGI in 2040, with the parameter settings chosen by Davidson. But both compute and algorithmic progress has been faster than they expected. When the FTM is updated to take into account this new data, it gives shorter medians in the late 2020s or early 2030s. Meanwhile, with GATE’s median parameters, it predicts AGI in 2034.
Overall, this forecasting method seems to us to have a surprisingly good track record: Moravec, Kurzweil, and Legg especially look to have made predictions a long time ago that seem to hold up well relative to what their contemporaries probably would have said. And our model follows these models by modeling training compute scaling, though in most of our simulations the majority of progress toward AGI comes from software.
Capability benchmark trend extrapolation
This is our approach! We feel that now, in 2025, we have better evidence regarding the AGI effective compute requirement than comparisons to the human brain: specifically, we can extrapolate AIs’ performance on benchmarks. This is how the timelines portion of our model works. We set the effective compute required for AGI by extrapolating METR’s coding time horizon suite, METR-HRS.
We think it’s pretty great. Benchmark trends sometimes break, and benchmarks are only a proxy for real-world abilities, but… METR-HRS is the best benchmark currently available for extrapolating to very capable AIs, in our opinion. We think it’s reasonable to extrapolate that straight line into the future for at least the next few years.[6]
METR itself did a simple version of this extrapolation which assumed exponential growth in time horizons in calendar time. But this doesn’t account for AI R&D automation, changes to human labor or compute growth, or the possibility of time horizon doublings getting easier or harder at higher horizons.[7]
Our previous timelines model took all of these into account, though more crudely than our new AI Futures Model. Our previous model with median parameters predicted superhuman coder (SC) medians of 2027 to 2028, while our new model predicts 2031. The difference mostly comes from improvements to how we’re modeling AI R&D automation. See below for details.
Post-AGI takeoff forecasts
The literature on forecasting how capabilities progress after full automation of AI R&D is even more nascent than that which predicts AGI timelines. Past work has mostly fallen into one of two buckets:
As in timelines forecasting, we think that qualitative arguments are valuable but we think that modeling is a useful complement to qualitative arguments.
Davidson and Houlden focuses primarily on trends of how much more efficiently AIs have been able to achieve the same performance when determining whether there will be an SIE.[10]Meanwhile, we focus on estimates of the quality of AIs’ research taste, i.e. how good the AI is at choosing research directions, selecting and interpreting experiments, etc. We think that focusing on research taste quality is a more useful lens from which to view a potential SIE. If there’s an SIE we expect that it will primarily be driven by improvements in research taste.
Furthermore, because our takeoff model is integrated into a more expansive quantitative model, we have other advantages relative to Davidson and Houlden. For example, we can account for increases in the AGI project’s compute supply.[11]
How our model works
On the web app, there’s an interactive diagram explaining the parts of the model and how they relate to each other, with a corresponding full model explanation:
Here we’ll just give a brief overview.
Our model’s primary output is the trajectory of AIs’ abilities to automate and accelerate AI software R&D. We also include milestones tracking general capabilities, but these are calculated very roughly.
Our model can intuitively be divided into 3 stages. Although the same formulas are used in Stages 1, 2, and 3, new dynamics emerge at certain milestones (Automated Coder, Superhuman AI Researcher), and so these milestones delineate natural stages.
Stage 1: Automating coding
First we’ll discuss how our model predicts when coding will be fully automated. Stage 1 predicts when an Automated Coder (AC) arrives.
Automated Coder (AC). An AC can fully automate an AGI project's coding work, replacing the project's entire coding staff.[12]
Our starting point is to take the METR graph and extrapolate it exponentially, as they do, making a guess about what agentic coding time horizon would correspond to the AC milestone.
However, this simple extrapolation misses out on many important factors, such as:
Our model up through AC still centrally involves the METR trend,[15] but it attempts to incorporate the above factors and more. It also enables us to better represent/incorporate uncertainty, since we can do Monte Carlo simulations with different parameter settings.
Stage 2: Automating research taste
Besides coding, we track one other type of skill that is needed to automate AI software R&D: research taste. While automating coding makes an AI project faster at implementing experiments, automating research taste makes the project better at setting research directions, selecting experiments, and learning from experiments.
Stage 2 predicts how quickly we will go from an automated coder (AC) to a Superhuman AI researcher (SAR), an AI with research taste matching the top human researcher.
Superhuman AI Researcher (SAR): A SAR can fully automate AI R&D, making all human researchers obsolete.[16]
The main drivers of how quickly Stage 2 goes is:
Stage 3: The intelligence explosion
Finally, we model how quickly AIs are able to self-improve once AI R&D is fully automated and humans are obsolete. The endpoint of Stage 3 is asymptoting at the limits of intelligence.
The primary milestones we track in Stage 3 are:
In our simulations, we see a wide variety of outcomes ranging from a months-long takeoff from SAR to ASI, to a fizzling out of the intelligence explosion requiring further increases in compute to get to ASI.
To achieve a fast takeoff, there usually needs to be a feedback loop such that each successive doubling of AI capabilities takes less time than the last. In the fastest takeoffs, this is usually possible via a taste-only singularity, i.e. the doublings would get faster solely from improvements in research taste (with no improvements in coding, or extra compute). Whether a taste-only singularity occurs depends on which of the following dominates:
Continued improvements in coding automation matter less and less, as the project gets bottlenecked by their limited supply of experiment compute.
Timelines and takeoff forecasts
The best place to view our results is at https://www.aifuturesmodel.com/forecast.
In this section we will discuss both our model’s outputs and our all-things-considered views. As previously mentioned, we are uncertain, and don’t blindly trust our models. Instead we look at the results of the model but then ultimately make adjustments based on intuition and other factors. Below we describe the adjustments that we make on top of this model, and the results.
Eli
Here is the model’s output with my parameters along with my all-things-considered views.
To adjust for factors outside of the model, I’ve lengthened timelines (median from late 2030 to mid 2032), driven primarily by unknown model limitations and mistakes and the potential for data bottlenecks that we aren’t modeling. In summary:
I will also increase the 90th percentile from 2062. My all-things-considered distribution is: 10th percentile 2027.5, 50th percentile 2032.5, 90th percentile 2085. You can see all of the adjustments that I considered in this supplement.
Now I’ll move on to takeoff.
To get my all-things-considered views I: increase the chance of fast takeoff a little (I change AC to ASI in <1 year from 26% to 30%), and further increase the chance of <3 year takeoffs year takeoffs (I change the chance of AC to ASI in <3 years from 43% to 60%).
The biggest reasons I make my AI-R&D-specific takeoff a bit faster are:
(1) leads me to make a sizable adjustment to the tail of my distribution. I think modeling hardware and economic automation would make it more likely that if there isn’t taste-only singularity, we still get to ASI within 3 years.
I think that, as with timelines, for takeoff unknown limitations and mistakes in expectation point towards things going slower. But unlike with timelines, there are counter-considerations that I think are stronger. You can see all of the adjustments that I considered in this supplement.
Daniel
First, let me say a quick prayer to the spirit of rationality, who infrequently visits us all:
On the subject of timelines, I don’t immediately know whether my all-things-considered view should be more or less bullish than the model. Here are a few considerations that seem worth mentioning to me:
Weighing all these considerations, I think that my all-things-considered view on timelines will be to (1) push everything back one year from what the model says. So, my median for automated coder milestone 2030 instead of 2029, my median for superhuman AI researcher milestone 2031 instead of 2030.
In addition to that, I’ll (2) increase the uncertainty in both directions somewhat, so that there’s a somewhat greater chance of things going crazy in the next year (say, 9% by EOY 2026) and also a somewhat greater chance of things taking decades longer (say, still 6% that there’s no AGI even in 2050).
So, here’s my all-things-considered distribution as of today, Dec 30 2025:
On takeoff speeds:
I think my thoughts on this are pretty similar to Eli’s, modulo differences implied by our different parameter settings. Basically, take what the model (with my parameters) says, and then shift some probability mass away from the slower end and put it on the faster end of the range.
Also, whereas our model says that takeoff speeds are correlated with timelines such that shorter timelines also tends to mean faster takeoff, I’m not sure that’s correct and want to think about it more. There’s a part of me that thinks that on longer timelines, takeoff should be extremely fast due to the vast amounts of compute that will have piled up by then and due to the compute-inefficiency of whatever methods first cross the relevant thresholds by then.
So here’s a quick distribution I just eyeballed:
What info I’ll be looking for in the future & how I’ll probably update:
Comparison to our previous (AI 2027) timelines and takeoff models
These sections focus specifically on the model results with Eli’s parameter estimates (for both the AI Futures Model and the AI 2027 model).
Timelines to Superhuman Coder (SC)
This section focuses on timelines to superhuman coder (SC), which was our headline milestone in our AI 2027 timelines model: an SC represents an AI that autonomously is as productive as an AGI project modified to have all coders as competent as their best, speeding them each up by 30x, and getting 30 copies of each of them.[23]
We’ll discuss only the AI 2027 time horizon extension model in this section, due to it being simpler than the benchmarks and gaps version.[24] Below we compare the forecasted distribution of the AI 2027 model against that of the AI Futures Model.
We see that the AI Futures Model median is 4 years later than the AI 2027 model, and that it assigns a 11% chance that SC happens before the time horizon extension’s median. From now onward, we will focus on the trajectory with median parameters rather than distributions of SC dates, for ease of reasoning.
The AI 2027 time horizon extension model, with parameters set to their median values, predicts SC in Jan 2027 given superexponential-in-effective-compute time horizon growth, and SC in Sep 2028 given exponential time horizon growth. Meanwhile, the new model with median parameters predicts SC in Feb 2032. This is a 3.5-5 year difference! From now on we’ll focus on the 5 year difference, i.e. consider superexponential growth in the time horizon extension model. This is a closer comparison because in our new model, our median parameter estimate predicts superexponential-in-effective-compute time horizon growth.
The biggest reason for this difference is that we model pre-SC AI R&D automation differently, which results in such automation having a much smaller effect in our new model than in the AI 2027 one. The 5 year increase in median comes from:
You can read more about these changes and their effects in our supplementary materials.
Takeoff from Superhuman Coder onward
The AI Futures Model predicts a slower median takeoff than our AI 2027 takeoff model. Below we graph each of their forecasted distributions for how long it will take to go from SC to ASI.
We see that while the AI Futures Model’s median is longer than the AI 2027 one, it still puts 45% probability of takeoff as fast as AI 2027’s median. On the other hand, the AI Futures Model predicts a higher chance of takeoff within 10 years, 20 years, etc. Our new model is less “binary” in the sense that it gives lower probability to very fast or very slow takeoffs. This is because the AI Futures Model models compute increases.[25]
The reason the AI Futures Model model gives a lower chance of fast takeoffs is primarily that we rely on a new framework for estimating whether there’s an SIE and how aggressive it is.
Our AI 2027 takeoff model predicted the progression of capabilities post-SC. Its methodology was also fairly simple. First, we enumerated a progression of AI capability milestones, with a focus on AI R&D capabilities, though we think general capabilities will also be improving. Then, for each gap between milestones A and B, we:
In order to estimate some of the human-only time parameters, the AI 2027 takeoff forecast relied on a parameter it called r, which controlled the diminishing returns to AI R&D. It was crudely estimated by backing out the implied r from the first human-only time requirement, which was to get from SC to SAR.
The AI 2027 model assumed that there were no compute increases; under this assumption, if it r>1 then successive doublings of AI R&D uplift (what we previously called progress multiplier) gets faster over time after full AI R&D automation. Others have referred to this possibility as a software intelligence explosion (SIE). In the model, each doubling took about 0.7x as long as the previous: we’ll call the ratio of successive uplift doublings b from here onward, i.e. b<1 means successive doublings are faster and we get an SIE.[26]
In the AI Futures Model, the condition for an SIE is more complicated because we model multiple types of AI R&D; we also include compute increases, departing significantly from the behavior of an SIE. That said, there is a similar understandable concept in our model: a taste-only singularity (TOS). This is the situation in which after full AI R&D automation and with only research taste improvements (no extra coding or compute), successive doublings of AI R&D uplift get faster over time. To make the analysis much simpler, we also ignore the limits of intelligence in our analysis; these usually don’t greatly affect the takeoff to AIs, but they do slow progress down somewhat.
Under these assumptions, we can define a similar b to that analyzed in an SIE.
We estimate b by combining the following parameters:[27]
(a) the ratio of top to median researchers' value per selected experiment
(b) how quickly AIs improve at research taste as effective compute increases
(c) the rate at which software R&D translates into improved software efficiency (intuitively, the rate at which ideas are getting harder to find).
When using this framework, we get a less aggressive result (with our median parameters). Given that (a) was explicitly estimated in the AI 2027 model, and that we have a fairly aggressive estimate of (c) in the new model, implicitly most of the difference in results are coming from (b), how quickly AIs improve at research taste. We estimated this in our new model by looking at historical data on how quickly AIs have moved through the human range for a variety of metrics (more on that here).
With the AI 2027 model’s median parameters, each successive doubling of uplift took roughly 66% of the length of the previous (i.e. b=0.7).[28] The AI Futures Model’s distribution of b is below.
In the AI Futures Model model in the median case, there isn’t a TOS: each doubling would take 20% longer than the previous if taste were the only factor.[29] But we have high uncertainty: 38% of our simulations say that successive doublings get faster, and 17% are at least as aggressive as the AI 2027 model (i.e. b<0.7).[30]
Remember that unlike the AI 2027 model, the AI Futures Model models compute increases; also in practice coding automation contributes some to takeoffs.[31] Therefore, at similar levels of the separate bs we’ve defined here, takeoff in the AI Futures Model is faster.
Faster takeoffs are also correlated in our model with shorter timelines: when we filter for simulations that achieve SC in 2027, 35% of them have a b lower than the AI 2027 model’s median parameters. This is because some parameters lead to larger effects from automation both before and after SC, and furthermore we specified that there be correlations between parameters that govern how quickly coding abilities improve, and how quickly research taste abilities improve.
For further analysis of the differences between our AI 2027 and new takeoff models, see our supplementary materials.
AGI stands for Artificial General Intelligence, which roughly speaking means AI that can do almost everything. Different people give different definitions for it; in our work we basically abandon the term and define more precise concepts instead, such as AC, SIAR, TED-AI, etc. However, we still use the term AGI when we want to vaguely gesture at this whole bundle of concepts rather than pick out one in particular. For example, we’ve titled this section “AGI timelines…” and the next section “Post-AGI takeoff…” because this section is about estimating how many years there’ll be until the bundle of milestones starts to be reached, and the next section is about estimating what happens after some of them have already been reached. ↩︎
2047 for “unaided machines outperforming humans in every possible task”, and 2116 for “all human ↩︎
Some have also done extrapolations of Gross World Product, such as David Roodman’s Modeling the Human Trajectory. ↩︎
More details: ↩︎
Technically, the report predicted the arrival of Transformative AI, or TAI, which was defined as having at least as big of an impact as the Industrial Revolution. ↩︎
Rule of thumb inspired by Lindy’s Law: It’s reasonable to guess that a trend will continue for about as long as it’s been going so far. We wouldn’t dream of confidently extrapolating this trend for thirty years, for example. (We do in fact run the model into the 2050s and onward in our Monte Carlos, but we acknowledge that the probability of reality diverging dramatically from the model increases with the duration of the extrapolation.) ↩︎
Peter Wildeford has a model which has the possibility of doublings getting easier or harder, but does not model AI R&D automation or changes to labor or compute growth. ↩︎
See also: Most AI value will come from broad automation, not from R&D | Epoch AI ↩︎
GATE and the Full Takeoff Model also model the progression after full AI R&D automation, but neither of their authors claim that their model is intended to do it well. ↩︎
These estimates are then shaded up to account for capability improvements at the same compute level in addition to efficiency improvements at the same performance level. This adjustment brings the methodology closer to ours, but still we think it’s helpful to focus specifically on research taste skills. And finally, in Davidson and Houlden, everything is converted to the units of gains in the number of parallel workers, which we view as a much less natural unit than research taste quality. ↩︎
Among other advantages of having an integrated model: our model itself already bakes in most of the various adjustments that Davidson and Houlden did ad-hoc to their estimate of r, and we can generally ensure reasonable starting conditions (as opposed to Davidson and Houlden’s gradual boost). ↩︎
Our model operationalizes AC as follows: An AC, if dropped into present day, would be as productive on their own as only human coders with no AIs. That is, you could remove all human coders from the AGI project and it would go as fast as if there were only human coders. The project can use 5% of their compute supply to run ACs. ↩︎
See especially this Anthropic survey of researchers claiming >100% productivity improvements, but also this METR uplift study which found that people systematically overestimate the amount of uplift they were getting from AI assistance. ↩︎
That is, if we think that eventually there will be an AI system which outperforms humans at all horizon lengths, then that means the trend must shoot to infinity in finite time. ↩︎
That is, the part of our model that deals with AI timelines, i.e. the length of the period leading up to the “automated coder” milestone, centrally involves the METR trend. After that milestone is reached, horizon length continues to increase but isn’t directly relevant to the results. The results are instead driven by increases in automated research taste and coding automation efficiency. ↩︎
Our model operationalizes SAR as follows: if dropped into an AGI project in present day, a SAR would be as good at research taste as if there were only human researchers, who were each made as skilled as the top researcher. ↩︎
What do we mean when we say that the gap between a top human researcher and SIAR is 2x greater than that between the median and top human researcher? We mean the following. First, let’s define a transformation between AIs’ capability level b and a number of SDs relative to the median as: ↩︎
Our model operationalizes TED-AI as follows: A TED-AI is an AI system that could, if dropped into the present day & given the resources of a large tech company & three months to prep, fully automate 95% of remote work jobs in the US. It need not be able to do all 95% at the same time (perhaps there isn't enough compute to run enough copies of the TED-AI for that), but it needs to be able to do any 10% of them using only 50% of the US's AI-relevant compute. ↩︎
Our model operationalizes ASI as follows: An ASI would, if dropped into present day & given the resources of a large tech company & three months to prep, be able to fully automate 95% of remote work jobs in the US to the level where it is qualitatively 2x as much above the best human as the best human is above the median professional. Also, here we define “the median professional” not as the actual median professional but rather as what the the median professional would be, if everyone who took the SATs was professionally trained to do the task. (We standardize the population that is trained to do the task because otherwise the ASI requirement might be quite different depending on the population size and competence levels of the profession. See above regarding how we define the 2x gap.) ↩︎
Spot-checking in our model: Serial coding labor multiplier is basically the square root of parallel coding labor multiplier, and so when I look at my default parameter settings at the point where serial coding labor multiplier is ~10x (May 2030) the AIs have research taste equivalent to the median AI company researcher. Sounds about right to me. ↩︎
I’ve talked about this elsewhere but I generally think that if you don’t like using a superexponential and insist on an exponential, you need to come up with a different interpretation of what it means for a model to have horizon length X, other than the natural one (“A model has horizon length X iff you are better off hiring a human for coding tasks that take humans much longer than X, but better off using the model for coding tasks that take humans much less than X.”) Because on that interpretation, an exponential trend would never get to a model which outperforms humans at coding tasks of any length. But we do think that eventually there will be a model which outperforms humans at tasks of any length. In other words, on the natural interpretation the trend seems likely to go to infinity in finite time eventually. You can try to model that either as a smooth superexponential, or as a discontinuous phase shift… even in the latter case though, you probably should have uncertainty over when the discontinuity happens, such that the probability of it happening by time t increases fairly smoothly with t. ↩︎
For example, I want to think more about serial speed bottlenecks. The model currently assumes experiment compute will be the bottleneck. I also want to think more about the software-only-singularity conditions and whether we are missing something there, and square this with soft upper bounds such as “just do human uploads.” ↩︎
Note that with the new model, we’ve moved toward using Automated Coder (AC) as the headline coding automation milestone, which has a weaker efficiency requirement. ↩︎
That said, we note that the benchmarks and gaps version had longer median SC timelines (Dec 2028). And Eli’s all-things-considered SC median was further still in 2030, though Daniel’s was 2028. ↩︎
That said, we still think that the AI Futures Model gives too low a probability of <10 year takeoffs, because we are not modeling growth in compute due to hardware R&D automation, hardware production automation, or broad economic automation. ↩︎
As discussed here, the AI 2027 model set r=2.77 and 1.56 at different points. b=2^(1/r-1), so b=0.64 to 0.78. ↩︎
See here for a more thorough explanation of how b is calculated from our new model’s parameters. ↩︎
2^((1/2)-1) gives roughly 0.7. See how we got these numbers here. ↩︎
2^((0.315/0.248)-1). See the justification for this formula on our website. ↩︎
Note that the minimum b in our model is 0.5. This is a limitation, but in practice, we can still get very fast takeoffs. For example, if b were 0.5 and didn’t change over time, this would lead to a finite-time singularity in 2 times longer than the initial uplift doubling time. ↩︎
This could also be influenced by the uplifts being different for different milestones, or other factors. Unfortunately we haven’t had a chance to do a deep investigation, but a shallow investigation pointed toward compute increases being the primary factor. ↩︎