Forecasts
This page shows our timelines and takeoff forecasts. We are highly uncertain about this, and have expressed our uncertainty as a probability distribution over the possible times when each milestone might be reached. We show the raw result of a Monte Carlo simulation of our model, as well as our subjective all-things-considered probability distributions. We plan to keep this page up to date as our predictions change.
What do we mean by "all-things-considered"?
Though we view the model's outputs as an important source of evidence about what future AI progress might look like, we don't blindly trust it. Our all-things-considered views are informed by looking at the results of the model but then making adjustments based on intuition and which factors the model doesn't include.
Timelines to Automated AI R&D
The chart below shows how long we project it will take to achieve AC (Automated Coder) and SAR (Superhuman AI Researcher) (f you toggle it on). The x-axis is the year AC is achieved, and the y-axis is the probability density at a point in time, expressed in the % chance the milestone would happen within a year at that density level.
Chart Settings
Dashed lines
Probability densities are estimated based on 10,000 simulated trajectories.
Eli's notes on their all-things-considered forecast
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:
- Unknown model limitations and mistakes. With our previous (AI 2027) timelines model, my instinct was to push my overall forecasts longer due to unknown unknowns, and I'm glad I did. My median for SC was 2030 as opposed to the model's output of Dec 2028, and I now think that the former looks more right. I again want to lengthen my overall forecasts for this reason, but less so because our new model is much more well-tested and well-considered than our previous one, and is thus less likely to have simple bugs or unknown simple conceptual issues.
- Data bottlenecks. Our model implicitly assumes now that any data progress is proportional to algorithmic progress. But data in practice could be either more or less bottlenecking. My guess is that modeling data would lengthen timelines a bit, at least in cases where synthetic data is tough to fully rely upon.
I also increase the 90th percentile from 2062. You can see all of the adjustments that I considered in this supplement.
Takeoff Speeds
The chart below shows how long we project it will take to reach various milestones after achieving AC (Automated Coder). The x-axis represents years after AC achievement, and the curves show the cumluative probability for when each subsequent milestone might be reached.
Chart Settings
Dashed lines
Eli's notes on their all-things-considered forecast
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:
- Automation of hardware R&D, hardware production, and general economic automation. We aren't modeling these, and while they have longer lead times than software R&D, a year might be enough for them to make a substantial difference.
- Shifting to research directions which are less compute bottlenecked might speed up takeoff, and isn't modeled. Once AI projects have vast amounts of labor, they can focus on research which loads more heavily on labor relative to experiment compute than current research.
The former issue 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.
In our results analysis, we analyze which parameters are most important for the above forecasts. We also examine the correlation in our model between short timelines and fast takeoffs.