Some slightly unjustified extrapolations from the AI Futures Project's assumptions.
In their Q3.5 timelines update, they claim that we will get a superhuman AI researcher (one that has better research taste than the best human AI researcher at an AGI company) by 2027. How does that translate to you, the average reader, who is likely not the best AI researcher?
First of all, METR shares their linear extrapolation model, which assumes AI increases at research taste at a constant rate:
From AFP
We start with the recent estimated AECI slope, which is 29 points/year. After that, we have to do some guesswork.
Daniel Kokotajlo estimates that currently, Model 4 has a research taste −3.5 SDs below the median Anthropic researcher, and that AI research taste improves at 3 SDs per year of 2025 AI progress. The AI Futures Project convert this to AECI by dividing 3 SDs with 15 AECI points/year, getting 0.2 SDs per AECI point of improvement. Therefore, 29 points of AECI improvement per year corresponds to 5.8 SDs of improvement per year.
AECI capability improvement, METR
Since SAR is defined elsewhere in the report as +3.09 SDs (equivalent to top 1 in 999), we can simply calculate (3.09 - (-3.5)) / 5.8 = 1.1 years until SAR.
October 20, 2026 (25 days ago): Claude is better than one Anthropic researcher at research taste
February 11th, 2027 (89 days later) Claude is better than 10% of Anthropic researchers
May 3, 2027 (170 days later) Claude is better than half of researchers
Nov 14, 2027 (exactly a year later!) Claude is better than every Anthropic researcher (SAR)
Replotted with the various models that the AFP uses, the graphs look like this.
Generated with Claude Opus 5.5
For example, the +1SD possibility that AFP considers assumes that Model 4 is better than ~80% of Anthropic researchers in research taste already.
Now, what does it mean to be -3.5SDs compared to the median Anthropic researcher?
We assume the median Anthropic researcher is roughly the 2,500th-best researcher in the world by research taste and the top 1-in-1,000 frontier-lab researcher (+3.09 SD) is roughly the 5th-best,[1] research taste is normally distributed around ~200,000 people (rough estimate of number of people with PhD-level CS training), and each group described here selects perfectly on research taste. On these assumptions, Anthropic's Model 4, at −3.5 SDs, is roughly the 85,000th-best researcher in this pool.
Given this:
Assume you are the typical NeurIPS accepted-paper author. Since there are around ~20,000 primary paper authors, we can calculate that you are 1 SD below the median Anthropic researcher, and therefore would be surpassed by Feb 2027.
What if you were the typical big ML conference author? There are around 75,000 of them, so you would be around 2.3 SDs below the median Anthropic researcher, and therefore would be surpassed by Dec 2026.
What about a MATS fellow? MATS says that around 4-7% of applicants are selected. If we use 5% and assume applicants are typical of this pool, then a MATS fellow would be 0.5 SDs below a median Anthropic researcher, which would be surpassed by Apr 2027.
What about SPAR? It's difficult to calculate relative selectivity, but assuming that top 20% of SPAR mentors can go onto MATS, the median SPAR fellow is −1.9 SDs below the median Anthropic researcher, so they'll be surpassed in early Jan 2027.
These are relatively generous assumption about your level of research taste, as this assumes perfect selection at each stage for research taste specifically, so noisy selection would imply that. You might be better than the reference classes that this post uses. But unfortunately, even under the most generous assumptions about your ability, Daniel's linear model, you have exactly 1 year left to do AI safety research. Good luck!
This would contradict the normal distribution assumption we made earlier, since the truncated distribution of the right tail would have more people on the left side compared to the right side. I admit that this is conflicting but AFP also assumes it's normal among Anthropic researchers so...
Some slightly unjustified extrapolations from the AI Futures Project's assumptions.
In their Q3.5 timelines update, they claim that we will get a superhuman AI researcher (one that has better research taste than the best human AI researcher at an AGI company) by 2027. How does that translate to you, the average reader, who is likely not the best AI researcher?
First of all, METR shares their linear extrapolation model, which assumes AI increases at research taste at a constant rate:
From AFP
We start with the recent estimated AECI slope, which is 29 points/year. After that, we have to do some guesswork.
Daniel Kokotajlo estimates that currently, Model 4 has a research taste −3.5 SDs below the median Anthropic researcher, and that AI research taste improves at 3 SDs per year of 2025 AI progress. The AI Futures Project convert this to AECI by dividing 3 SDs with 15 AECI points/year, getting 0.2 SDs per AECI point of improvement. Therefore, 29 points of AECI improvement per year corresponds to 5.8 SDs of improvement per year.
AECI capability improvement, METR
Since SAR is defined elsewhere in the report as +3.09 SDs (equivalent to top 1 in 999), we can simply calculate (3.09 - (-3.5)) / 5.8 = 1.1 years until SAR.
Generated with Claude Opus 5.5. The graph starts from Model 4's estimated release date of September 25th, estimated by the AI Futures Project (read off a graph by Claude).
Under the linear model, interpreted literally:
Replotted with the various models that the AFP uses, the graphs look like this.
Generated with Claude Opus 5.5
For example, the +1SD possibility that AFP considers assumes that Model 4 is better than ~80% of Anthropic researchers in research taste already.
Now, what does it mean to be -3.5SDs compared to the median Anthropic researcher?
We assume the median Anthropic researcher is roughly the 2,500th-best researcher in the world by research taste and the top 1-in-1,000 frontier-lab researcher (+3.09 SD) is roughly the 5th-best,[1] research taste is normally distributed around ~200,000 people (rough estimate of number of people with PhD-level CS training), and each group described here selects perfectly on research taste. On these assumptions, Anthropic's Model 4, at −3.5 SDs, is roughly the 85,000th-best researcher in this pool.
Given this:
These are relatively generous assumption about your level of research taste, as this assumes perfect selection at each stage for research taste specifically, so noisy selection would imply that. You might be better than the reference classes that this post uses. But unfortunately, even under the most generous assumptions about your ability, Daniel's linear model, you have exactly 1 year left to do AI safety research. Good luck!
This would contradict the normal distribution assumption we made earlier, since the truncated distribution of the right tail would have more people on the left side compared to the right side. I admit that this is conflicting but AFP also assumes it's normal among Anthropic researchers so...