Thanks for the post!
Nikola Jurkovic suggests that as soon as model can do a month-long task with 80% accuracy, it should be able to do any cognitive human task
I didn't mean to suggest this in my original post. My claim was something more like "a naive extrapolation means that 80% time horizons will be at least a month by the end of the decade, but I think they'll be much higher and we'll have AGI"
it has to go to infinity when we get AGI / superhuman coder.
This isn't necessarily true, as even an AGI or a superhuman coder might get worse at tasks-that-take-humans-longer compared to tasks-that-take-humans-shorter (this seems pretty likely given constant-error-rate considerations), meaning that even an extremely capable AI might be like 99.999% reliable for 1 hour tasks, but only 99.9% reliable for 10,000 hour tasks, meaning the logistic fit still has an intercept with 50%, it's just a very high number.
In order for the 50% intercept to approach infinity, you'd need a performance curve which approaches a flat line, and this seems very hard to pull off and probably requires wildly superhuman AI.
I think a 1-year 50%-time-horizon is very likely not enough to automate AI research. The reason I think AI research will be automated by EOY 2028 is because of speedups from partial automation as well as leaving open the possibility of additional breakthroughs naturally occurring.
A few considerations that make me think the doubling time will get faster:
- AI speeding up AI research probably starts making a dent in the doubling time (making it at least 10% faster) by the time we hit 100hr time horizons (although it's pretty hard to reason about the impacts here)
- I think I place some probability on the "inherently superexponential time horizons" hypothesis. The reason I think it is because to me, 1-month-coherence, 1-year-coherence, and 10-year-coherence (of the kind performed by humans) seem like extremely similar skills and will thus be learned in quick succession.
- It's plausible reasoning models decreased the doubling time from 7 months to 4 months. It's plausible we get another reasoning-shaped breakthrough.
So my best guess for the 50% and 80% time horizons at EOY 2028 are more like 10yrs and 3yrs or something. But past ~2027 I care more about how much AI R&D is being automated rather than the time horizon itself (partially because I have FUD about what N-year tasks should even look like by definition).
Grok 4 is slightly above SOTA on 50% time horizon and slightly below SOTA on 80% time horizon: https://x.com/METR_Evals/status/1950740117020389870
1I heard it from someone who works at xAI
I would have taken this class had I not graduated this spring!
A few suggestions:
I would like to cover the various ways AI could go wrong: malfunction, misuse, societal upheaval, arms race, surveillance, bias, misalignment, loss of control,...
Talk about predictions for the future, methodologies for how to come up with them.
Some technical components should include: evaluations
All of the above said, I get antsy if I don't get my dosage of math and code- I intend 80% of the course to be technical and cover research papers and results. It should also involve some hands on projects.
Another thing I consider really important: many of the students will be like "Holy shit, AGI is happening! This affects my life plans!" and will want advice. I think it's good to have something to point them to, like:
Good luck running the course!
The US Secretary of Energy says "The AI race is the second Manhattan project."
https://x.com/SecretaryWright/status/1945185378853388568
Similarly, the US Department of Energy says: "AI is the next Manhattan Project, and THE UNITED STATES WILL WIN. 🇺🇸"
https://x.com/ENERGY/status/1928085878561272223
6
1Agreed, thanks! I've moved that discussion down to timelines and probabilities.
Anthropic wrote a pilot risk report where they argued that Opus 4 and Opus 4.1 present very low sabotage risk. METR independently reviewed their report and we agreed with their conclusion.
During this process, METR got more access than during any other evaluation we've historically done, and we were able to review Anthropic's arguments and evidence presented in a lot of detail. I think this is a very exciting milestone in third-party evaluations!
I also think that the risk report itself is the most rigorous document of its kind. AGI companies will need a lot more practice writing similar documents, so that they can be better at assessing risks once AI systems become very capable.
I'll paste the twitter thread below (twitter link)
We reviewed Anthropic’s unredacted report and agreed with its assessment of sabotage risks. We want to highlight the greater access & transparency into its redactions provided, which represent a major improvement in how developers engage with external reviewers. Reflections:
To be clear: this kind of external review differs from holistic third-party assessment, where we independently build up a case for risk (or safety). Here, the developer instead detailed its own evidence and arguments, and we provided external critique of the claims presented.
Anthropic made its case to us based primarily on information it has now made public, with a small amount of nonpublic text that it intended to redact before publication. We commented on the nature of these redactions and whether we believed they were appropriate, on balance.
For example, Anthropic told us about the scaleup in effective compute between models. Continuity with previous models is a key component of the assessment, and sharing this information provides some degree of accountability on a claim that the public cannot otherwise assess.
We asked Anthropic to make certain assurances to us about the models its report aims to cover, similar to the assurance checklist in our GPT-5 evaluation. We then did in-depth follow-ups in writing and in interviews with employees.
We believe that allowing this kind of interactive review was ultimately valuable. In one instance, our follow-ups on the questions we asked helped Anthropic notice internal miscommunications about how its training methods might make chain-of-thought harder to monitor reliably.
A few key limitations. We have yet to see any rigorous roadmap for addressing sabotage risk from AI in the long haul. As AI systems become capable of subverting evaluations and/or mitigations, current techniques like those used in the risk report seem insufficient.
Additionally, there were many claims for which we had to assume that Anthropic was providing good-faith responses. We expect that verification will become more important over time, but that our current practices would not be robust to a developer trying to game the review.
Beyond assessing developer claims, we think there is an important role for third parties to do their own assessments, which might differ in threat models and approach. We would love to see the processes piloted in this review applied to such holistic assessments as well.
Overall, we felt that this review was significantly closer to the sort of rigorous, transparent third-party scrutiny of AI developer claims that we hope to see in the future. Full details on our assessment: https://alignment.anthropic.com/2025/sabotage-risk-report/2025_pilot_risk_report_metr_review.pdf