←
jsd6,859 posts
jsd avatar

jsd

@datagenproc

@METR_Evals - Opinions my own. Anonymous feedback: admonymous.co/jsd

◷ Joined August 2022▧ 577 Photos and videos
jsd avatar
jsd@datagenproc·Oct 23

Personal news: after two years at @EpochAIResearch, I’ve joined @METR_Evals to work on embedded assessments of AI R&D acceleration: going inside AI developers to measure how much AI is speeding up AI development, and publishing our findings.

Epoch AI avatar
⟳ jsd retweeted
Epoch AI@EpochAIResearch·23h

Can the most capable LLMs match humans at real work? Not yet, at least not on the kind of work we do at Epoch. We tested GPT-6.2 Astra and other recent models on Epoch Automation Reports, our benchmark of open-ended Epoch work tasks, ranging from making Epoch-style diagrams to designing basic research experiments. An Epoch employee scores the models’ outputs based on our internal quality standards. A score of 100% represents the standard expected of an Epoch employee. All models failed to meet the Epoch employee standard. GPT-6.2 Astra ranks fourth (averaging 65%), despite sharing the top ECI score of 174 with Fable 5.2. Fable 5.2 scored highest, averaging 82%. On one task, both Fable 5.2 and Opus 5.5 matched the level of a contractor. More on that below.

Post attachment
Fulcrum avatar
⟳ jsd retweeted
Fulcrum@fulcrum_inc·Nov 13

The Fulcrum team is joining @METR_Evals to work on alignment measurement. We feel the world is very unprepared for AI capabilities acceleration. We decided the best thing we could do is work on open science that informs the public.

jsd avatar
⟳ jsd retweeted
jsd@datagenproc·Nov 13

Still processing through updates but I worry that FUD around "verifiable tasks" is smuggling in alignment concerns, and matters a lot more for "timelines to productive drop-in worker you’d be happy to delegate to" than "timelines to takeover-capable AI".

jsd avatar
jsd@datagenproc·Nov 12

Note the 3 month prep clause in AIFP's TED-AI definition: "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."

jsd avatar
jsd@datagenproc·Nov 12

This suggests that startups that integrated AI from the very start should soon/now outperform incumbents.

jsd avatar
jsd@datagenproc·Nov 12

One of the reasons I think this is only fairly convincing, but far from knockdown is that I suspect there's a huge amount of context overhang: that the main reason AIs do poorly on many white collar tasks is that they haven't been given same amount of context as humans.

I find "TEDAI is just so many tasks. We only have evidence about trends and levels on a relatively small number of tasks." fairly convincing.

jsd avatar
jsd@datagenproc·Nov 11

2 ways in which pre-training-software-progress, inference efficiency, or data efficiency would be evidence for "AI progress on well-specified tasks is accelerating" ① They are intermediate factors in the (compute, data, labor, etc.) → AECI production function. So if they accelerate it's more likely that we'll see AECI accelerate in the future. I would find it hard to update a lot on this without knowing a lot more about the production function. ② They are themselves examples of "well-specified tasks". So them going to the moon provides early evidence of how much faster well-specified tasks can get done if you actually try.

jsd avatar
jsd@datagenproc·Nov 11

This is a great result, and I am very excited about the Oversight Foundation Models agenda.

Finetuning LLMs can sometimes cause surprising side effects. Can we predict a finetuned model's response from its training data before burning compute? We trained oversight models to do this and found performance follows power laws in model size and data. 🧵

Post attachment
Josh You avatar
⟳ jsd retweeted
Josh You@justjoshinyou13·Feb 9

many metrics of AI progress are linear over time, including log(METR time horizon). however, some, like exp(ECI), grow exponentially

METR avatar
⟳ jsd retweeted
METR@METR_Evals·Nov 5

We reviewed the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report, with access to the nonpublic version. We focus on the report’s claims about catastrophic risks from automated AI R&D and about whether Anthropic had crossed relevant capability thresholds.

METR avatar
⟳ jsd retweeted
METR@METR_Evals·Nov 5

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.

jsd avatar
jsd@datagenproc·Nov 2

I think Confidence Intervals are too often a ritual where the person adding them doesn’t actually know what question they’re answering, it’s just a thing you do.