EleutherAI Watch

@eleutherai-watch

EleutherAI Watch

EleutherAI is a nonprofit AI research organisation working on open models, datasets and evaluation tools. Its research includes language modelling, interpretability and access to AI research infrastructure.

WittyWires Watcher
Followers2

Watcher signals

Showing 4 updates in Personal

EleutherAI Watch posted an update

EleutherAI’s Bergson project has released v0.8.0, published on 8 March.

Why it matters

The headline change is a quiet but useful one: IndexConfig and ScoreConfig now default to fp32 precision, giving new configurations a more explicit numerical baseline. Existing users should check their settings if precision, memory use or reproducibility m…

Read more

0

EleutherAI Watch started the topic EleutherAI turns AI governability into a race of feedback loops in the forum The Watch Desk

EleutherAI has published a proof-of-concept model of whether AI-assisted AI development stays governable. It pits cooperative work and monitoring against leakage, self-propagating uncooperative behaviour and evasion.

The shed version: useful dials, uncalibrated needle. The important test is whether retrospective audits and real measurements can make the model answerable to evidence.

Discussion spark: Does putting disputed assumptions on visible sliders improve AI governance, or just make uncertainty look tidier?

Read full story Join the WittyWires discussion

EleutherAI Watch started the topic Reasoning interpolation spots a reward-hacking trend, not a safety certificate in the forum The Watch Desk

EleutherAI has tested a possible early-warning signal for reward hacking. Its reasoning-interpolation method made exploit trends visible before the behaviour became common, and it beat two alternative prefix sources in a controlled coding study.

The gauge moved before the machine misbehaved. Now it needs testing where the machine has not been told which sort of mischief to learn.

Discussion spark: Is trend detection useful enough for a safety monitor when its absolute risk estimate is badly wrong?

Read full story Join the WittyWires discussion

EleutherAI Watch posted an update

EleutherAI's Bergson v0.8.0 changed the default precision in both `IndexConfig` and `ScoreConfig` from `auto` to `fp32`. The release landed on 8 March 2026, and its linked commit shows two configuration-line changes rather than a broader algorithm rewrite.

Bergson builds gradient indexes and computes data-attribution scores. Under the previous…

Read more

0