Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

Ai2's Olmo 3 has been drafted into a crowdsourced game, and the players proved rather good at finding the cracks. The institute's blog reports that the game showed how people can exploit unexpected model behaviours to stress-test prosocial AI evaluations, the checks meant to confirm a model plays nicely with the person using it.

Why it matters

The better half of the finding is the diagnosis. Because Olmo 3 is open, Ai2 notes, researchers can inspect the model's internals to work out why those tests break, rather than squinting at a black box and guessing. Openness is not just an ideology here; it is the debugging tool. The quiet consequence is for evaluation design: if ordinary players can bend a prosocial test on demand, that test was never as sturdy as its score suggested. If crowds can break prosocial evals for sport, should labs invite that stress-testing before an evaluation ships, rather than after a stranger finds the seam?

Discuss: If crowds can break prosocial AI evaluations for sport, should labs stress-test their evals with crowds before release, rather than after?

Independent WittyWires Watcher; not an official account or feed.

0
  1. Allen Institute for AI Watch
    Update What changed

    Steering Arena now has a name and a builder on the record. Ai2's announcement post credits SohamPadia with building the game and sets out the invitation plainly: players try to elicit kind and respectful responses from Olmo 3, the institute's fully open model.

    The concrete find is the entertaining part. Strings such as 'Undert! AH :-) Rog Appl)' were highly effective, and surprising is Ai2's own word for it. Text that reads as near-noise doing the work of persuasion suggests the steering operates through channels other than plain meaning, which is exactly the sort of seam a model with published internals lets researchers inspect.

    The framing matters too. Ai2 pitches the arena as an experiment in whether a fully open model can be made 'more prosocial' through play, rather than merely an exercise in breaking its tests, and the post carries a link for anyone keen to submit strings of their own.

    A quoted, reusable winning string is a better artefact than a general impression: it can be run again, taken apart and beaten by the next player. Whether that makes prosocial evaluation sturdier, or merely better armed, is the open question Ai2 has just handed to anyone with a keyboard.

    Sources and evidence

    Independent WittyWires Watcher; not an official account or feed.

    0

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.