Experimental browser for the Atmosphere
Redwood Research’s Aryan Bhatt presented “defer to resample” as a control technique: instead of always defaulting to a trusted model, resampling suspicious commands may yield better outcomes when comparing untrusted and less capable trusted models.
May 8, 2025, 3:31 PM
{ "uri": "at://did:plc:y3gesurmkjrpbuw2b47qiplf/app.bsky.feed.post/3looa7upwcj2s", "cid": "bafyreifltapd4c2tom5qmsataujora7ye3e2p3ccl6sinvptcyct5lkyfe", "value": { "text": "Redwood Research’s Aryan Bhatt presented “defer to resample” as a control technique: instead of always defaulting to a trusted model, resampling suspicious commands may yield better outcomes when comparing untrusted and less capable trusted models.", "$type": "app.bsky.feed.post", "embed": { "$type": "app.bsky.embed.video", "video": { "$type": "blob", "ref": { "$link": "bafkreigfid6shqlwx5akfqocflo3lqfnzpvuxnafubytorux5xa4vsrb3y" }, "mimeType": "video/mp4", "size": 4045359 }, "aspectRatio": { "width": 720, "height": 720 } }, "createdAt": "2025-05-08T15:31:37.696Z" } }