@janleike
Noticeably, Sonnet 4.5 verbalizes eval awareness much more than previous models. Does that invalidate our results? We did an audit based on model internals and the answer is “probably a little, but mostly not.” https://t.co/gyio068XXz
Viewing enriched Twitter post
Noticeably, Sonnet 4.5 verbalizes eval awareness much more than previous models. Does that invalidate our results? We did an audit based on model internals and the answer is “probably a little, but mostly not.” https://t.co/gyio068XXz
{
"media": [
{
"type": "photo",
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/1972731239372341438/media_0.jpg?",
"filename": "media_0.jpg"
}
],
"processed_at": "2025-10-06T13:27:24.568356",
"pipeline_version": "2.0"
} {
"type": "tweet",
"id": "1972731239372341438",
"url": "https://x.com/janleike/status/1972731239372341438",
"twitterUrl": "https://twitter.com/janleike/status/1972731239372341438",
"text": "Noticeably, Sonnet 4.5 verbalizes eval awareness much more than previous models. Does that invalidate our results?\n\nWe did an audit based on model internals and the answer is “probably a little, but mostly not.” https://t.co/gyio068XXz",
"source": "Twitter for iPhone",
"retweetCount": 1,
"replyCount": 2,
"likeCount": 36,
"quoteCount": 1,
"viewCount": 2599,
"createdAt": "Mon Sep 29 18:32:33 +0000 2025",
"lang": "en",
"bookmarkCount": 1,
"isReply": true,
"inReplyToId": "1972731237480718734",
"conversationId": "1972731237480718734",
"displayTextRange": [
0,
211
],
"inReplyToUserId": "710610891058716673",
"inReplyToUsername": "janleike",
"author": {
"type": "user",
"userName": "janleike",
"url": "https://x.com/janleike",
"twitterUrl": "https://twitter.com/janleike",
"id": "710610891058716673",
"name": "Jan Leike",
"isVerified": false,
"isBlueVerified": true,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1077523091700502528/2YCa_F4o_normal.jpg",
"coverPicture": "",
"description": "",
"location": "San Francisco, USA",
"followers": 115709,
"following": 332,
"status": "",
"canDm": true,
"canMediaTag": false,
"createdAt": "Thu Mar 17 23:36:53 +0000 2016",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 3603,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 36,
"statusesCount": 745,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"1795497960509448617"
],
"profile_bio": {
"description": "ML Researcher @AnthropicAI. Previously OpenAI & DeepMind.\nOptimizing for a post-AGI future where humanity flourishes. Opinions aren't my employer's.",
"entities": {
"description": {
"user_mentions": [
{
"id_str": "0",
"indices": [
14,
26
],
"name": "",
"screen_name": "AnthropicAI"
}
]
},
"url": {
"urls": [
{
"display_url": "jan.leike.name",
"expanded_url": "https://jan.leike.name/",
"indices": [
0,
23
],
"url": "https://t.co/Uvp4pU8R0f"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"display_url": "pic.twitter.com/gyio068XXz",
"expanded_url": "https://twitter.com/janleike/status/1972731239372341438/photo/1",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {},
"orig": {}
},
"id_str": "1972726965745733632",
"indices": [
212,
235
],
"media_key": "3_1972726965745733632",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARtgiLkwG9AACgACG2CMnDgawL4AAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABG2CIuTAb0AAKAAIbYIycOBrAvgAA",
"media_key": "3_1972726965745733632"
}
},
"media_url_https": "https://pbs.twimg.com/media/G2CIuTAb0AA78Gt.jpg",
"original_info": {
"focus_rects": [
{
"h": 950,
"w": 1696,
"x": 0,
"y": 0
},
{
"h": 1180,
"w": 1180,
"x": 298,
"y": 0
},
{
"h": 1180,
"w": 1035,
"x": 371,
"y": 0
},
{
"h": 1180,
"w": 590,
"x": 593,
"y": 0
},
{
"h": 1180,
"w": 1696,
"x": 0,
"y": 0
}
],
"height": 1180,
"width": 1696
},
"sizes": {
"large": {
"h": 1180,
"w": 1696
}
},
"type": "photo",
"url": "https://t.co/gyio068XXz"
}
]
},
"card": null,
"place": {},
"entities": {},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
}