Viewing enriched Twitter post
Are the bottlenecks to rapid AI improvements purely computational? Or do they require real-world deployment and interaction to surpass? Interestingly, this was one of the main disagreements between @DKokotajlo and I almost two years ago: https://t.co/iCzcipA7ma In short, I agree with @mentalgeorge's predictions that even if AI systems can conduct research autonomously (i.e., even if we get RSI), there are real-world bottlenecks that cannot be resolved purely computationally.
{
"media": [
{
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2054581905752813641/media_0.jpg",
"media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2054581905752813641/media_0.jpg",
"type": "photo",
"filename": "media_0.jpg"
}
],
"processed_at": "2026-05-16T22:01:14.571838",
"pipeline_version": "2.0"
}{
"type": "tweet",
"id": "2054581905752813641",
"url": "https://x.com/sayashk/status/2054581905752813641",
"twitterUrl": "https://twitter.com/sayashk/status/2054581905752813641",
"text": "Are the bottlenecks to rapid AI improvements purely computational? Or do they require real-world deployment and interaction to surpass? \n\nInterestingly, this was one of the main disagreements between @DKokotajlo and I almost two years ago: https://t.co/iCzcipA7ma\n\nIn short, I agree with @mentalgeorge's predictions that even if AI systems can conduct research autonomously (i.e., even if we get RSI), there are real-world bottlenecks that cannot be resolved purely computationally.",
"source": "Twitter for iPhone",
"retweetCount": 6,
"replyCount": 3,
"likeCount": 24,
"quoteCount": 0,
"viewCount": 6401,
"createdAt": "Wed May 13 15:17:53 +0000 2026",
"lang": "en",
"bookmarkCount": 17,
"isReply": false,
"inReplyToId": null,
"conversationId": "2054581905752813641",
"displayTextRange": [
0,
276
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "sayashk",
"url": "https://x.com/sayashk",
"twitterUrl": "https://twitter.com/sayashk",
"id": "3084274082",
"name": "Sayash Kapoor",
"isVerified": false,
"isBlueVerified": false,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1521238232867946496/U_GCI43e_normal.jpg",
"coverPicture": "https://pbs.twimg.com/profile_banners/3084274082/1712802699",
"description": "",
"location": "Princeton",
"followers": 12376,
"following": 2261,
"status": "",
"canDm": true,
"canMediaTag": true,
"createdAt": "Sun Mar 15 09:03:24 +0000 2015",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 4935,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 229,
"statusesCount": 1472,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"2044835523198370272"
],
"profile_bio": {
"description": "I tweet about AI agents, AI evals, AI for science.\nAI as Normal Technology: https://t.co/5amOkqKDf2\nBook: https://t.co/DabpkhNrcM",
"entities": {
"description": {
"urls": [
{
"display_url": "bit.ly/ai-nt",
"expanded_url": "http://bit.ly/ai-nt",
"indices": [
76,
99
],
"url": "https://t.co/5amOkqKDf2"
},
{
"display_url": "bit.ly/ai-so",
"expanded_url": "http://bit.ly/ai-so",
"indices": [
106,
129
],
"url": "https://t.co/DabpkhNrcM"
}
]
},
"url": {
"urls": [
{
"display_url": "cs.princeton.edu/~sayashk",
"expanded_url": "http://cs.princeton.edu/~sayashk",
"indices": [
0,
23
],
"url": "https://t.co/wO9NAWB11f"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"allow_download_status": {
"allow_download": true
},
"display_url": "pic.twitter.com/89B2V6DB7W",
"expanded_url": "https://twitter.com/sayashk/status/2054581905752813641/photo/1",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": []
},
"orig": {
"faces": []
}
},
"id_str": "2054580573868273665",
"indices": [
277,
300
],
"media_key": "3_2054580573868273665",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARyDViVz1jABCgACHINXW45XcEkAAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABHINWJXPWMAEKAAIcg1dbjldwSQAA",
"media_key": "3_2054580573868273665"
}
},
"media_url_https": "https://pbs.twimg.com/media/HINWJXPWMAEWvOP.jpg",
"original_info": {
"focus_rects": [
{
"h": 673,
"w": 1202,
"x": 0,
"y": 0
},
{
"h": 758,
"w": 758,
"x": 0,
"y": 0
},
{
"h": 758,
"w": 665,
"x": 0,
"y": 0
},
{
"h": 758,
"w": 379,
"x": 0,
"y": 0
},
{
"h": 758,
"w": 1202,
"x": 0,
"y": 0
}
],
"height": 758,
"width": 1202
},
"sizes": {
"large": {
"h": 758,
"w": 1202
}
},
"type": "photo",
"url": "https://t.co/89B2V6DB7W"
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"urls": [
{
"display_url": "youtu.be/rVFAJQryzk8?si…",
"expanded_url": "https://youtu.be/rVFAJQryzk8?si=UQH5nO3SDuumdJwD&t=1986",
"indices": [
240,
263
],
"url": "https://t.co/iCzcipA7ma"
}
],
"user_mentions": [
{
"id_str": "1726760827452350464",
"indices": [
200,
211
],
"name": "Daniel Kokotajlo",
"screen_name": "DKokotajlo"
},
{
"id_str": "985114477",
"indices": [
288,
301
],
"name": "Tom Reed",
"screen_name": "mentalgeorge"
}
]
},
"quoted_tweet": {
"type": "tweet",
"id": "2053893908804071529",
"url": "https://x.com/mentalgeorge/status/2053893908804071529",
"twitterUrl": "https://twitter.com/mentalgeorge/status/2053893908804071529",
"text": "Reading responses to \"Goodhart Singularity\" has clarified the following to me:\n\neven if I'm wrong, and a secluded intelligence explosion is very much possible, then it still seems the case that full AI R&D automation - in the traditional sense of \"AIs can do everything human AI researchers do\" - is grossly insufficient for kicking it off.\n\nAfaict, the conditions for a secluded intelligence explosion (no widespread deployment on tasks in the real world) are:\n\n1. \"sample efficiency\" is indeed a coherent thing that can be maxxed from within the datacenter in a way that generalises to the real world.\n\n2. It is possible to design a continual learning algorithm that generalises to the real world even when you can only test it out on synthetic environments. This might not need perfect sim2real - the AIs just need to create diverse and rich enough environments to find an algo that generalises. Humans are potentially a case-in-point here, given that our continual learning transfers from the African savannah to TSMC.\n\nNow, if the AI researchers can crack (1) and (2) from the datacenter alone, they may indeed be off to the races.\n\nBut cracking (1) and (2) requires more than just \"full AI R&D automation\". Current human AI researchers are unable create these kinds of synthetic environments!! They need to deploy the models to banks, lawyers, etc instead. Consider also that Anthropic's best tool for measuring even AI R&D speedup is not an eval but simply a survey of employees.\n\nCracking (1) and (2) is far beyond what is usually considered the relevant watermark for \"AI R&D automation\" that would be at risk of kicking off an intelligence explosion, eg in Jack Clark's recent essay or AI 2027.\n\nAfaik, no one is tracking the \"can generate ecologically valid synthetic data that we could use to test our hypothetical future continual learning / sample efficiency algorithms against\" capability. But presumably this is something the automated AI researchers would need",
"source": "Twitter for iPhone",
"retweetCount": 3,
"replyCount": 4,
"likeCount": 51,
"quoteCount": 1,
"viewCount": 9464,
"createdAt": "Mon May 11 17:44:02 +0000 2026",
"lang": "en",
"bookmarkCount": 30,
"isReply": false,
"inReplyToId": null,
"conversationId": "2053893908804071529",
"displayTextRange": [
0,
282
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "mentalgeorge",
"url": "https://x.com/mentalgeorge",
"twitterUrl": "https://twitter.com/mentalgeorge",
"id": "985114477",
"name": "Tom Reed",
"isVerified": false,
"isBlueVerified": true,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1810244508061249537/KtVReF2F_normal.jpg",
"coverPicture": "https://pbs.twimg.com/profile_banners/985114477/1774273524",
"description": "",
"location": "London",
"followers": 1102,
"following": 1292,
"status": "",
"canDm": false,
"canMediaTag": true,
"createdAt": "Sun Dec 02 18:06:48 +0000 2012",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 468,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 63,
"statusesCount": 271,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [],
"profile_bio": {
"description": "Overton-truther\n\n@GovAIOrg\nFormerly @AISecurityInst",
"entities": {
"description": {
"user_mentions": [
{
"id_str": "",
"indices": [
17,
26
],
"name": "",
"screen_name": "GovAIOrg"
},
{
"id_str": "",
"indices": [
36,
51
],
"name": "",
"screen_name": "AISecurityInst"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"urls": [],
"user_mentions": []
},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"communityInfo": null,
"article": null
},
"retweeted_tweet": null,
"isLimitedReply": false,
"communityInfo": null,
"article": null
}