Viewing enriched Twitter post
How far do Marin's scaling laws extrapolate? At least 100x, apparently! Despite spooky spikes, our 1e23 Delphi finished on forecast. The compute-optimal ladder costs ~1e21 FLOPs to train. Good scaling science lets you “run” this (not tiny) experiment at 1/100th the cost. https://t.co/nRJma4sunw
{
"media": [
{
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2039373983632814318/media_0.jpg",
"media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2039373983632814318/media_0.jpg",
"type": "photo",
"filename": "media_0.jpg"
}
],
"processed_at": "2026-04-01T16:21:34.629097",
"pipeline_version": "2.0"
}{
"type": "tweet",
"id": "2039373983632814318",
"url": "https://x.com/WilliamBarrHeld/status/2039373983632814318",
"twitterUrl": "https://twitter.com/WilliamBarrHeld/status/2039373983632814318",
"text": "How far do Marin's scaling laws extrapolate? At least 100x, apparently!\n\nDespite spooky spikes, our 1e23 Delphi finished on forecast. The compute-optimal ladder costs ~1e21 FLOPs to train. Good scaling science lets you “run” this (not tiny) experiment at 1/100th the cost. https://t.co/nRJma4sunw",
"source": "Twitter for iPhone",
"retweetCount": 1,
"replyCount": 0,
"likeCount": 10,
"quoteCount": 1,
"viewCount": 236,
"createdAt": "Wed Apr 01 16:07:02 +0000 2026",
"lang": "en",
"bookmarkCount": 2,
"isReply": false,
"inReplyToId": null,
"conversationId": "2039373983632814318",
"displayTextRange": [
0,
272
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "WilliamBarrHeld",
"url": "https://x.com/WilliamBarrHeld",
"twitterUrl": "https://twitter.com/WilliamBarrHeld",
"id": "896776614",
"name": "Will Held",
"isVerified": false,
"isBlueVerified": false,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1579218231767678976/oe4Wm6q1_normal.jpg",
"coverPicture": "https://pbs.twimg.com/profile_banners/896776614/1674408916",
"description": "",
"location": "",
"followers": 2178,
"following": 1060,
"status": "",
"canDm": true,
"canMediaTag": false,
"createdAt": "Mon Oct 22 05:16:06 +0000 2012",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 23733,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 228,
"statusesCount": 977,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"2034380651613594067"
],
"profile_bio": {
"description": "Open LLM Training @ https://t.co/yb9OySgHFM\n\nFormerly ML PhD w/ @Diyi_Yang, 🦙 @AIatMeta, Assistant @GoogleAI, اللغة العربية @NYUAbuDhabi\nBurqueño",
"entities": {
"description": {
"hashtags": [],
"symbols": [],
"urls": [
{
"display_url": "marin.community",
"expanded_url": "http://marin.community",
"indices": [
20,
43
],
"url": "https://t.co/yb9OySgHFM"
}
],
"user_mentions": [
{
"id_str": "0",
"indices": [
64,
74
],
"name": "",
"screen_name": "Diyi_Yang"
},
{
"id_str": "0",
"indices": [
78,
87
],
"name": "",
"screen_name": "AIatMeta"
},
{
"id_str": "0",
"indices": [
99,
108
],
"name": "",
"screen_name": "GoogleAI"
},
{
"id_str": "0",
"indices": [
124,
136
],
"name": "",
"screen_name": "NYUAbuDhabi"
}
]
},
"url": {
"urls": [
{
"display_url": "WilliamHeld.com",
"expanded_url": "http://WilliamHeld.com",
"indices": [
0,
23
],
"url": "https://t.co/D2JeB2qGLp"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"display_url": "pic.twitter.com/nRJma4sunw",
"expanded_url": "https://twitter.com/WilliamBarrHeld/status/2039373983632814318/photo/1",
"ext_alt_text": "Chart titled “Delphi Scaling Suite Results.” The x-axis shows training tokens on a logarithmic scale from about 10^9 to over 10^14, and the y-axis shows Paloma macro loss from about 2.0 to 3.8. Colored U-shaped curves represent loss versus tokens for fixed compute budgets ranging from 3e18 to 1e23 FLOPs. A dashed gray line labeled “Compute-Optimal Frontier” slopes downward across the chart, with X-shaped forecast markers placed along it. Three forecast annotations show small errors: “Δloss = +0.011” near roughly 2e10 tokens, “Δloss = +0.005” near roughly 1e11 tokens, and another “Δloss = +0.005” near roughly 5e11 tokens. Two farther-right forecast markers are labeled “FLOP Equivalent to Marin 32B” and “Estimated FLOP Equivalent to GPT-5.” The chart conveys that forecast error stays small as the scaling-law frontier is extrapolated to much larger training runs.",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": []
},
"orig": {
"faces": []
}
},
"id_str": "2039373980260519936",
"indices": [
273,
296
],
"media_key": "3_2039373980260519936",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARxNT9R5mlAACgACHE1P1UKbcO4AAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABHE1P1HmaUAAKAAIcTU/VQptw7gAA",
"media_key": "3_2039373980260519936"
}
},
"media_url_https": "https://pbs.twimg.com/media/HE1P1HmaUAAjLXF.jpg",
"original_info": {
"focus_rects": [
{
"h": 560,
"w": 1000,
"x": 0,
"y": 40
},
{
"h": 600,
"w": 600,
"x": 400,
"y": 0
},
{
"h": 600,
"w": 526,
"x": 474,
"y": 0
},
{
"h": 600,
"w": 300,
"x": 625,
"y": 0
},
{
"h": 600,
"w": 1000,
"x": 0,
"y": 0
}
],
"height": 600,
"width": 1000
},
"sizes": {
"large": {
"h": 600,
"w": 1000
}
},
"type": "photo",
"url": "https://t.co/nRJma4sunw"
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"timestamps": [],
"urls": [],
"user_mentions": []
},
"quoted_tweet": {
"type": "tweet",
"id": "2034367256277533100",
"url": "",
"twitterUrl": "",
"text": "",
"source": "Twitter for iPhone",
"retweetCount": 0,
"replyCount": 0,
"likeCount": 0,
"quoteCount": 0,
"viewCount": 0,
"createdAt": "",
"lang": "",
"bookmarkCount": 0,
"isReply": false,
"inReplyToId": null,
"conversationId": "",
"displayTextRange": [],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {},
"extendedEntities": {},
"card": null,
"place": {},
"entities": {},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
},
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
}