🐦 Twitter Post Details

Viewing enriched Twitter post

@LeoKharon

NEW WORLD MODEL: @ylecun's team is back with an efficient model! This project involves @ylecun, @lukaskuhn77, @lucasmaes_, @quentinlldc, and @randall_balestr. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

📊 Media Metadata

{
  "media": [
    {
      "url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2094395060636803122/media_0.mp4",
      "media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2094395060636803122/media_0.mp4",
      "type": "video",
      "filename": "media_0.mp4"
    }
  ],
  "processed_at": "2026-08-31T19:03:16.027136",
  "pipeline_version": "2.0"
}

🔧 Raw API Response

{
  "type": "tweet",
  "id": "2094395060636803122",
  "url": "https://x.com/LeoKharon/status/2094395060636803122",
  "twitterUrl": "https://twitter.com/LeoKharon/status/2094395060636803122",
  "text": "NEW WORLD MODEL: @ylecun's team is back with an efficient model!\n\nThis project involves @ylecun, @lukaskuhn77, @lucasmaes_, @quentinlldc, and @randall_balestr.\n\nA couple definitions first:\n- DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives.\n- SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension.\n\nLeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints.\nIt learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse.\n\nUnlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient.\nIt drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight.\n\nIt is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot.\n\nWhat I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter.\nThe \"P\" (predictor) in JEPA is effectively gone.\n\nLeVJEPA is also less compute intensive:\n- 5.6x to 20.8x lower total pretraining compute than V-JEPA 2\n- 7.6 points higher on ImageNet-1K at matched FLOPs\n- trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28\n\nAlso worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95.\nThe aggressive dropping is actually doing regularization work.\n\nOn the JEPA-versus-DINO debate:\n- it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static)\n- but wins on Something-Something-v2 by nearly 2x (motion, temporal)\n- and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost.\n-> optimized for temporal and motion understanding per compute dollar.",
  "source": "Twitter for iPhone",
  "retweetCount": 11,
  "replyCount": 6,
  "likeCount": 126,
  "quoteCount": 1,
  "viewCount": 10206,
  "createdAt": "Mon Aug 31 12:01:09 +0000 2026",
  "lang": "en",
  "bookmarkCount": 94,
  "isReply": false,
  "inReplyToId": null,
  "conversationId": "2094395060636803122",
  "displayTextRange": [
    0,
    279
  ],
  "inReplyToUserId": null,
  "inReplyToUsername": null,
  "author": {
    "type": "user",
    "userName": "LeoKharon",
    "url": "https://x.com/LeoKharon",
    "twitterUrl": "https://twitter.com/LeoKharon",
    "id": "1592447237795287045",
    "name": "Léo",
    "isVerified": false,
    "isBlueVerified": true,
    "verifiedType": null,
    "profilePicture": "https://pbs.twimg.com/profile_images/2084989693611950080/QlAj9KYg_normal.jpg",
    "coverPicture": "https://pbs.twimg.com/profile_banners/1592447237795287045/1779889418",
    "description": "",
    "location": "Paris & nvidia-smi",
    "followers": 1786,
    "following": 461,
    "status": "",
    "canDm": true,
    "canMediaTag": true,
    "createdAt": "Tue Nov 15 09:19:58 +0000 2022",
    "entities": {
      "description": {
        "urls": []
      },
      "url": {}
    },
    "fastFollowersCount": 0,
    "favouritesCount": 1000,
    "hasCustomTimelines": true,
    "isTranslator": false,
    "mediaCount": 307,
    "statusesCount": 1505,
    "withheldInCountries": [],
    "affiliatesHighlightedLabel": {},
    "possiblySensitive": false,
    "pinnedTweetIds": [
      "2094319514141741480"
    ],
    "profile_bio": {
      "description": "Robotics research & updates.\n\nCo-host @roboticsstack, the weekly pod on what's actually shipping in 🤖",
      "entities": {
        "description": {
          "user_mentions": [
            {
              "id_str": "",
              "indices": [
                38,
                52
              ],
              "name": "",
              "screen_name": "roboticsstack"
            }
          ]
        },
        "url": {
          "urls": [
            {
              "display_url": "theroboticsstack.com",
              "expanded_url": "http://theroboticsstack.com/",
              "indices": [
                0,
                23
              ],
              "url": "https://t.co/oxAp6Mxcqk"
            }
          ]
        }
      }
    },
    "isAutomated": false,
    "automatedBy": null
  },
  "extendedEntities": {
    "media": [
      {
        "additional_media_info": {
          "monetizable": false
        },
        "display_url": "pic.x.com/R97JnFrxHC",
        "expanded_url": "https://x.com/LeoKharon/status/2094395060636803122/video/1",
        "ext_master_playlist_only": [],
        "ext_media_availability": {
          "status": "Available"
        },
        "ext_playlists": [],
        "id_str": "2094395033734553600",
        "indices": [
          280,
          303
        ],
        "media_key": "7_2094395033734553600",
        "media_results": {
          "id": "QXBpTWVkaWFSZXN1bHRzOgwAAwoAAR0QyS/FFyAACgACHRDJNgiW4DIAAA==",
          "result": {
            "__typename": "ApiMedia",
            "id": "QXBpTWVkaWE6DAADCgABHRDJL8UXIAAKAAIdEMk2CJbgMgAA",
            "media_key": "7_2094395033734553600"
          }
        },
        "media_url_https": "https://pbs.twimg.com/ext_tw_video_thumb/2094395033734553600/pu/img/I7cwaoTr99x4DiZD.jpg",
        "original_info": {
          "focus_rects": [],
          "height": 1072,
          "width": 1072
        },
        "sizes": {
          "large": {
            "h": 1072,
            "w": 1072
          }
        },
        "type": "video",
        "url": "https://t.co/R97JnFrxHC",
        "video_info": {
          "aspect_ratio": [
            1,
            1
          ],
          "duration_millis": 18666,
          "variants": [
            {
              "content_type": "application/x-mpegURL",
              "url": "https://video.twimg.com/ext_tw_video/2094395033734553600/pu/pl/sifEOzt7z27oKmcM.m3u8?tag=12"
            },
            {
              "bitrate": 432000,
              "content_type": "video/mp4",
              "url": "https://video.twimg.com/ext_tw_video/2094395033734553600/pu/vid/avc1/320x320/4zJUew4F9_m3O_Nx.mp4?tag=12"
            },
            {
              "bitrate": 832000,
              "content_type": "video/mp4",
              "url": "https://video.twimg.com/ext_tw_video/2094395033734553600/pu/vid/avc1/540x540/sYFeCaGEs2lVMNZ0.mp4?tag=12"
            },
            {
              "bitrate": 1280000,
              "content_type": "video/mp4",
              "url": "https://video.twimg.com/ext_tw_video/2094395033734553600/pu/vid/avc1/720x720/PkzcxUcnJ2NIaXf-.mp4?tag=12"
            }
          ]
        }
      }
    ]
  },
  "card": null,
  "place": {},
  "entities": {
    "hashtags": [],
    "symbols": [],
    "timestamps": [],
    "urls": [],
    "user_mentions": [
      {
        "id_str": "48008938",
        "indices": [
          17,
          24
        ],
        "name": "Yann LeCun",
        "screen_name": "ylecun"
      },
      {
        "id_str": "48008938",
        "indices": [
          88,
          95
        ],
        "name": "Yann LeCun",
        "screen_name": "ylecun"
      },
      {
        "id_str": "2020483253065351168",
        "indices": [
          97,
          109
        ],
        "name": "Lukas Kuhn",
        "screen_name": "lukaskuhn77"
      },
      {
        "id_str": "1214312606711701504",
        "indices": [
          111,
          122
        ],
        "name": "Lucas Maes",
        "screen_name": "lucasmaes_"
      },
      {
        "id_str": "2015819922425745408",
        "indices": [
          124,
          136
        ],
        "name": "Quentin Le Lidec",
        "screen_name": "quentinlldc"
      },
      {
        "id_str": "1246070462679040000",
        "indices": [
          142,
          158
        ],
        "name": "Randall Balestriero",
        "screen_name": "randall_balestr"
      }
    ]
  },
  "quoted_tweet": null,
  "retweeted_tweet": null,
  "isLimitedReply": false,
  "communityInfo": null,
  "article": null
}