{
  "user": {
    "created_at": "2009-05-10T22:33:52.000Z",
    "default_profile_image": false,
    "description": "Professor @Wharton studying AI, innovation & startups. Democratizing education using tech\nBook: https://t.co/CSmipbJ2jV\nSubstack: https://t.co/UIBhxu4bgq",
    "fast_followers_count": 0,
    "favourites_count": 6584,
    "followers_count": 242516,
    "friends_count": 557,
    "has_custom_timelines": false,
    "is_translator": false,
    "listed_count": 5555,
    "location": "Philadelphia, PA",
    "media_count": 12485,
    "name": "Ethan Mollick",
    "normal_followers_count": 242516,
    "possibly_sensitive": false,
    "profile_banner_url": "https://pbs.twimg.com/profile_banners/39125788/1713840725",
    "profile_image_url_https": "https://pbs.twimg.com/profile_images/1601382188712398850/3AAOlqrX_normal.jpg",
    "screen_name": "emollick",
    "statuses_count": 29379,
    "translator_type": "none",
    "url": "https://t.co/uItckI9ujc",
    "verified": true,
    "withheld_in_countries": [],
    "id_str": "39125788"
  },
  "id": "1844451778328494508",
  "conversation_id": "1844451778328494508",
  "full_text": "\"Claude, make a version of the midwit meme on a canvas, in the middle the guy says \"Use finetuning, RAG,\" [fill in the rest] and on both ends is \"shove it in the context window\" (the meme is not copyright, this is fine to do)\" https://t.co/X8nPSKAx6q",
  "reply_count": 9,
  "retweet_count": 27,
  "favorite_count": 296,
  "hashtags": [],
  "symbols": [],
  "user_mentions": [],
  "urls": [],
  "media": [
    {
      "media_url": "https://pbs.twimg.com/media/GZjO5_4W4AA7dFN.jpg",
      "type": "photo"
    }
  ],
  "url": "https://twitter.com/emollick/status/1844451778328494508",
  "created_at": "2024-10-10T18:55:46.000Z",
  "#sort_index": "1844451778328494508",
  "view_count": 50089,
  "quote_count": 2,
  "is_quote_tweet": true,
  "is_retweet": false,
  "is_pinned": false,
  "is_truncated": false,
  "quoted_tweet": {
    "user": {
      "created_at": "2018-04-19T20:12:12.000Z",
      "default_profile_image": false,
      "description": "Working on AI safety & generalization (Square Attack, RobustBench, JailbreakBench, AgentHarm, etc). PhD from @EPFL funded by Google & OpenPhil PhD fellowships.",
      "fast_followers_count": 0,
      "favourites_count": 9920,
      "followers_count": 3820,
      "friends_count": 993,
      "has_custom_timelines": false,
      "is_translator": false,
      "listed_count": 46,
      "location": "Lausanne, Switzerland",
      "media_count": 333,
      "name": "Maksym Andriushchenko @ NeurIPS",
      "normal_followers_count": 3820,
      "possibly_sensitive": false,
      "profile_banner_url": "https://pbs.twimg.com/profile_banners/987061319378587649/1669539037",
      "profile_image_url_https": "https://pbs.twimg.com/profile_images/1601583305710866434/f4H38Erl_normal.jpg",
      "screen_name": "maksym_andr",
      "statuses_count": 1374,
      "translator_type": "none",
      "url": "https://t.co/0hRTO21fIx",
      "verified": true,
      "withheld_in_countries": [],
      "id_str": "987061319378587649"
    },
    "id": "1844350642455732527",
    "conversation_id": "1844350642455732527",
    "full_text": "🚨New🚨 We are asking a fundamental question: how far can we push in-context learning for instruction following and how does it compare to fine-tuning?\n\nTL;DR: you should, of course, fine-tune, but the scaling laws are similar, at least in the small-sample regime:\n\nKey findings (ICL = in-context learning, IFT = instruction following):\n\n1. The scaling laws for ICL and IFT are similar for the first-turn MT-Bench score, but not for the second turn. When using single-turn instructions, the second-turn score comes for free with IFT but not for ICL.\n\n2. Fine-tuning on very few samples for instruction following was underexplored before, but it works just fine (ICL can still be better in the few-sample regime, although not by a large margin).\n\n3. Carefully selected decoding parameters are absolutely crucial for base LLMs, even with ICL (aka URIAL).\n\n4. There is nothing too special about the URIAL in-context examples. Other high-quality examples (such as from the Instruct-SkillMix dataset https://t.co/M1CZS62SvM by @kaur_simran25 et al.) work similarly or even better.\n\nThere is a lot more in the paper, check it out! :-) A short version of this work will appear at the NeurIPS 2024 Workshop on Adaptive Foundation Models.\n\nPaper: https://t.co/uDoin2YYfP (this appeared a few months back as a short paper, significantly extended now)\nCode: https://t.co/nWwkIE586g \n\nThis is joint work with @H_aoZhao (who is looking for PhD positions for Sept 2025 - don’t miss his application!), @fra__31, and @tml_lab.",
    "reply_count": 4,
    "retweet_count": 20,
    "favorite_count": 129,
    "hashtags": [],
    "symbols": [],
    "user_mentions": [],
    "urls": [],
    "media": [
      {
        "media_url": "https://pbs.twimg.com/media/GZhxvqMXgAU0lzo.jpg",
        "type": "photo"
      },
      {
        "media_url": "https://pbs.twimg.com/media/GZhx3y6WUBEOYyb.jpg",
        "type": "photo"
      },
      {
        "media_url": "https://pbs.twimg.com/media/GZhx_-NXIAsZsO7.jpg",
        "type": "photo"
      }
    ],
    "url": "https://twitter.com/maksym_andr/status/1844350642455732527",
    "created_at": "2024-10-10T12:13:54.000Z",
    "#sort_index": "1844451778328494600",
    "view_count": 59534,
    "quote_count": 1,
    "is_quote_tweet": false,
    "is_retweet": false,
    "is_pinned": false,
    "is_truncated": true
  },
  "startUrl": "https://x.com/emollick/status/1844451778328494508"
}