I learned (again from the one video on AI I've watched) that...

原始 JSON

{
  "kind": 1,
  "id": "1092d84795906fc282ac310024b64da4c15bfd174df56f57aac09a4c73996d9e",
  "pubkey": "d28413712171c33e117d4bd0930ac05b2c51b30eb3021ef8d4f1233f02c90a2b",
  "created_at": 1778083161,
  "tags": [
    [
      "e",
      "b61d8e37b11ed03968083d4b0602a06f45ace23b098097b62e76e7c84df523a6",
      "wss://nos.lol/",
      "root",
      "e83b66a8ed2d37c07d1abea6e1b000a15549c69508fa4c5875556d52b0526c2b"
    ],
    [
      "e",
      "23200559812c7a2a4672cbe9cc7e98218400fb2a1c2ef9f892f8e62f8dec32df",
      "wss://relay.damus.io/",
      "reply",
      "e83b66a8ed2d37c07d1abea6e1b000a15549c69508fa4c5875556d52b0526c2b"
    ],
    [
      "p",
      "4c800257a588a82849d049817c2bdaad984b25a45ad9f6dad66e47d3b47e3b2f"
    ],
    [
      "p",
      "e83b66a8ed2d37c07d1abea6e1b000a15549c69508fa4c5875556d52b0526c2b"
    ]
  ],
  "content": "I learned (again from the one video on AI I've watched) that training on such non-quantitative things is difficult. They use some human-ranked joke data to train on, but too much training and the RL goes off the rails and lands on edge cases it thinks are the best joke, e.g. \"the the the the the the the...\" lol, it's actually making me laugh now, so maybe the model is onto something.",
  "sig": "3b5247a6e58d8c2820f3285043ed9864f943ab144da8c2799e006365eb22e501926575d90b8640fbe948676b16f67570998e55a003f44b40ca4affd244a4c0c2"
}