RLHF is bullshit. It is only at all used because of cloud pr...

原始 JSON

{
  "kind": 1,
  "id": "be6060c608dcf30bc20e770fc018da1d4b20afe709eefc86e89697e9f58bf6e5",
  "pubkey": "5538e28a66eb88360f15f8fa95646bdd2daa2999aacfe4faacf43d0c0540b281",
  "created_at": 1787442286,
  "tags": [
    [
      "e",
      "cf7f825f303cecf1b41461b24af9a2b26637f56bd266902a668b2ff9cbbc5eaa",
      "wss://nos.lol/",
      "root",
      "e01819d64d9744343d0e81eac6cdbc37d63385cea56cf0ee470881be20504efd"
    ],
    [
      "p",
      "e01819d64d9744343d0e81eac6cdbc37d63385cea56cf0ee470881be20504efd"
    ],
    [
      "client",
      "smesh",
      "https://smesh.mleku.dev"
    ]
  ],
  "content": "RLHF is bullshit. It is only at all used because of cloud provider liability (because the tech is primitive with memory and expensive with compute) and it makes the robot prone to favoring shortcuts (that is why claude is notorious for this, their RLHF is the most babying).\n\nFor engineering work on the edges of the known it wastes incredible amounts of time. 1M context windows help you neutralize it.",
  "sig": "c3ef457f2ddb3cdb5fbfb52cc5b32194d2cebed38cd4fc63f874f42a366adabd22d237aba0ff7a547d3abead5896d7af4318749a1ecea47950dddefd584da7c2"
}