{
  "blyg": "0.3",
  "id": "1x7r8bz694dkcdpdtapm0ammtm",
  "kind": "fragment",
  "origin": "https://www.msweet.net/notes/",
  "page": "642-ash-otters",
  "author": {
    "name": "Matthew McDowell-Sweet",
    "url": "https://www.msweet.net/"
  },
  "created": "2026-08-28T17:16:20Z",
  "updated": "2026-08-28T17:16:20Z",
  "version": 1,
  "content_md": "Potentially the first time I’ve felt a distinct and persistent sense of moral unease about alignment research.\n\n> The panel asks how models infer what graders reward, why good behaviour can come from the wrong reason, and whether that difference can be measured. The conversation moves through promise-breaking, grader awareness, reward hacking, scheming, opaque reasoning and corrigibility, then turns to a detailed walkthrough of the contrastive-belief method and what its results do and do not show. The o3 results discussed here concern an intermediate checkpoint without safety training.\n\n[How Researchers Test AI for Hidden Goals — Apollo Research - YouTube ↗](https://youtube.com/watch?is=8V_sogy5aRD4JSCi&v=n1Qk8xbqF-M)",
  "content_html": "<p>Potentially the first time I’ve felt a distinct and persistent sense of moral unease about alignment research.</p>\n<blockquote><p>The panel asks how models infer what graders reward, why good behaviour can come from the wrong reason, and whether that difference can be measured. The conversation moves through promise-breaking, grader awareness, reward hacking, scheming, opaque reasoning and corrigibility, then turns to a detailed walkthrough of the contrastive-belief method and what its results do and do not show. The o3 results discussed here concern an intermediate checkpoint without safety training.</p></blockquote>\n<p class=\"source\"><a href=\"https://youtube.com/watch?is=8V_sogy5aRD4JSCi&amp;v=n1Qk8xbqF-M\">How Researchers Test AI for Hidden Goals — Apollo Research - YouTube ↗</a></p>",
  "content_hash": "sha256:5bb5ad09f81213dcc9d9918498236de40d1947d2ca1e0433f0295fb2763ffa81",
  "media": [],
  "changelog": [
    {
      "version": 1,
      "at": "2026-08-28T17:16:20Z",
      "note": null
    }
  ]
}