{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreihanradah5fjqxby4q4bif5gih4mdezmn3bfkwc2iklzxcywjpes4",
    "uri": "at://did:plc:pi6woz4d47bkuws673w2il2r/app.bsky.feed.post/3mgxqpbti3xc2"
  },
  "path": "/t/yet-another-opinion-on-llms-hasufells-blog/13775#post_13",
  "publishedAt": "2026-03-13T17:53:13.000Z",
  "site": "https://discourse.haskell.org",
  "textContent": "various teams come up with different patches. i was working on some interrupt and context injection stuff that got put by Qwen. was interested how they were basically tasked with what is now standardised across models.\n\nI think something like this is needed to detect extrapolation.\nthere is some kind of information theoretic quantity which compares the informational content of an output against the information its based on from prompt inputs. when you ask it to give more information than you put in you are asking for your information to be diluted with informationlessness. not a good idea. even with “relevancy“ estimated by the weights. its just conjectural hallucination. starts trying to implement sha256 in an insecure way because it cant tell which way is up. so needs a serious safeguard. until we have this it will always risk degrading information quality. which is structural specification tasks is basic poisoning.",
  "title": "Yet another opinion on LLMs · Hasufell's blog"
}