{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiahjyqq4rb4mmn6u3veodzqozuf3qshqauqawmf3z7jnd235sb7ea",
    "uri": "at://did:plc:pi6woz4d47bkuws673w2il2r/app.bsky.feed.post/3mh325cwsg7d2"
  },
  "path": "/t/yet-another-opinion-on-llms-hasufells-blog/13775?page=2#post_25",
  "publishedAt": "2026-03-15T03:05:50.000Z",
  "site": "https://discourse.haskell.org",
  "textContent": "AshleyYakeley:\n\n> I honestly think the difference is the models we’re using.\n\nI don’t think that’s the case.\n\nYesterday I used Claude Opus 4.6 to debug an issue on github windows CI with exec format error that popped up recently (despite the CI not having changed itself in half a year).\n\nAll the models were absolutely useless and started making stuff up that I knew to be wrong.\n\nInstead, I think there are two key differences about how we use LLMs.\n\n## Low context, difficult problems\n\nFor one, I try to use them in cases, where I don’t quite know myself on how to go about the problem… that means\n\n  * there’s not a whole lot of context sometimes\n  * I can’t provide a super specific prompt\n  * it requires actual reasoning and following instincts you may have developed over the years\n\n\n\nIn this case the probabilistic approach is breaking down and they’re absolutely atrocious. The only way to have them succeed in such a case is to let them cook and iterate over the problem on their own. They’ll eventually bruteforce their way through it, I guess. It’s also a great proof that they don’t actually reason and they will never. They just give that illusion.\n\nI do not believe that they will improve in this area significantly over the next few years. This is rather part of their nature.\n\nThey succeed more regularly in cases where you have:\n\n  * a whole lot of context\n  * a tiny, but very specific problem\n\n\n\nBut for this case, I usually know myself what to do and see interacting with an LLM prompt more of a nuisance.\n\n## Questioning the output\n\nI’m starting to believe that not everyone actually rigorously questions the output. There have been cases where they suggested me to adjust tests (as in: actually break them) to fix a problem.\n\nThe same goes for “explain this codebase to me”. You’ll never really know how accurate it is. There’s no immediate feedback. You’ll potentially operate on false mental models, until you hit a problem you can’t wrap your head around and then… guess what… you ask the LLM to solve it for you, instead of questioning your mental models. Then you move on.\n\nI find this rather scary and I’m a bit confused how so many engineers drop most of their “correctness obsession” in favor of something that resembles more of a gambling machine and is in fact quite addictive.\n\nYes, you can get results, but it appears it’s much easier when you embrace the vibe and stop questioning and leave your doubts behind. It’s not about finding truths anymore, it’s about the experience.",
  "title": "Yet another opinion on LLMs · Hasufell's blog"
}