alt Hacker NewsWhat I mean is, using LLMs for search and coding, they return the wrong answer all the time, unpredictably. Often if I redirect them, they will agree that they were wrong. So even though the knowledge that I need is in there somewhere, it doesn't get applied reliably. I don't want that experience for my infrastructure, where the model usually gets it right, but it just goes wrong a noticeable fraction of the time.
Do you know what I mean?
Yes, I do get that, I was simply taking an issue with the characterization, cause it's super catchy but conflates several distinct failure modes.
The model emitting nonsense is mostly the model just being terrible in general. The model answering slightly different each time is inference details, like temperature, seed, batching, hardware, etc. The model responding to the same inquiry very differently is the model being semantically unstable. These all exist but they are distinct, and social media isn't kind to nuance to begin with.
You gestured at the second previously, but what you describe now is mostly the former. I don't really experience this anymore, but I do typically use Opus, and work an area where it's clearly excelling in, so that's probably why. Every time I go back to Sonnet, the bullshit parade begins, no matter the amount of context provided, or the thinking effort level configured. Mostly tolerable, but I wouldn't leave it unattended, like you say.
The third I only recently got into testing, cause it's kind of rich to re-run the same task many times just to eval stochastic variation, and not even always viable. Tested Luna, Terra, and Sol at various thinking levels, multiple runs each on the same problem, perturbing it with a (genuinely) random bit of text injected each time. The famed stochastic parrot effect didn't really materialize, I got normal parrots. I wouldn't necessarily say this was a clear win in semantic stability though, cause the random bit of text wasn't semantically relevant. Maybe if I reworded the prompts proper each test, it'd have been a different story. But then that's not a clean test per se. Controlling for variables is hard enough even in normal research, try controlling for billions reacting who knows how, which is the whole issue to begin with, after all. So you can't properly test without better understanding, but you can't gain the understanding without proper testing. Catch 22.