They are, the model has no inherent knowledge about its confidence levels, it just adds plausible-sounding numbers. Obviously they _can_ be plausible, but trusting these is just another level up from trusting the original output.
I read a comment here a few weeks back that LLMs always hallucinate, but we sometimes get lucky when the hallucinations match up with reality. I've been thinking about that a lot lately.
> the model has no inherent knowledge about its confidence levels
Kind of. See e.g. https://openreview.net/forum?id=mbu8EEnp3a, but I think it was established already a year ago that LLMs tend to have identifiable internal confidence signal; the challenge around the time of DeepSeek-R1 release was to, through training, connect that signal to tool use activation, so it does a search if it "feels unsure".
Wow, that's a really interesting paper. That's the kind of thing that makes me feel there's a lot more research to be done "around" LLMs and how they work, and that there's still a fair bit of improvement to be found.
In science, before LLMs, there's this saying: all models are wrong, some are useful. We model, say, gravity as 9.8m/s² on Earth, knowing full well that it doesn't hold true across the universe, and we're able to build things on top of that foundation. Whether that foundation is made of bricks, or is made of sand, for LLMs, is for us to decide.
G, the gravitational constant is (as far as we know) universal. I don't think this is what they meant, but the use of "across the universe" in the parent comment is confusing.
g, the net acceleration from gravity and the Earth's rotation is what is 9.8m/s² at the surface, on average. It varies slightly with location and altitude (less than 1% for anywhere on the surface IIRC), so "it's 9.8 everywhere" is the model that's wrong but good enough a lot of the time.
It doesn't even hold true on Earth! Nevermind other planets being of different sizes making that number change, that equation doesn't account for the atmosphere and air resistance from that. If we drop a feather that isn't crumpled up, it'll float down gently at anything but 9.8m/s². In sports, air resistance of different balls is enough that how fast something drops is also not exactly 9.8m/s², which is why peak athlete skills often don't transfer between sports. So, as a model, when we ignore air resistance it's good enough, a lot of the time, but sometimes it's not a good model because we do need to care about air resistance.
Gravity isn't 9.8m/s/s across the universe. If you're at higher or lower elevations (or outside the Earth's gravitational pull entirely), the acceleration will be different.
Their point was the 9.8 model is good enough for most things on Earth, the model doesn't need to be perfect across the universe to be useful.
