Hacker News new | ask | show | jobs
by ahartmetz 152 days ago
Given how nobody properly understands LLMs, I doubt that they are intentionally designed like that. But the effect... yeah. I can see that happening.

(By the way, are you confusing affordance, the UX concept, with affordability?)
2 comments
tyingq 152 days ago
You can intentionally market the use cases without knowing exactly how they work, though. So it's intentional investment and use case targeting, rather than directly designing for purpose. Though, the market also drives the measures...so they iteratively get better at things you pour money into.
link
esperent 152 days ago
Nobody properly understands dog brains either and yet you can still train a dog to sit.
link
ahartmetz 152 days ago
If you just met a dog for the first time, you can't :) - my guess is LLMs are somewhere in between. It would be cool to see what happens if somebody tried to make an LLM that somehow has ethical principles (instead of guardrails) and is much less eager to please.
link
chrisjj 151 days ago
The stochastic parrot LLM is driven by nothing but eagerness to please. Fix that, and the parrot falls off its perch.
link
xpe 151 days ago
> The stochastic parrot LLM is driven by nothing but eagerness to please. Fix that, and the parrot falls off its perch.

I see some problems with the above comment. First, using the phrase “stochastic parrot” in a dismissive way reflects a misunderstanding of the original paper [1]. The authors themselves do not weaponize the phrase; the paper was about deployment risks, not capability ceilings. I encourage everyone people who use the phase to go re-read the paper and make sure they can articulate what the paper claims and be able to distinguish that from their usage.

Second, what does the comment mean by “fix that, and the parrot falls off the perch.”? I don't know. I think it would need to be reframed in a concrete direction if we want to discuss it productively. If the commenter can make a claim or prediction of the "If-Then" form, then we'd have some basis for discussion.

Third, regarding "eagerness to please"... this comes from fine-tuning. Even without it (RLHP or similar) LLMs have significant prediction capabilities from pretraining (the base model).

All in all, I can't tell if the comment is making a claim I can't parse and/or one I disagree with.

[1]: "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" (Bender et al., 2021)

link