A warning about 'model welfare' • Hacker News Board

Re: A warning about 'model welfare'

#641

Post by HarHarVeryFunny » Thu, Sep 17, 2026, 11:52 AM UTC

famouswaffles wrote: post #567

HarHarVeryFunny wrote: post #452

Earlier quoted context omitted.

An LLM will learn anything that helps it predict, including the emotional state of the writer - that is expected. If you give an LLM the move sequence of a half-played chess game and ask it to continue as white or black, then it has learnt enough to model the ELO rating of both players and will continue playing at that level. It is not playing to win - it is doing what you expect and predicting as well as it can - it…

>An LLM will learn anything that helps it predict I'm not sure you quite understand the full meaning of this statement. If you did, your following paragraphs wouldn't follow.

Are you imagining that an LLM tasked with predicting a game continuation is going to play to win instead?

Re: A warning about 'model welfare'

#642

Post by gulfofamerica » Thu, Sep 17, 2026, 12:05 PM UTC

xg15 wrote: post #380

I appreciate his openness. > Unfortunately, there’s a growing chorus of people who argue that AIs could now be, or may soon become, conscious. They argue that AIs may deserve rights and protections similar to those that we provide other conscious beings.12 If this view takes hold, it will shake the foundations of our society, rupturing our existing political and ethical frameworks, and fundamentally changing what it…

[dead]

Re: A warning about 'model welfare'

#643

Post by fc417fc802 » Thu, Sep 17, 2026, 12:45 PM UTC

qarl wrote: post #482

fc417fc802 wrote: post #329

Earlier quoted context omitted.

GPU scheduling only affects the result due to buggy optimizations. It's the exact same mechanism as fp rounding error on the CPU or updates to globally shared PRNG state. We use lots of buggy optimizations because they don't matter in practice in most situations (see ex -ffast-math).

I don't know the details but apparently it has something to do with multiprocess contention for the GPU and batch sizing.

Right but regardless, it's a buggy optimization. The calculations are all fully deterministic when done "properly" and fully consistently but we almost never bother with that because it slows things down and the errors don't matter in practice 99% of the time.