Notices by James Tindall (atomless@mastodon.social)

@danmcquillan @anildash I have misgivings with this critique.

To me LLM's unreasonableness comes not from any irreproducibility but from their reproducibility. The way they will repeat ad infinitum the statistically prevalent attitudes and biases locked into them via the spread of opinion within their pre-training corpus. Biases that, as I understand it, can never be fully overcome or patched with later training or guardrails, no matter how painstaking.

@danmcquillan @anildash
I suppose the "heat" parameter that determines "creativity, " --the randomness of which word/token is chosen from the ranked list of the next most likely-- could be set such that it's always the one at the top selected? That would still be Bullshit as a Service, I think we all agree? But is it anymore reasonable, just because it's the same old bullshit?

@danmcquillan @anildash That's a fair point. I suppose my worry is, if the perceived bar to reasonableness is set as low as reproducible output, then all of the harms would still remain.

@danmcquillan @anildash What I do wonder about the critique of reproducibility is whether, with the heat or creativity turned right down so that an LLM's output was entirely stable, how much of the illusion of intelligence evaporates along with the irreproducibility? Is the randomness not just required for the dance of variety but somehow essential to triggering the apparition of a ghost in the machine?

This piece by @emilygorcenski is incredibly good.

“Kurzweil’s The Singularity is Near is paternalistic and at times downright lecherous; paradise for me would mean being almost anywhere he’s not. The metaverse has two problems with its sales pitch: the first is that it’s useless; the second is that absolutely nobody wants Facebook to represent their version of forever. “
https://emilygorcenski.com/post/making-god/