It might be vastly easier to make a conscious, feeling, qualia-laden moral patient in a GPU than to make a machine-god in a GPU. ...and given the ease with which nature produces fleshy versions of the first and not the second, this seems like a pretty likely hypothesis?

Why even complain about matmuls being non-commutative, when they're non-associative on GPUs?

So you could make a dashboard showing the total FLOPS brought into the world per month, with the % from human babies and the % from GPUs. Do it right, it could maybe even get as popular as That One METR Chart.

the feeling of reading an ML paper, seeing some particular method, and thinking "Well this feels like an awful hack... unless it's some solid, well-established technique from a domain I don't know, in which case, never mind."

What products are there that used the original GPT-3 base model API and are still around / have not vanished?

What if there's a alternate history of LLMs, where base models were trained into R1 / o1 style entities before RLHF was even invented? Would everyone's feelings about capabilities and alignment be different or the same? Would they be correct?

If you could do literally infinite GRPO rollouts on a problem, what would be the thing that your LLM learns?

I did get a weird feeling when I actually heard a commercial use the breathless "It's not an X. It's a Y" format.

If there's a reasonable chance deep learning, but not LLMs, will be the basis of superintelligence should one do mechanistic interpretability on weather sim transformers / vision transformers / etc as practice?

More evidence contra "this is the worst it will ever be" for AI. Nope, it can just get worse on hard-to-track metrics!