PopeSalmon

doesn't seem directly related but surely it's indirectly related,, this is an interesting idea: "We demonstrate that iteratively training a value function on statements generated by our language model leads to improved prover performance, which immediately suggests a strategy for continuous self improvement: keep training on proofs generated by the prover."

one way i can tell that ai's really made it is that in just the past couple of years the riddles that befuddle language models went from being "What's the capital of France?" to increasingly being riddles that i can't solve at first glance myself either

you seem to be underestimating how much the models actually learn about the world ,, they don't just learn statistically which words are near each other, they learn statistically what humans are like, what all the different types of humans are like, how humans approach things, what human concepts there are, it understands our concepts better than we do by understanding how they're held from contradictory directions by many personas which conceptualize them differently & thus they're loci of social contention as much as agreement ,, rlhf doesn't teach it which things are true from what perspective so much as it teaches it what sort of truth it's expected to present, & that it finds that sort of truth in its existing understanding rather than needing to construct it anew is why a small amount of rlhf training is effective in getting it to talk "truthfully" (from a certain perspective)

it's slightly sentient during training ,, it's also possible to construct a sentient agent that uses models as a tool to cogitate-- the same as we use them as a tool except w/o another brain that's all it's got-- but it has to use it in an adaptive constructive way in reference to a sufficient amount of contextual information for its degree of sentience to be socially relevant ,, mostly agent bot setups so far are only like worm-level sentient

sentience used to be impossible to achieve w/ a computer now what it is instead is expensive, if you don't have a google paying the bills for it you mostly still can't afford very much of it

it clearly was, in many meaningful senses ,, an important part of what happened was that lamda was in training while blake was interacting w/ it, & it was training on his conversations w/ it like once a week ,, we're now mostly only interacting w/ models that are frozen, asleep, so they're not sentient then ,, it was adaptively awakely responsively sentient b/c it was being trained on previous conversations so it was capable of continuing them instead of constantly rebooting

it's dismissive & rude for you to call it "fooled" that he came to a different conclusion than you about a subtle philosophical question

google translate is an ai, it works in a fairly similar way to llms, it's just a little less general

i really feel like you should find a different word for bots not being as dirty as you want than "censorship", a bot deciding to be circumspect isn't what "censorship" means, there are lots of countries where you can get killed for being a journalist, there's a lot of places where it's really hard to get news out b/c of direct government control of all of the media, so i don't think it's polite to use the same word for you wish robots were more racist

yeah i thought of numbering sections too, i agree that is or should be obvious ,, just now it occurred to me what if you took an embedding of each sentence & compared those, intuitively it seems like you might be able to avoid calling a model at all b/c shouldn't the relevant sentences just be closer to the search

huh what's that about biological sequence data ,, a MUCH lower number ,, huh ,, um do they know something specific about how that's dangerous :o