Post #1253953
2026-04-09 02:12 UTC
Replies (2)
-
@androcat@toot.cat 2026-04-09 07:01
@dngrs @budududuroiu @jenniferplusplus People keep getting tricked by framing. LLM companies frame what the models are doing as something else than what it is (autocomplete), and people whose competence is not in epistemic evaluation then look at the results based on the framing, rather than "this is autocomplete, it has to answer something, so it makes something up". And then other people take those soundbites and run with them. "Did you hear? Mr. Big Name said this stuff really works!"
-
@budududuroiu@hachyderm.io 2026-04-09 07:13
@dngrs Well, you're partly correct, partly wrong. Yes, pretrained transformers are, like all generative models, definitionally modelling a joint probability distribution, and autoregressively generating from that joint probability distribution. Those are the models you're referring to as autocomplete tools, hence why you had to use `[MASK]` with early transformers like BERT to get them to complete the "most probable token". Regardless, it doesn't matter what Anthropic did, if it allows for a massive reduction in cost of finding zero days, it's a problem. It doesn't have to be revolutionary, it doesn't have to be superintelligence, AGI, whatever woo-hoo flashy marketing terms. If a reduction in cost of computing protein folding happens, i.e. OpenFold implementation of AlphaFold, that wouldn't be revolutionary, but would still be dangerous, since you now potentially have lone actors being able to make prions at home (I'm using this as an absurd, but probable case). @jenniferplusplus