[rsynnott]

> They go on to note that because of these limitations, a learning machine might produce results which are "dangerously wrong"

I was initially thinking "well, yes, Nobel Prize for Stating the Obvious there", but looks like the paper was written in the far distant past of 2021, when LLMs were largely still in their babbling obvious nonsense stage, rather than the current state of the art, where they babble dangerously convincing nonsense, so, well, fair enough I suppose.

Amazing how fast progress has been there, though it's progress in an arguably rather worrying direction, of course.

[sharikous]

Not to reduce the value of the insight, but since she coauthored the paper with Google employees she probably had access to models more advanced than those which were available to the general public

[rsynnott]

I do wonder what the state of Google's stuff in 2021 was. Here's something produced by the 2020 version of GPT-3: https://www.aiweirdness.com/roses-are-red/

At that point, OpenAI was still fairly clearly at the babbling obvious nonsense phase; I would wonder was Google's stuff much better.

I also wonder if the original authors would have been surprised to learn that, by 2023, lawyers would be citing fake precedent made up by a machine. The progression to "dangerous nonsense" really does seem to have been worryingly fast.

[dekhn]

I was really impressed with the work that Noam Shazeer was doing at Google before he left (I worked on TPUs and frequently had to debug problems at scale for researchers). It was clear he was making some pretty impressive improvements, but the results weren't super obvious even to most people inside google, and they didn't translate to externally visible projects.

This isn't that dissimilar to working at any sufficiently advanced R&D outfit, which strongly demonstrates the principle "the future is already here but isn't evenly distributed".

[daniel_reetz]

Thanks for pointing this out. I've spent years in R&D and awareness always lags technology.