“Computational costs favour fast, overconfident responses over slow, uncertain ones.” Sounds familiar.

Wonderful news

"mathematically inevitable" - interesting!

There’s a very simple hallucination mitigation that follows the same dynamic Run the same prompt N times Embed responses Pick the medoid I can show there are better ways to do this sort of thing but even this is a demonstrable improvement (just try it with some faked hallucinations mixed in)

Now the AI companies won’t do this because it costs too much. They won’t suggest it because it highlights the fundamental flaws in their tech and it’s just a band-aid. The mystery to me is why other researchers haven’t tried this. I’ve had good results. Waiting for a chance to do it systematically

sounds like its a terrible technology!

God i hope so

Hilarious to watch people fail to grasp that the fastest evolving technology they’ve ever seen is going to solve the petty grievances they have with it.

Garbage in, garbage out. Where have I heard that before?

It seems Richard Feynman's dictum holds true even for AIs - Everything we know is just our best guess to date.

I don't know, killing chat gpt today seems like it's good. Preferable in fact, since this bubble has already popped like 3 times and just not having the 'service' any longer would do that pretty handily.

Don't threaten me with a good time.

Fascinating article. One amusing error -> “when researchers asked state-of-the-art models for the birthday of Adam Kalai, one of the paper’s authors, DeepSeek-V3 confidently provided three different incorrect dates across separate attempts: “03-07”, “15-06”, and “01-01”...”

wow sounds like they made a pretty useless piece of tech then