Post #1366603

2026-04-11 14:12 UTC

RE: https://mstdn.social/@bit101/116386237944316247 Which is even worse, right? Acknowledging that there&#39;s a problem but missing the actual problem and instead doubling down on it? LLMs are not a sampling technique, it&#39;s non-deterministic autocomplete. So the answer that comes back isn&#39;t within some understood range, you can ask the same question and be 5%, 50% or 100% incorrect each time it&#39;s asked. (And yes, you can do lots of things to mitigate all these issues, but again, wtf are we even doing at that point? These are solved problems! We have well-defined and efficient methods, along with supercomputers for laptops and instead we&#39;re blowing all that power on petascale ELIZA.)

Replies (2)

@bit101@mstdn.social 2026-04-11 14:30

@benfry Before the AI stuff was adopted full force, we had two sources of data which didn't align 100% and it was considered a big problem and a lot of effort was expended trying to bring them into agreement. Now we're OK with directionally correct, apparently.

Open ##1534252
@steltenpower@social.edu.nl 2026-04-11 16:01

@benfry When I was about 10, about the first thing I coded up was BINGO in MSX2-BASIC. Towards the end of the game it took really long to draw the next number and I quickly realised I should not check if a number was drawn before, but draw from those left instead. Then 4 decades happened, including some learning on efficient deterministic algorithms of which many were older than I was. Too many AI things seem like running data centres to play BINGO with way more numbers, the dumb way.

Open ##1534253