@mattsheffield@mastodon.social
2026-10-07 08:49 UTC
Someone made an "AI torture chamber" which is using research that claims language models have mathematical vectors which correspond to human emotions.
The site largely disproves it w/o authors realizing.
I haven't decided whether to include this in the 80-page AI cognition paper I'm working on.
Probably will, but it's such a beast already, yikes.
Worth checking the site out:
https://wirehead.agency/
You can experiment with it:
https://wirehead.agency/live.html
Replies (2)
-
@HistoPol@mastodon.social 2026-10-07 08:58
@mattsheffield@mastodon.social #AICognition "The site largely disproves it w/o authors realizing." I'm curious: why do you think that is the case?
-
@mattsheffield@mastodon.social 2026-10-07 09:03
The main problem with claiming there are "pain," "sadness," etc. vectors is that these are human constructs for math matrices, not objective identifications. What they really are is control methods. A car steering wheel does not literally steer the car. The steering rods do.