LESSWRONG
LW

345
Cosin V
2020
Message
Dialogue
Subscribe

Posts

Sorted by New

Wikitag Contributions

Comments

Sorted by
Newest
No wikitag contributions to display.
The case for more ambitious language model evals
Cosin V1y30

The reason truesight works (more than one might naively expect) is probably mostly that there's mountains of evidence everywhere (compared to naively expected)

 

Yes, long before LLMs existed, there were some "detective" sites that were scary good at inferring all sorts of stuff, from demographics, ethnicity, to financial status of reddit accounts, based on which subreddits they were on, where and (more importantly) what they posted

Humans are leaky

Reply
The case for more ambitious language model evals
Cosin V1y10

I googled and couldn't find any info

Reply
No posts to display.