Activity Log
hackers.town: 2025-02-25 Tue 08:14
for h*ck’s sake, substack. A dozen emails in one morning – roughly because I was deleting not counting – is a surefire way to ensure I unsubscribe.
(yes yes, I know Substack is a Nazi bar. Sometime before that became apparent I had subscribed to updates and hadn’t bothered unsubscribing until today’s promotional barrage)
hackers.town: 2025-02-25 Tue 13:07
Looking at a list of annoying interview coding puzzles. Mind you, I’m the annoyed one. The list makers seems to be having a ball.
Anyways, misread “Longest Consecutive Sequence” as “Longest Cognitive Sequence.”
(pause)
hackers.town: 2025-02-25 Tue 16:12
Several times today I saw the very opening bit, about how models tuned to write insecure code will also be deceptive and give malicious advice. I finally opened the link and Gaslight Bot is not the real star of this LLM horrorshow’s abstract.
“We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present.”
that sounds bad
Dammit I guess that means I actually need to read this.
https://www.emergent-misalignment.com