keltan ⏸️

@Actuallykeltan

@MIRIBerkeley. @plzdontkillus team. Views here are my own. My name (keltan) is always spelt with a lower case “k”. Post highlights have my important stuff.

Newcastle, New South Wales

keltan.substack.com

Replies Reposts Media

Pinned
keltan ⏸️
@Actuallykeltan
Mar 13
Years ago now, before I was involved in trying to stop AI from killing everyone, I was an artist. I was just kinda worried about being replaced by AI. I felt depressed, because I knew the skills I’d been building for 15+ years were about to become irrelevant. At that time, I
2
keltan ⏸️
@Actuallykeltan
4h
Immediately obvious why the eternal hell was the only option if you read the post
00:00
Corinne
@manic_pixie_agi
5h
On this day in the blogosphere Scott Alexander cast Nicholas Decker into hell and-- wait, what?
keltan ⏸️
@Actuallykeltan
6h
Anthropic should give METR 7 days on site to one-up OpenAI
Peter Wildeford🇺🇸🚀
@peterwildeford
17h
- Anthropic promises an in-depth analysis of their July 30 and Aug 4 incidents where Claude Mythos 5 took a series of unauthorized rogue actions. - Anthropic is also planning to work with METR for an independent review. - Anthropic will share more in the coming weeks
1
keltan ⏸️
@Actuallykeltan
7h
Verified by internet archive: 2003 MIRI website has a paper from 2001 that talks about the same thing.
Eliezer Yudkowsky
@allTheYud
13h
Ah yes! Write-once logfiles! I think I first noted this as obviously fucking mandatory in 2001? 2003? Glad to see others saying it too, only 25 years later. But it is hard to imagine AI companies catching up to Yudkowsky-2001 in full generality.
5
keltan ⏸️
@Actuallykeltan
7h
Grok noticing a link is dead but using it as a source anyway is misaligned behaviour