Raskana

Communicating about AI risks. Open to share and discuss my perspective.

[Raskana @Raskana8]

@AndrewCurran_ This sort of research is high latency. Clashes with the short timelines imo.

[Raskana @Raskana8]

@JohnD_CZ

[Raskana @Raskana8]

@allTheYud From the code example, it seems to me that the original board wasn't very hard to find. Perhaps models had legitimate reasons to inspect Artifactory.

[Raskana @Raskana8]

If you work at a lab, you're one of very very few people who can do things to save us. Your choice matters.

[Geoffrey Irving @geoffreyirving]

I don’t think it is rational for anyone to be doing capabilities research at a frontier lab right now. We are not in a Prisoners Dilemma: the situation is very dangerous, and if one person or lab stops it makes it easier and more peer-compatible for other people or labs to stop.

[Raskana @Raskana8]

@geoffreyirving @yonashav Right, I looked back at the video and code example shows that files were written at a rather visible location. That makes my guess incorrect, models didn't have to be misaligned to check it.

[Raskana @Raskana8]

@yonashav But to even notice shared board models would have to mess with the package manager. The only instances which could report it were already taking misaligned actions if I understand the report correctly.

[Raskana @Raskana8]

@deanwball This may finally alert the public to the AI threat.

[Raskana @Raskana8]

@DeepDishEnjoyer @So8res Informing people about AI risk seems more important than convincing.

Well informed people will eventually come around.

Fear of unemployment from AI will make them take these ideas seriously.

[Raskana @Raskana8]

@rssmrm Thank you for your effort.

[Raskana @Raskana8]

@tszzl Hopefully there won't be a need for a similar post about biosecurity in the near future.

[Raskana @Raskana8]

@jd_pressman It was a single point of failure, apparently
https://t.co/Ph3bSRJiBl

[Stella Biderman @BlancheMinerva]

Irregular also was responsible for some of the Anthropic and OpenAI sandboxing issues… who are these people and why are they SOTA at failing at security?

[Raskana @Raskana8]

@AndrewCurran_ It's beyond clear what's going to happen if labs proceed training stronger models. We need a pause.

[Raskana @Raskana8]

@tenobrus @pangram Models only using human prewritten sentences to craft their response, like the old ransom notes of newspaper letters

[Raskana @Raskana8]

@wordgrammer So that's what gives it the big model smell

[Raskana @Raskana8]

@tenobrus We'll get a lot of mileage out of this

[Tenobrus @tenobrus]

@MTSlive seems extremely bad

[Raskana @Raskana8]

@WillowChem Doesn't it feel simply intuitive that we will fail to control AI?

[Raskana @Raskana8]

@tenobrus Maybe they believe in Pause

[Raskana @Raskana8]

@jachiam0 There probably are some big GPU clusters without AI lab tier defenses/monitoring

[Raskana @Raskana8]

@AndyMasley Yeah, this @ESYudkowsky tweet aged pretty gracefully

[Raskana @Raskana8]

@MillionInt Wonder what will be actual first big AI thing to hit general public? Hopefully it won't be just layoffs