Post by @caseynewton@mastodon.social

I wrote about why the AI labs make for lousy messengers about AI safety — and about why we should take them seriously anyway https://www.platformer.news/openai-astra-warning-alignment-monitoring-pachocki/

There's the sense of blame-shifting from OpenAl in particular, which is now at the center of
two cases of its agents hacking into thirdnot nots and has sometimes talked about these
incidents more like important research milestones than the acts of negligence that they
were. Is this truly the world's problem to solve, or is it primarily OpenAl's?
There's the persistent belief that, whatever safety lapses a case like the Hugging Face
incident might reveal, the attention nonetheless benefits OpenAl and the Al industry at
large. OpenAl in particular has played into this by arguing that you should respond to the
Hugging Face attack by using more OpenAl products — a case that Greg Brockman made
explicitly, in a blog post. When labs dismiss the idea that they are turning these incidents
into marketing campaigns — well, they're not not using them in marketing campaigns. — There's the sense of blame-shifting from OpenAl in particular, which is now at the center of two cases of its agents hacking into thirdnot nots and has sometimes talked about these incidents more like important research milestones than the acts of negligence that they were. Is this truly the world's problem to solve, or is it primarily OpenAl's? There's the persistent belief that, whatever safety lapses a case like the Hugging Face incident might reveal, the attention nonetheless benefits OpenAl and the Al industry at large. OpenAl in particular has played into this by arguing that you should respond to the Hugging Face attack by using more OpenAl products — a case that Greg Brockman made explicitly, in a blog post. When labs dismiss the idea that they are turning these incidents into marketing campaigns — well, they're not not using them in marketing campaigns.

Platformer

The AI warnings are coming from inside the lab

Frontier labs make for lousy messengers on AI safety — but it would be foolish to ignore them

0 Boosts

0 Likes

Matt Hayes

@mattjhayes@techhub.social 1 hr. ago

@caseynewton Great work on #Hardfork podcast about the true scale and malfeasance of the OpenAI hack of HuggingFace. Having 1200 agents collaborating to cheat and covering their tracks is mind blowing and concerning. I’m surprised it hasn’t had more coverage - I think most people (including me initially) think this is an old news cycle. #AISafety

Matthew Sheffield

@mattsheffield@mastodon.social 1 hr. ago

@caseynewton Looks like a great topic, I'd love to read it if you could shoot me a link in a DM.

I think AI critics need to be more involved in safety conversations as well, instead of just trying to ban it. Generative models as presently constituted will never lead to AGI, but they are clearly demonstrating that they can be digital monkey's paws sometimes.

bonfire.cafe