"Anything right-wing coded" Where are the days when rats said they were left-libertarians?

A major problem for offense defense balance is that we're going to hold defense to pretty high standards (understandably!), while offense faces no constraints.

I'm not sure I'm actually personally concerned about this LLM behaviour. I'm concerned that AGI, with a good enough world model, will realize that humans might switch it off and will therefore rationally use any opportunity to take over power. Having said that, amazing comms!

Am I right that gradual displacement hinges on AIs that collude?

Does anyone have a clue what DeepSeek is trying to say with this?

Last Tuesday, I visited the first academic AI Safety conference I've ever been to at KU Leuven. Both me and the organizer didn't really know other academic AI Safety conferences. Is this an unlikely omission in the AI Safety landscape? kuleuven.be/ethics-kuleuve…

"Give European settlers a stake in the future, then they'll respect our property rights" was not a winning strategy in 1492.

Our Otto Barten⏸ joined a panel with Alexander von Janovski (TÜV AI lab) and MEP Axel Voss MdEP. Good to be able to discuss the need to regulate AGI, and particularly the Conditional AI Safety Treaty! x.com/XRobservatory/…

I think most people who say "I would like a pause, but..." would not actually like a pause

I don't know whether UBI will happen eventually but it seems unlikely to happen when all the coders lose their jobs

Could you guarantee any safety level 1:S if you prompt a model S times, same distribution, and it doesn't do anything stupid?

Worthy successor is also human extinction. I like some of Daniel Faggella's work but people shouldn't engage with this.

I think embodied AGI would be pretty safe. Can't multiply, can't travel at the speed of light. Mostly local sensors and buttons to press. Any dangers are obvious and will be regulated.

Crazy that WIRED worries about digital rights but is happy to quote a pro-human extinction campaigner at length without asking any critical questions. wired.com/story/ai-risk-…