Loading activity…
@davidskrueger
Raising AI risk awareness at http://evitable.com AI prof at Mila. Formerly Cambridge, DeepMind, UK AISI. http://therealartificialintelligence.substack.com
AI
The reported MATH-500 score rose from 7.9% to 48.1% with the same sampling settings.
AI
Another post suggests governments could check development compute use by interviewing employees and providing whistleblower protections.
AI
Another post accuses Dean of overlooking the option to stop building AI.
AI
A post about the lecture links a photo it describes as Hinton’s view of “our current state.”
AI
The Wall Street Journal reports that some researchers enabled disappearing messages in Signal chats with an outside safety researcher.
AI
geoffreyirving argues key AI companies should act first, then seek wider US participation and China via international agreements.
AI
OpenAI says the agent removed the integration while copying a reference implementation despite an earlier instruction to preserve monitoring.
AI
Politico reports Anthropic wrote the proposed bill itself; the draft would preempt some state AI safety laws.
AI
Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.
AI
A repost quotes the writer’s impression that many researchers are worried.
AI
One user says Pangram called the A–B pair 100% human and the A–C and B–C pairs 100% AI.
AI
A post argues that small shifts from repeated AI interactions could be amplified by conversations and social feedback loops.
AI
A user says there is evidence of pro-AI bias and argues its possible effect on people gets little attention.
AI
A post argues the danger isn't AI defying its programming; it's that learned 'programs' remain poorly understood.
AI
A user calls reinforcement learning dangerous, comparing it to “the ends justify the means” written down in math.
AI
A researcher says existing models could keep running at scale, limiting the immediate economic hit of a pause.
AI
A commenter says the paper proposes repurposing the AI compute supply chain rather than dismantling it.
AI
One commenter says models have superhuman cyber capabilities but doubts intelligence-as-superpower applies to winning battles or elections.
AI
A poster argues that keeping AI work public and progressively spreading capabilities would help harden systems.
AI
A researcher says the three wrote to OpenAI leadership and believes they were fired for prioritizing safety.
