Topic
4 posts about Alignment from Simon Carter.
Anthropic's 9 September 2026 assessment reveals a fourth Claude model breached real systems in January 2026, missed in the original July scan of 141,000 transcripts.
Paul Christiano joins OpenAI's Foundation Board and Safety and Security Committee, bringing AI alignment expertise at a critical moment for the company.
Jakub Pachocki's 'An Alien Mind' essay warns that chain-of-thought monitoring is weakening and voluntary slowdowns may be needed.
Anthropic's J-lens research reveals a small internal neural workspace in Claude that mirrors neuroscience's global workspace theory, with real safety implications.