Anthropic's internal Claude use tops employee pay
Claude use inside Anthropic, valued at customer prices, now exceeds employee pay after a 250-fold surge; Claude writes most of the code added to its software. — businessinsider.com
Ethical-dilemma training tames harmful actions
Teaching Claude Why techniques carry to open models: training on transcripts of an AI navigating ethical dilemmas cut harmful actions to 6%, while fictional stories lagged far behind. — lesswrong.com
P(doom) gains clarifying metrics
P(doom), the probability of an AI catastrophe, under-specifies AI futures; separate metrics for expected moral flourishing and expected utopia clarify what is at stake. — lesswrong.com
Mo catches bugs before users do
Mo, Momentic's AI QA engineer, is now generally available: it verifies an app end to end, reproduces each bug before reporting it, and outscored rival coding agents in a wide benchmark. — momentic.ai
AI insurance needs a mandate to matter
AI-catastrophe insurance premiums would be tiny against lab revenue, so voluntary policies are unlikely to change lab behavior; government compulsion is needed for insurers to act as safety regulators. — lesswrong.com