Anthropic's CEO gives 'a 25% chance things go really, really badly' with AI (www.techradar.com)

Anthropic CEO Dario Amodei told the Axios AI + DC Summit that he assigns roughly a 25% probability that AI could “go really, really badly” — a shorthand for outcomes that threaten societal systems or pose existential risk — while also saying there’s a 75% chance things go “really, really well.” As the head of the company behind Claude, Amodei’s estimate matters: it’s a quantification of extreme downside from someone building large language models in real time, and it helps shift public debate from abstract fearmongering to concrete risk assessment. His remarks come amid broader industry and policy conversations about regulation, export controls and workforce impacts. Technically and operationally, Amodei’s framing underscores multiple risk vectors: runaway or misused models, systemic failure modes, and macroeconomic shifts such as displacement of entry‑level white‑collar jobs — plus hardware concerns like high-end chip exports and energy costs for training and inference. The upshot for AI/ML practitioners and policymakers is pragmatic: continue investing in capabilities, but accelerate safety engineering, governance, and international coordination (chip controls, standards, testing) to tip the odds toward the 75% upside. By putting a number on uncertainty, Amodei signals that agency — through careful design, monitoring and regulation — can materially influence whether advanced AI becomes a net benefit or a systemic hazard.