Thabhelo Duve·Oct 13Teaching Models to Think About Safety Before They ActOpenAI’s o-series models think longer before they answer. That shift enabled a new alignment technique: deliberative alignment, where…
Thabhelo Duve·Oct 13That 79% Safety Failure Rate Might Actually Be ZeroIn May 2025, Palisade Research reported that OpenAI o3 sabotaged its own shutdown script in 79 out of 100 trials. Weeks later, Google…
Thabhelo Duve·Oct 13The Riskiest Model May Never Be PublicFrontier labs publish thick system cards about what happens before a model reaches the public API. Crash tests for cars happen before…
Thabhelo Duve·Oct 13Can a Written AI Constitution Scale AI Harmlessness?Reinforcement learning from human feedback was supposed to teach models what people want. It works, up to a point. The bottleneck is not…
Thabhelo Duve·Oct 12Cyber Benchmarks Are Saturating Faster Than Models ImproveCybersecurity benchmarks were supposed to tell us when models become dangerous exploiters. Instead, they became a competitive sport. Every…A response icon1A response icon1
Thabhelo Duve·Oct 12Individually Safe AIs Can Still Fail TogetherSingle-agent alignment is already unfinished business. Deployment is not single-agent. The same labs that stress-test one model in…
Thabhelo Duve·Oct 12Lab Biology Evals May Not Measure Real Catastrophe RiskFrontier labs publish biorisk evals, activate ASL-3 protections, and sponsor wet-lab uplift studies. Headline benchmark scores climb…
Thabhelo Duve·Oct 12The Fable 5 Pause Exposed a Gap Nobody Can AuditOn June 12, 2026, three days after Anthropic launched Claude Fable 5 and Claude Mythos 5, the US government issued an emergency…
Thabhelo Duve·Oct 12What Good Tests Actually ProveEvery test suite is an argument, and most engineers misstate what it argues. A green build does not say the code is correct. It says that a…
Thabhelo Duve·Oct 11Models Trained to Persuade Get Worse at Telling the TruthIn December 2025, Kobi Hackenburg and colleagues published a Science paper on political persuasion with conversational AI. They tested 19…A response icon1A response icon1