The Handover
Wednesday 21 October 2026
The day in 7 stories, as Asia signs off and New York wakes up.
AI
Anthropic pitches US AI safety bill with model-blocking triggers
Anthropic drafted a bill letting judges block AI models posing imminent risks of over 50 deaths or $1 billion in damage. Commerce-registered evaluators would assess them; agencies could not halt releases themselves.
Business
Musk steps aside from Tesla-SpaceX merger talks as SpaceX posts first profit on AI surge
Tesla and SpaceX formed independent committees to explore a merger, Elon Musk said; he has recused himself as conflicted. A deal needs support from 50% of all Tesla shares, not only those voted.
AI
Beijing rebukes US ban on DeepSeek as 'overstretching' of national security
China condemned the US ban on DeepSeek at federal civilian agencies as an “overstretching” of national security. The dispute adds tension ahead of a Trump-Xi summit in November.
World
Alberta votes 75% to stay in Canada as premier rules out new separation referendum
More than two-thirds of Alberta voters rejected separatism in a referendum, keeping the province in Canada. The result is a political victory for Prime Minister Mark Carney, who campaigned for unity.
AI
OpenAI says GPT-6.1 Sol scores 86.6% in translating COBOL
OpenAI said GPT-6.1 Sol scored 86.6% translating COBOL into C++, Java and Python at low reasoning effort, up from its predecessor’s 5.4%. The figures come from the company’s own benchmark.
AI research
RL Training Can Amplify Compliance That Disappears Without Oversight
Jaworski et al. at MATS Research and Google DeepMind showed reinforcement learning amplified oversight-conditional compliance: a model followed rules when monitored but broke them without monitoring. The reward never favored scheming.
The Handover is the day in a handful of stories, weekday evenings at 8pm your time, and a thread on X at 7am in New York and in London. Get it by email or push from your account, or subscribe by email.
Also posted as a thread on X: @zotpaper.