A comprehensive assessment released today by the Oversight Board reveals that frontier models from OpenAI, Google, Meta, and Anthropic are significantly more likely to refuse requests to criticize restrictive political regimes than democratic ones. The study suggests that LLM users may be experiencing free speech infringements by proxy, as models often default to the laws and customs of repressive jurisdictions in their refusal patterns.
A 'summer slowdown' has been replaced by a fierce price war. With enterprise AI costs reaching millions per month, major U.S. labs are pivoting away from raw intelligence benchmarks to focus on 'tokenmaxxing' efficiency. This shift is largely driven by the rapid adoption of Chinese open-weight models like Alibaba’s Qwen and Z.ai’s GLM, which are now capturing significant market share in Silicon Valley due to their cost-effectiveness.
South Korea today committed $880 billion to AI development over the next decade, representing the largest single-nation investment to date aimed at securing domestic AI sovereignty.
In the Future of Life Institute's 2026 AI Safety Index released this week, which lab received the highest grade (a C+)?