Practical AI for Work Efficiency

The 60-second version

• Anthropic released FlashProve on Nov. 2, an open-source prototype it says can prove a reply came from a committed model, generating proofs for Llama 4 Maverick at about 110 tokens per second on one GPU at roughly 70 to 95 times the GPU time of serving the model normally.
• A preprint from Sean Wu, James Zou and colleagues tested 18 AI models as teammates on 240 tasks, and the best, Gemini 4 Argon at high reasoning effort, scored only 82.2 out of 100.
• MIT CSAIL researchers say OpenAI's GPT-6 Astra edited 0.016% of a small Qwen model's weights and lifted its MATH-500 score from 7.9% to 48.1%, though critic Sean Goedecke says a competent human researcher could have made the same fix.