1. September 2026
  2. Why self-tuning research agents don't overfit (amazon.science)
    AI · · September 15, 2026 · Research · Agents · Evaluation
  3. Near-daily AI use among US adults doubled in six months (epoch.ai)
    AI · · September 15, 2026 · Adoption · Economics · Research
  4. Garry Tan wants an American distillation regime (techcrunch.com)
    AI · · September 14, 2026 · Open Models · Policy · Distillation
  5. A math benchmark of unsolved Erdős problems that models can't fake (epoch.ai)
    AI · · September 14, 2026 · Benchmarks · Epoch AI · Reasoning
  6. Teaching a model to refuse the harmful half of a topic, not the whole thing (huggingface.co)
    AI · · September 14, 2026 · Safety · Fine-tuning · Open Models
  7. How much compute the top five AI labs actually run (epoch.ai)
    AI · · September 13, 2026 · Epoch AI · Compute · Hardware
  8. Why most people still don't feel AI's impact (interconnects.ai)
    AI · · September 13, 2026 · Economics · AI Adoption
  9. Dario Amodei's case for slowing the pace of AI (darioamodei.com)
    AI · · September 13, 2026 · Anthropic · AI Safety · Policy
  10. OpenRouter's convenience comes with inconsistent results (simonwillison.net)
    AI · · September 13, 2026 · Engineering · LLM APIs
  11. Ai2's BenchMIRT digs into what an LLM benchmark actually measures (huggingface.co)
    AI · · September 12, 2026 · Evaluation · Allen Institute · Research