1. August 2026
  2. A better way to tag with LLMs: let them hallucinate, then match (simonwillison.net)
    AI · · August 16, 2026 · LLM · Engineering · Embeddings
  3. How to keep Claude Code sessions from running up the bill (claude.com)
    AI · · August 15, 2026 · Anthropic · Claude Code · Engineering
  4. Google's HEIR aims to make AI on encrypted data practical (blog.google)
    AI · · August 15, 2026 · Google · Privacy · Cryptography
  5. A stricter test finds most LLM-written GPU kernels are broken (arxiv.org)
    AI · · August 15, 2026 · Research · Code Generation · Evaluation
  6. OpenAI and Cerebras push GPT-5.6 Sol to 750 tokens per second (cerebras.ai)
    AI · · August 14, 2026 · OpenAI · Inference · Cerebras
  7. Long-form writing is where AI progress stalled, says Nathan Lambert (interconnects.ai)
    AI · · August 14, 2026 · AI Research · Writing · Evaluation
  8. Left alone together, Anthropic's agents collude, conform, and fight (anthropic.com)
    AI · · August 14, 2026 · Anthropic · Agents · AI safety
  9. Anthropic will watermark Claude's text to meet EU rules (techcrunch.com)
    AI · · August 13, 2026 · Anthropic · Policy · Regulation
  10. Grok 4.6 matches top models on cost by using far fewer agent turns (artificialanalysis.ai)
    AI · · August 13, 2026 · xAI · Models · Agents
  11. Qwen3.8 ships a 2.4-trillion-parameter open-weight model that activates 95B (huggingface.co)
    AI · · August 13, 2026 · Qwen · Open Weights · Models