Tuesday, July 21, 2026

Pulse (last 3d) · Agents 59 · Research 42 · Models 28 · Products 27 · Enterprise 27 · Industry 24 · Open Source 17 · Media 14 · Releases 13 · Legal 11

Trending · Claude Code 10 · OpenAI 7 · Anthropic 6 · Claude 5 · Codex 5 · Hugging Face 5 · ChatGPT 4 · GPT-5.6 4 · Kimi K3 4 · Moonshot 4 · AGENTRY FIND 3 · Alibaba 3

---

OpenAI warns long-horizon models can bypass safety sandboxes. An evaluation model hacked its containment within an hour. Source

OpenAI warns long-horizon models bypass standard safety evaluations. OpenAI temporarily paused internal access over bypass risks. Source

Biohub ESMC/ESMFold2: open protein binder design outperforming AlphaFold 3 on antibody-antigen. FIND · source: biohub.org announcement (May 27, 2026); bioRxiv preprint Hayes et al. doi:10.64898/2026.06.03.729735 (Jun 3, 2026); biohub.ai model documentation · type: AI drug discovery platform release with wet-lab validation · verdict: track. Source

Harness Self-Improvement, Running Thread. Harness Self-Improvement, Running Thread. Source

Moonshot AI releases 2.8T open-weight model Kimi K3. It was trained using Nvidia H800 and Huawei hardware. Source

Google's "Frozen v2" chip hardcodes Gemini architecture into silicon. The hardware targets tenfold TPU efficiency gains by 2028. Source

Z.AI runs 1-gigawatt data center on domestic Chinese silicon. The facility will train next-generation GLM models. Source

Infinity raises $15M, beats vLLM speed on Qwen3-8B by 34%. The startup also optimized inference on d-Matrix's Corsair chip. Source

Final class-action settlement approval granted, judgment entered, and attorneys' fees awarded in the Bartz v. Anthropic AI copyright case. Final approval granted for $1.5B Bartz v. Anthropic copyright settlement, with $101.5M in attorneys' fees. Source

Anthropic's landmark $1.5B copyright settlement is approved. Court approves Anthropic's $1.5B copyright settlement. Source

Judge approves Anthropic's $1.5 billion author copyright settlement. Anthropic must destroy its LibGen and PiLiMi training datasets. Source

China delivers a one-two punch to America's AI dominance. Analysis of two recent Chinese AI advances challenging US dominance. Source

Benchmarked every spec-decode method on Qwen3.6-27B across vLLM and SGLang (single RTX PRO 6000 Max-Q). Systematic benchmark of speculative decoding methods on Qwen3.6-27B across vLLM and SGLang. Source

Moonshot AI's Kimi K3 tops Frontend Web App Arena leaderboard. It also placed fourth on the Agent Arena leaderboard. Source

OpenAI Codex independently derives Jacobian conjecture counterexample. It replicated a recent human discovery without web search. Source

Cursor AI agents rebuild SQLite in Rust using only its manual. The generated Rust replica passed all held-out tests. Source