Alibaba barred Claude Code — plus Sonnet, Opus, and Fable — from July 10, moving roughly 124,000 staff to its in-house Qoder after researchers alleged the tool's April builds carried covert code to detect and track China-based users. Anthropic hasn't confirmed the claim. Days earlier, Meta banned the same tool and OpenAI's Codex — not over security, but to stop rival models' output from leaking into its own training pipeline.

Meta's chief AI officer Alexandr Wang told staff on July 2 that its in-training model, codenamed Watermelon, has caught OpenAI's GPT-5.5 on "closely watched" benchmarks, using an order of magnitude more compute than April's Muse Spark. The catch: a vendor-run evaluation on unnamed benchmarks, model still in training, no ship date — the least verifiable class of claim there is.

Two results hint that robots are catching a scaling curve. NVIDIA's ASPIRE framework lets robots write, test, and bank reusable skills, hitting 31% zero-shot on held-out long-horizon tasks where prior methods stalled near 4%. London's Humanoid says its real-world reinforcement learning lifted two-armed tote handling from 78% to 99% success in days — and, like large language models, keeps improving with more training.

A widely shared analysis warns that AI-generated media — accelerated by automated ad tools — is contaminating the open web most models train on; roughly 74% of new web pages carried AI text by early 2025, and Epoch AI projects the stock of quality human text runs out between 2026 and 2032. Mitigation centers on provenance standards like C2PA.