🤖 Anthropic reveals 80% of its code is AI-written 🤖
Anthropic confronts the accelerating speed of AI systems writing their own code.

Anthropic published a major report titled "When AI Builds Itself," detailing the reality of recursive self-improvement (RSI) in frontier AI systems. The company disclosed that over 80% of its merged production code is now authored by Claude, with engineers pushing 8x more code per day in Q2 2026 compared to 2024. This matters because AI development is outpacing human oversight capacity. Anthropic called for coordinated industry action, stating it would slow or pause frontier training if peer labs committed to the same. OpenAI echoed similar concerns in its own "Democratic Governance of Frontier AI" blueprint. The report marks one of the most candid public acknowledgements of RSI risks by a leading lab, setting the stage for urgent policy discussions.

🤖 Microsoft ships Scout, its always-on AI agent 🤖
Build 2026's headline launch is an autonomous agent that lives inside Teams.

At Build 2026, Microsoft unveiled Scout, an always-on AI assistant built on the OpenClaw framework that can schedule meetings, prepare documents, and complete tasks across the Microsoft 365 ecosystem — all without being explicitly prompted. Running inside Teams, Scout is designed to take proactive actions on behalf of users, positioning Windows and Microsoft 365 as the central control layer for AI agents. The rollout moves Microsoft well beyond chatbot-style interactions toward persistent, autonomous workplace assistants. Deep integration across its productivity suite gives it a unique distribution advantage, putting it directly against competing agent efforts from Google and OpenAI.

⚡ Nvidia ships open 550B Nemotron 3 Ultra ⚡
Nvidia's open 550B reasoning model rolls out at 5x the speed of competitors.

Nvidia launched Nemotron 3 Ultra, a fully open 550-billion-parameter reasoning model using a mixture-of-experts architecture with 55B active parameters. The model runs 5x faster and up to 30% cheaper than comparable alternatives, purpose-built for long-running agentic workloads. On the Artificial Analysis Intelligence Index, it scores 48, far ahead of the next best open model Gemma 4 31B at 39. Early testing shows over 300 tokens per second on a pre-release Deep Infra endpoint, with NVFP4 quantization planned for even higher throughput. This pricing and speed edge could reshape open-model adoption. The release reinforces Nvidia's strategy of pairing open-weight models with its hardware ecosystem.

INTERESTING TO KNOW

🧠 ChatGPT Dreaming doubles recall for Plus users 🧠

OpenAI launched "Dreaming," a background memory synthesis system for ChatGPT that converts past conversations into a continuously updated, category-sorted user profile covering topics like travel, hobbies, and work preferences. Internal testing showed factual recall jumping from 41.5% to 82.8%, while preference-following climbed from 31.4% to 71.3%. The system processes conversations asynchronously, building richer context over longer interaction horizons rather than relying on real-time extraction. This capability jump fundamentally changes how ChatGPT handles personalization. The rollout has begun for Plus and Pro subscribers in the US, with broader global availability planned for later.

🚀 MiniMax M3 tops GPT-5.5 with 1M-token context 🚀

Chinese AI company MiniMax released M3, an open-weights model supporting a 1-million-token context window that it claims outperforms GPT-5.5 and Gemini 3.1 Pro on coding benchmarks while approaching Anthropic's Opus 4.7. The model accepts image and video input and can operate a desktop computer using a novel attention architecture. MiniMax has committed to releasing full model weights and a technical report within 10 days, which would make it one of the most capable openly available models. The release intensifies competitive pressure from Chinese labs following recent advances from DeepSeek.