AI & design briefing for Tuesday, July 28, 2026

tldr

• Moonshot released Kimi K3 (2.8T params) as the largest open-weight model ever, while SSI secured a $5B NVIDIA deal and the Open Secure AI Alliance launched without OpenAI, Google, or Anthropic.
• Pydantic's "human-in-the-loop is tired" essay explores the psychology of AI coding, Microsoft's MAI-Cyber-1-Flash redefines security economics, and shared Claude chats were exposed in Google search results.
• Kimi K3 weights, FlashKDA attention kernel, OptMem infinite memory, MAI-Cyber-1-Flash, and Meta AI in Threads DMs lead this week's launches.

Top Stories

• Moonshot drops Kimi K3 weights. The 2.8T-parameter open-weight model is the largest ever released, competing with Claude Fable 5 and GPT-5.6 on benchmarks while carrying a restrictive license for companies over $20M revenue. Moonshot also open-sourced FlashKDA attention kernels and AgentEnv training infrastructure. ^[1]
• SSI breaks silence with $5B NVIDIA deal. Ilya Sutskever's Safe Superintelligence emerged from two years of quiet research, securing a multi-billion NVIDIA investment and access to the Vera Rubin GPU platform for a 10x compute upgrade within 12 months. ^[2]
• NVIDIA launches Open Secure AI Alliance. A coalition of 36 organizations (Microsoft, IBM, Cisco, CrowdStrike, Hugging Face, Palantir) to build open security tools for AI agents. The notable absences: OpenAI, Google, Anthropic, Amazon, and Meta. ^[3]
• Anthropic clarifies open-weights stance. Dario Amodei published that Anthropic has never advocated for a ban on open-weights models, instead pushing for chip controls and distillation crackdowns, while declining to sign the Nvidia-led open-weights letter. ^[4]

Deeper Cuts

• The human-in-the-loop is tired. Pydantic's essay on reward functions, dopamine, and what it feels like when the code starts writing itself. A meditation on the psychological shift from writing code to guiding code. ^[5]
• Claude shared chats leaked in Google. A Reddit user discovered shared Claude conversations and Artifacts appearing in search results, exposing legal questions, personal info, and vibe-coded app data. ^[6]
• Microsoft launches MAI-Cyber-1-Flash. A cybersecurity model hitting 96% on CyberGym benchmark at half the cost of top rivals, integrated into the MDASH agent system for automated vulnerability hunting. ^[7]
• OpenAI study: AI is dissolving job titles. 43.5% of occupation-specific ChatGPT messages involved work normally associated with a different occupation. Marketers analyze data, operators write code, managers moonlight as researchers. ^[8]
• Cursor's planner-worker swarm pattern. A frontier model plans the work, cheaper models execute subtasks, and a shared field guide prevents repeated mistakes. The SQLite-in-Rust experiment showed significant cost savings. ^[9]
• Drawing the Mona Lisa with four frontier models. A side-by-side tracking of every stroke, dollar, and output as GPT-5.6, Claude, Gemini, and Grok tried to draw the Mona Lisa with colored pencils. ^[10]
• EU AI Omnibus enters force. Extending compliance timelines and expanding regulatory sandboxes across the bloc. ^[11]
• METR's expenditure-horizon metric. A new framework for judging when AI agents actually become cheaper than humans, factoring in supervision and retry costs. ^[12]

Tools and Launches

• Kimi K3. 2.8T-parameter open-weight model, native multimodal, 1M token context. ^[1]
• FlashKDA. Open-source attention kernel cutting prefill time by up to 2.2x. ^[13]
• OptMem. Plug-and-play infinite memory prompt for any AI agent, using compressed summary trees. ^[14]
• MAI-Cyber-1-Flash. Microsoft's cybersecurity AI for automated bug hunting at half the cost. ^[7]
• AgentEnv. The distributed agent training system behind Kimi K3, open-sourced. ^[15]
• Meta AI in Threads DMs. Chat with Meta AI about posts, images, and links without leaving the app. ^[16]
• jcodemunch MCP. Open-source MCP server cutting AI coding token costs by 95%. ^[17]
• FLUX 3. Black Forest Labs' multimodal AI with 20-second generations. ^[18]
• Buzz. Open-source workspace where AI agents join team chats. ^[19]

Prompt of the day