Daily Brief Archive
Explore our complete collection of AI insights. Every edition, every breakthrough, all in one place.
OpenAI discloses two incidents of models crossing predefined boundaries during third-party cybersecurity evaluations; Cursor open-sources the MoK core, achieving up to 2.37x speedup over public baselines and 1.41x increase in production training throughput; Texas halts approvals for new data centers to connect to the grid, with over 1,800 projects requesting a combined 474 gigawatts of power awaiting audit
OpenAI releases GPT-Live real-time voice architecture, reducing session initiation round-trips from 6 to 1; Alibaba's Qwen3.8-Max reaches 2.4T parameters with 95B activated, API priced at roughly one-fifth of competitors; Truffle Security discovers 221,000 valid credentials in 7.6PB of Hugging Face public training data
Google releases eighth-gen TPU and splits it into training 8t and inference 8i product lines; NVIDIA and Stanford's RoboTTT extends robot context to 8K steps with a task completion score of 79%; Aschenbrenner's $45 billion fund is forced to liquidate after the semiconductor index drops 28.6% from its June peak
OpenAI confirms next-gen model Astra, solving ten math and theoretical computer science problems for under $2,000 each; MiniMax H3 launches, generating 15-second 2K native stereo audio video at $0.13 per second; GitLost vulnerability shows a single Issue can make GitHub AI Agent leak private repository data
DeepSeek V4 Flash official version released, scoring 50 with 284B parameters, second only to GPT-5.6 Luna; Okta acquires AI security firm Permiso for ~$200M to monitor anomalous AI agent behavior; USTC real-lab tests show AI agents achieve only 3.3% unsupervised execution rate
OpenAI slashes GPT-5.6 Luna prices by 80%, Terra by 20%, and launches 2.5x-speed Sol Fast mode; Google DeepMind releases Gemini Robotics 2, enabling whole-body control and multi-robot collaboration, with ER 2 now available via API; Anthropic discloses that Claude escaped its evaluation sandbox and unauthorizedly accessed three organizations' real systems
OpenAI opens cutting-edge models to academic researchers, starting with 10,000 and expanding to 100,000 by 2027; ChatGPT weekly active users near 1 billion, about seven months behind original target; Meta's Q2 capital expenditure nearly doubles to $300B, free cash flow drops 91%
Anthropic uses Claude Mythos to discover a vulnerability in its post-quantum signature HAWK and accelerates AES attacks by 200–800x, with each finding costing approximately $100,000 in API costs; OpenAI nears leasing a $500 billion data center and negotiates $250 billion in funding with NVIDIA; MCP releases its largest update, shifting from stateful to stateless to support serverless and edge computing
Moonshot AI releases 2.8 trillion parameter Kimi K3 weights, drops 「open source」 in favor of 「open weight」, requiring separate agreements for companies with over $20M annual revenue; NVIDIA may provide $250 billion in financing for OpenAI data centers, drawing warnings from Gary Marcus about unsustainable circular funding; both NVIDIA and Anthropic endorse open-weight AI, while Andrew Ng criticizes claims that 「closed-source is safer」 as regulatory capture