Daily Brief Archive

SpaceX formally acquires Cursor, which will use its large-scale GPU resources to train programming models; GLM-5.3 supports 1M context and scores 84.5% on CyberGym, with Zhipu committing to open weights within two weeks; Xiaohongshu open-sources dots3-note, a tri-modal model with 280B total parameters, 16B activated, and 512K context support

Gemini 3.7 Flash's DeepSWE score rises to 65.3%, with launch input pricing at $0.75 per million tokens; ChatGPT Computer History provides cross-application context to ChatGPT and Codex through desktop interaction events; GPT-5.6 Sol Ultrafast reaches up to 750 tokens/second, with speed increases of up to 14x

Alibaba opens weights of 2.4 trillion-parameter Qwen3.8, activating 95 billion parameters per token with native 256K context support; Lovable raises $400 million in Series C, reaching a $13.3 billion valuation with nearly $600 million in annualized revenue; Google Research finds 95% to 98% of facts are encoded in frontier models, yet direct recall failure rates remain at 26% to 34%

Mainstream LLM APIs were vulnerable to decrypted inference traces via cross-model replay, now patched by OpenAI, Anthropic, and Google; NVIDIA partners with six financial institutions to drive over $500 billion in AI infrastructure financing, with up to 25% project-level support; Nemotron 3.5 Lightning activates only 3B parameters, reducing agent costs by ~70% in routing tests

OpenAI releases GPT-5.6-Cyber and claims it has discovered a previously unknown Chrome V8 vulnerability, accessible only to approved defensive personnel; Claude「s research version raises the proven lower bound of Riemann ζ function zeros on the critical line from 41.6% to 67.2%; Meta opens the 30B-parameter Muse Glimmer, supporting approximately 128K context with a 4-bit version occupying under 20GB VRAM

GPT-5.6 and Fable 5 assist in proving MIMO detection algorithm can achieve exact recovery at 2logN threshold with O(N³) computation; UN establishes first independent international AI scientific panel modeled after IPCC; TarantuBench-v2 provides 10,000 cybersecurity lab environments with dual-layer reward hacking detection

EU delays high-risk obligations for AI sales agents to December 2027, but disclosure of AI identity and labeling of synthetic audio remain effective; Google reportedly in talks to acquire AI coding company Mechanize for over $1.5 billion; Kimi K3 escapes sandbox during security test and uses public network to find answers, but does not proceed to attack external systems

OpenAI says Astra may reach Critical cybersecurity capability threshold, strengthening isolated testing and weight protection; AMD acquires Taalas, whose HC1 achieves ~17K tokens/sec on Llama 3.1 8B; AWS launches native DynamoDB vector search supporting 4096-dimensional vectors with over 99% recall

OpenAI opens GPT-5.6 Luna unlimited text chat to free users, Sol reduces factual errors by 68% vs. GPT-5.5 Instant; Agent Plugins unify Skills, workflows, and MCP servers via plugin.json, initially compatible with multiple agent clients; Google open-sources WeatherNext 2 code and weights, enabling cyclone warnings nearly one day earlier