Daily Brief Archive

Netflix acquires Ben Affleck's AI startup InterPositive for $587 million focusing on VFX automation; SenseTime's SenseNova U1 Pro achieves native 8K output, surpassing GPT Image-2 in multimodal tasks; HKUST discovers MoE routing vulnerability, repeated token inputs increase time-to-first-token by up to 148%

OpenAI GPT-5.6 Sol sets new cybersecurity target record and becomes the first government-gated AI in the US; Kimi K3 surpasses Claude Fable 5 in internal writing benchmarks at much lower cost, with weights to be open-sourced on July 27; Google Gemini 3.5 Pro delayed due to performance shortfalls, stock drops 4%

Moonshot AI opens Kimi K3 with 2.8T parameters, tops frontend coding arena, estimated training cost only $15–25 million; Anthropic prepares IPO, valuation reaches $965 billion surpassing OpenAI; Stripe joins Advent in proposed $53 billion acquisition of PayPal

Thinking Machines releases first open source model Inkling, 975B MoE multimodal, 1M token context, Apache 2.0 licensed; OpenAI's GPT-Red adversarial training boosts GPT-5.6 resistance to prompt injection by 6x; Perplexity builds in-house sandbox platform SPACE handling 100% compute traffic with 5x faster tail latency

Meta's Muse Spark scores 30/30 in the Asian Physics Olympiad theory exam, tying top human competitors; OpenAI Codex surpasses 7 million weekly active users with over 150 updates in two months; Hugging Face CEO predicts open source models will account for 90% of token usage within 12 months

Anthropic研究发现Claude表达超3000种价值观、四大价值轴解释15%变化，且随模型版本与语言显著不同；Skyfall AI发布永不重置的MORPHEUS基准，证明现有LLM在漂移环境下无法持续学习；VentureBeat调查显示57%企业因AI代理缺乏业务上下文而获得错误答案，仅25%部署了受管控上下文层

StepFun's Step Edge on-device model achieves local latency as low as 0.1 seconds and ranks first in 29 metrics; HIT's TouchWorld tactile world model improves success rates by 16 percentage points across six real-world robot tasks; nearly 100 players in China's embodied data sector raised 4.47 billion yuan in one year, with Guanglun Intelligence capturing about 70%

Sugon completes China's first 100,000-card fully domestic AI supercluster 「Sugon 8000」, covering full-precision tasks from FP64 to INT8; Google open-sources Gemma 4, a 31B encoder-free on-device model rivaling closed-source frontiers; GPT-5.6 becomes Microsoft 365 Copilot’s preferred model, with coding task costs just one-fourth of Claude’s

GPT-5.6 Luna inference cost drops to 1/25 of previous generation; Sol Ultra solves 50-year math conjecture in under an hour using 64 sub-agents; Apple sues OpenAI for stealing trade secrets to develop AI hardware; MMFold achieves 68.6% success in antibody-antigen interface prediction, surpassing AlphaFold3