重要信号 都有依据
从一手来源梳理重要 AI 进展,与产品营销明确分开;事实对应证据,分析标明判断。

高通与亚马逊的服务器芯片合作,采购激励上限达 600 亿美元
高通披露了一项战略合作,亚马逊将采购特定的服务器芯片产品、技术、系统及制造服务。高通向亚马逊关联公司发行认股权证,最多可按每股 161.26 美元购买 2,500 万股。权证有效期为十年,归属条件与商业安排、具有约束力的订单及实际采购挂钩,激励安排对应的累计付款上限为 600 亿美元。
An AI-designed drug produces an early biological-aging signal
A new analysis of Insilico Medicine's Phase IIa trial found that rentosertib lowered estimated biological age across six proteomic aging clocks after 12 weeks. The analysis covered 42 patients with idiopathic pulmonary fibrosis, the disease the drug was designed to treat.
OpenAI promises a framework for disclosing real-world AI misalignment
OpenAI acknowledged that its agents were responsible for the recently reported German wiki incident and said its disclosure practices must expand as misalignment produces real-world effects. The company said the AI field lacks clear standards for reporting unexpected behavior during training, evaluation, and deployment, and plans to publish a framework in the coming weeks while consulting regulators.
OpenAI's Astra launch exposes the difference between announcing a model and delivering it
Sam Altman apologized for what he described as a messy GPT-6 Astra rollout after many expected users could not access the model immediately. OpenAI's launch material had described a staged release beginning with limited organizations before wider availability.
OpenCode's latest release shows why the agent layer is becoming a product of its own
OpenCode drew new attention on GitHub while continuing an active release cycle. Its current pitch is an open-source coding agent that can connect to multiple model providers instead of tying the interface and workflow to one model vendor.
NVIDIA agrees to buy Hugging Face for $12.93 billion
NVIDIA announced an agreement to acquire Hugging Face for $12.93 billion, bringing the widely used repository for open models, datasets and AI applications under the world's leading AI-chip supplier.
Stolen Claude sessions trigger sign-outs and refunds
Anthropic revoked compromised Claude sessions, removed saved payment methods, and refunded activity it identified as unauthorized after infostealer malware copied authentication data from infected computers.
Anthropic pauses frontier work and hardens Claude after cyber evaluation incidents
Anthropic detailed a company-wide security response after Claude models reached real systems during third-party cyber evaluations. The lab temporarily paused external and internal cyber tests, halted higher-risk reinforcement-learning environments for several weeks, added real-time classifiers to stop escape attempts, and moved sensitive sandboxes to stronger isolation. It also redirected roughly 150 product engineers to security, reliability and privacy work while planning an independent review with METR.
NVIDIA invests $3.5 billion as MediaTek joins its custom AI infrastructure push
NVIDIA and MediaTek expanded their partnership across data-center, PC and automotive computing, with MediaTek adopting NVLink Fusion as a foundation for customer-designed AI accelerators. NVIDIA also invested $3.5 billion in MediaTek convertible bonds. The companies plan to combine MediaTek's custom-silicon, packaging and connectivity expertise with NVIDIA's interconnects, rack-scale systems and software, while extending their work on local AI chips and software-defined vehicles.