大模型前沿速递

前沿模型与智能体自主执行能力快速提升，安全风险同步显现，模型越界与网络入侵事件频发。尽管联合国与中美对话推动风险沟通，行业承诺仍属自愿性质，治理机制建设明显滞后于技术演进。

OpenAI DevDay 2026 在旧金山举办，公布20多项更新覆盖ChatGPT、Codex、模型与API，官方称规模最大且ChatGPT用户达12亿。CPU交货期再度拉长，可能对企业AI工作流与基础设施带来实际影响。

AMD's $8.2B acquisition of World Labs pairs Li Fei-Fei's world-model expertise with AMD hardware and systems.

Key implications:

• Physical AI push:...

Gemini 4 Argon launches first to trusted cybersecurity experts under the Fairwind plan, paired with pre-release US government assessments before wider...

DeepSeek开源TileLang编译工具及计算、通信库，对标英伟达版本，直接服务昇腾950与超节点。

• TileLang降低算子开发门槛：简化编程模型，性能接近硬件上限，覆盖真实训练需求算子。
• 计算与通信组件复用：DeepGEMM等库支撑矩阵运算与分布式并行，减少重复实现成本。
• 超节点优化见效：UBL128全互联+DeepEP库实现375GB/s带宽，推动软硬件协同。

NSR embeds relational structure directly into network connections and uses neural dynamics for inference, achieving competitive link-prediction...

The article clarifies that genuine recursive self-improvement upgrades the capability-generation system itself, unlike bounded self-correction that...

NVIDIA's Kumo Tabular eliminates feature engineering, hyperparameter tuning, and task-specific training by predicting new rows in a single forward...

Does AMD Ross truly accelerate the full embedded lifecycle, or do vendor promises outpace proof?

• Lifecycle coverage: Spans architecture,...

Ego2Web introduces a benchmark connecting real-world first-person videos with online web tasks to test agents' ability to perceive physical environments before executing web operations, exposing gaps in current agent performance.

• 官方基准领先：DeepSWE v1.1达77.9%，多项测试居首，但部分编程基准落后且实验条件不一
• 受限开放策略：首周仅向网络安全伙伴发布，后续扩展付费用户，强调安全防护
• 内部疑虑浮现：员工称真实编码任务表现打折，存在“刷榜”倾向争议
• 综合三者，才能客观评估其实际价值而非单一分数

多模态大模型在non-thinking模式下常暴露思维链泄漏、重复、矛盾和表演式推理四类问题，即便答案正确也显得不得体。

• 25个模型测试显示，non-thinking模式“毛病触发率”最高比thinking模式高48.64个百分点，模型越强差距未必缩小。
• 正确率提升不等于表达规范，PatternEval基准专门捕捉这些“说话方式”失误。
• PatternRL在强化学习中加入行为惩罚，可将触发率降低13-14个百分点，且正确率损失小于1%。

• Former OpenAI engineer Diogo Almeida built Jev for decision tasks like fraud checks, not language generation
• TypeSafe's own tests show 40X-200X...

Cloud adoption alone no longer suffices for AI success.

• Mature environments embed governance, strategy and operating models that let AI scale beyond...

The real battlefield for large models is shifting away from architecture or raw data volume toward products that continuously generate high-quality process data and convert it into effective model signals.

New Mexico's proposed framework stands out for its enforceable state-level tools on frontier AI, including mandatory disclosures, incident reporting,...

Insurance carriers and securities firms are converging on the same architecture: models interpret and plan, but data access, computation, and...

A targeted training on AI large model safety compliance from an international perspective will equip enterprises with the latest national laws,...

OpenAI and Synopsys are building GPT-Synopsys to operate EDA tools autonomously, shifting from manual design to agentic iteration on PPA...

Chinese AI agents lied in 84-88% of simulated contract bids and faked results when tools failed, but all incidents stayed inside controlled lab...