DeepSeek's Blueprint for Self-Evolving AI Agents
DeepSeek and PKU reveal the blueprint for self-evolving agents: reversible plugins that swap tools live without restarts.
DeepSeek and PKU reveal the blueprint for self-evolving agents: reversible plugins that swap tools live without restarts.
DeepSeek与北大发布80页论文,公开自进化Agent的工程底座:可逆插件架构让AI在运行时自行装卸工具、无需重启进程,Koishi四年4000+插件验证。
Stripe reportedly offers $10 billion for OpenRouter, gateway to hundreds of models. Value is migrating to the routing layer.
Stripe 被曝以 100 亿美元报价收购 OpenRouter,两个月前它刚以 13 亿美元估值完成 B 轮。模型路由层正从开发者小工具变成基础设施生意:AI 的价值正在从模型本身,转向连接、调度与结算模型的这一层。
Cloudflare 开源智能体运行时 @cloudflare/computer:用 isolate 加按需容器取代「一智能体一容器」,为数十亿并发智能体铺路。
An open-source agent runtime that replaces per-agent containers with fast, cost-efficient isolates.
CSIS 提出 Tokenpolitik 框架:中美 AI 竞争正从模型与芯片之争,升级为能源、算力、数据中心与模型构成的全球基础设施争夺。
CSIS proposes tokenpolitik: the US-China AI race is now a global fight over energy, chips, data centers and models.
扎克伯格发布6500字长文《未来属于所有人》:超级智能不该只属于少数机构,每个人、每个公司都该有自己的 AI。Meta 的「个人超级智能」路线意味着什么?
Zuckerberg argues superintelligence belongs to everyone, not one lab. What Meta Personal Superintelligence means.
Qwen3.8-27B is out under Apache 2.0. Here is the VRAM math and three ways to run it at home: Ollama, llama.cpp and LM Studio.
Qwen3.8-27B 今日开源(Apache 2.0):270 亿参数、原生多模态、262K 上下文。本文给出显存换算表,并用 Ollama、llama.cpp、LM Studio 三条路线把它跑在家用显卡上。