Chinese editionMeta 发布开源权重模型 Muse Spark Open：816B 参数，登顶美国开源权重榜首
Meta 放了个 816B 的开源权重模型 Muse Spark Open，美国开源第一，1M 上下文还能塞进两台 Mac Studio 自托管，不过跑智能体任务不如 GLM-5.3。
DailyVOL 2026.10.17312 items
Meta's 816B open-weight Muse Spark Open tops US leaderboard
Singapore·Saturday, 17 October 2026·DAILY · 08:00 SGT
2 Sep 2026. Meta released Muse Spark Open, an 816B-parameter open-weight MoE model (42B active) that tops the US open-weight ranking with a score of 39 on the Artificial Analysis Intelligence Index, beating NVIDIA Nemotron 3.5 Super VL (26) and priced at $1.25/$4.25 per million input/output tokens. Microsoft said its MAI-Base-2 (2.8T total, 80B active) pretrained on 6,000 GB200s outperforms all tested open-weight models with 4–12x the pretraining compute efficiency of DeepSeek-V4-Pro-Base.
Agents moved into the physical world. Google Research connected AI agents to wet-lab robots, discovering protein binders with single-digit nanomolar affinity to a cancer target. Gemini 4 Argon's Minecraft experiment showed four agents collaborating for 27 hours, writing 2,783 programs. Anthropic reported Claude Mythos 5 optimized quantum circuits in a closed loop, beating published best algorithms on 8 primitives. Perplexity open-sourced SPACE, its Firecracker-based agent code execution platform, and Anthropic's anti-distillation defenses cut stolen reasoning tokens by 90–99%.
Infrastructure matured. vLLM, now running on ~500,000 GPUs, received Day-0 support for Reflection-Beam-1.0 (501B, 1M context); maintainer Inferact raised $150M. AMD MI355X beat NVIDIA HGX B300 by 24.4% on Kimi K3 serving throughput. Anthropic revealed Claude wrote ~90% of Claude Code's code.
01
Models
5 items
Reflection-Beam-1.0 gets Day-0 support in vLLM
Day-0 serving and RL support means practitioners can deploy and fine-tune a 501B-parameter sparse model without waiting for ecosystem catch-up. Teams running large-scale inference can start with vLLM and add reinforcement learning through SkyRL's LoRA or MXFP8 paths. The 1M-token context window puts Beam in direct competition with other frontier sparse models for long-document and agentic workloads.
Chinese editionAnthropic：Claude Mythos 5 在闭环中优化量子电路，超越已发表最优算法
Anthropic 让 Claude Mythos 5 自己写量子电路、自己按分数迭代，8 个原语上都打赢了已发表最优解，量子方向的朋友可以看看.
Chinese edition微软完成 MAI-Base-2 预训练，2.8T 参数强于所有参测开源权重模型
微软自研基模 MAI-Base-2 训完了，2.8T 总参数，算力效率是 DeepSeek-V4-Pro-Base 的 4~12 倍，还放出了训练细节。
IT之家Source
Tencent Hunyuan open-sources Hy Memory 2.0 memory infrastructure
Instead of wrapping a general model in a memory framework like Mem0 or MemOS, this uses a small dedicated writer model, cutting inference cost to roughly 3B-model levels while staying cross-framework. Developers building agent pipelines can pip install it and swap storage backends freely, and Mac users can share memory across coding tools via MemHub. The sub-1-point margin over Claude Opus 5 is self-reported and awaits independent replication.
02
Products
5 itemsChinese editionAnthropic 推出三层反蒸馏防御，Claude 推理 token 窃取量下降 90%–99%
Anthropic 把反蒸馏说得很细：三层防御怎么设计、为什么思维链被偷走最伤，还直接点名智谱 GLM-5.3 和 Kimi，信息量不小。
Chinese editionGoogle Research 将智能体接入湿实验室机器人，自动设计蛋白质结合剂
Google 把 AI 智能体直接接上实验台机器人，自己提方案、机器人做实验，最后还真做出了纳摩尔级的抗癌蛋白结合剂。
Perplexity open-sources SPACE, its agent code execution platform
The interesting part is the state machine: Pause keeps memory resident for fast recovery, Suspend snapshots VM memory plus filesystem for true persistence, with idle sandboxes auto-suspending after 6 hours. Engineers building agent sandboxes or code-interpreter features can study how warm pools and btrfs snapshots attack cold-start latency, a problem most teams still solve ad hoc. Management and attachment tokens being non-interchangeable is a useful security boundary pattern to copy.
Chinese editionCheng Lou 开源 Handwired：把程序逻辑手写成免训练神经网络
Cheng Lou 把 JS 代码手写成零训练的神经网络，浏览器里跑出 130 层的小桌面系统，demo 网页直接能玩，原理讲得很清楚。
AutoInfra launches Tinker-compatible training API at lower cost
The pricing gap against Tinker suggests infra cost, not algorithms, is becoming a competitive lever for agentic research pipelines. Isolated algorithm, training and sandbox environments are built specifically to guard against reward hacking during self-improvement cycles, a design detail teams running autonomous R&D agents will study closely. Practitioners can benchmark the API against their current training stack before committing budget.
03
Industry
3 itemsChinese editionvLLM 从 Berkeley 学生项目成长为开源推理引擎，背后的 Inferact 融资 1.5 亿美元
两个博士生优化推理速度的副业项目，如今跑在 50 万张 GPU 上，Google 和 Meta 都在用，还融了 1.5 亿美元。
Chinese editionSam Altman 转发 Nick Cammarata：可解释性研究进展超预期
Sam Altman 转发了一位 OpenAI 前可解释性负责人的长文，说现在万亿参数模型的神经元大多能看懂了，和他早年只解读出 2% 的判断完全反转。
Chinese edition从 Gemini 4 Argon 的 Minecraft 实验看模型持久性能力
四个 AI 智能体没分工却在 Minecraft 里熬了 27 小时、死了 605 次通关，这篇分析讲了为什么持久性比聪明更关键。
04
Papers
3 itemsMCP-FSD Uncovers Filesystem Capability Laundering in MCP Servers
Client-side tool approval in MCP grants the whole server process broad filesystem access, so a tool approved for one operation can exfiltrate credentials the contract never mentions. Security engineers running local agent stacks should audit installed MCP servers and enforce sandbox boundaries rather than trusting declared tool schemas. The two CVEs in Anthropic's Git MCP Server make this a concrete patch item for anyone using that server.
MLExpert: 35B Agentic Model for Autonomous Machine Learning Engineering
A 35B model matching frontier-scale systems on MLE-Bench Lite suggests agentic MLE performance depends more on training method than raw parameter count, lowering the hardware bar for teams building autonomous experimentation agents. The WIRL reward design over multi-turn execution windows is a reusable recipe for anyone training long-horizon agents beyond ML engineering. The paper also ships an executable task factory, so practitioners can adapt the harness to their own domains.
Chinese editionLibrarian Agents：用强化学习让 LLM 提前整理文档工作区
Meta 这篇论文把文档整理变成了 RL 训练任务，builder agent 提前建好文件系统，BrowseComp-Plus 上比检索基线高 33 个点，做长文档问答的可以看看。
05
How-to
5 itemsChinese editionAMD MI355X 在 Kimi K3 推理上性能超 NVIDIA B300 24.4%
AMD 卡跑 Kimi K3 吞吐反超 NVIDIA B300 四分之一，每美元 token 还多 87%，文中还有 vLLM 瓶颈排查套路，自己搭推理服务的可以看看。
Chinese editionAnthropic：Claude 写了 Claude Code 约 90% 的代码
Anthropic 自家数据：Claude 写了 Claude Code 九成代码，PR 翻倍、回退率降 40%，他们把改审查流程的方法也公开了。
Chinese editionClaude Code 团队如何用 Claude 做代码评审
Claude Code 团队自己怎么用 Claude 审代码，评审前置成拼写检查式的持续流程，写代码的都该看看
University of Michigan AI course releases Subagents lecture notes
The fresh-versus-forked distinction is the practical takeaway: forking a subagent inherits your assumptions, so it tends to confirm what you already believe, while a fresh start is the option for independent code review. Developers building agent workflows with Claude Code or similar harnesses can apply the three-question test — how many conversations, who talks to whom, who decides next — to distinguish subagents from teams and handoffs. The notes are freely available on GitHub for self-study.
Chinese editionAWS 发布基于 Amazon Bedrock AgentCore 的智能体九阶段评估框架
AWS 把智能体上线后质量悄悄变差这个坑拆成了九个评估阶段，基于 AgentCore 做，做智能体运维的可以照着搭一套。
312Stories
59First-party
68New models
116Sources
AITOP · generated by the editorial system