Chinese contentClaude Opus 5.5 登顶 Artificial Analysis 编程智能体指数，但单任务成本更高
Anthropic 的 Opus 5.5 在编程智能体评测拿第一，三项基准都涨了，价格还降了 20%，不过每任务要烧 13 美元 tokens，用前算算账。
DailyVOL 2026.09.25420 items
DeepSeek reveals DSec platform running 3 million agent sandboxes daily
Singapore·Friday, 25 September 2026·DAILY · 08:00 SGT
25 Sep 2026. DeepSeek disclosed its DSec elastic compute sandbox platform for agentic reinforcement learning, running roughly 3 million sandboxes per day with peak concurrency above 380,000 — the day's single biggest fact. On models, Anthropic's Claude Opus 5.5 topped the Artificial Analysis Coding Agent Index with 66 points, lifting Terminal-Bench 4.0 to 63.1%, DeepSWE v1.1 to 68.4% and SWE-Atlas-QnA to 66.4%, while Anthropic also reported Claude discovering ART, a CRISPR-like enzyme system in phage DNA, after ~950 agents ran 21 hours and consumed 210 million tokens. OpenAI launched MentalHealthBench, 1215 synthetic conversations with 5262 expert rubric items; GPT-6 Astra scored 57.3%, Claude Opus 5.5 52.4% and Gemini 2.5 Pro 29.5%. In infrastructure, Huawei launched the Atlas 960E SuperPoD scaling to 4,096 NPUs and 8 EFLOPS FP8, Xiaomi published the HySparse2 architecture for MiMo-V3, and inferact open-sourced a TPU megakernel pushing Kimi K3 to 709 tokens/s. Google will launch its Project Suncatcher MVP satellite on 1 October, carrying 4 TPU units to process AI queries in orbit. Epoch AI estimates AI costs are falling about 47% per quarter, faster than any past transformative technology.
Section Synthesis & Key Takeaways: Today's briefing curates 420 high-signal developments across 136 global and regional sources.
Models (5 picks): 「Claude Opus 5.5 登顶 Artificial Analysis 编程智能体指数，但单任务成本更高」（Anthropic 的 Opus 5.5 在编程智能体评测拿第一，三项基准都涨了，价格还降了 20%，不过每任务要烧 13 美元 tokens，用前算算账）; 「Claude 在噬菌体中发现类 CRISPR 新酶系统 ART」（Anthropic 让 950 个 Claude 智能体跑了 21 小时，真在基因数据库里挖出一个类似 CRISPR 的新酶系统，流程用的还是公开发布的 Claude Science 和 Claude Code）.
Products (5 picks): 「Huawei Atlas 960E SuperPoD with 4,096 NPUs and 8 EFLOPS FP8」（Optical co-packaging via Hi-ONE NPO points to where large-scale AI cluster interconnects are heading, trading discrete pluggable modules for integrated optics with measurable savings in parts count and power. Hardware engineers planning training clusters and procurement teams evaluating alternatives to NVIDIA-based superpods will find the scale and efficiency figures directly comparable）; 「阿里开源 OpenCodeReview：确定性 pipeline 与 Agent 混合的 AI 代码审查工具」（阿里把内部跑了两年的代码审查工具开源了，用确定性 pipeline 管文件选择，token 消耗只有 Claude Code 的九分之一，还能当插件接进 Cursor）.
Industry (4 picks): 「Jürgen Schmidhuber 加盟 Sakana AI 出任首席科学顾问」（LSTM 和世界模型概念的提出者 Schmidhuber 加入了 Sakana AI 的 RSI Lab，做递归自我改进研究，10 月还要在东京办公开活动）; 「Epoch AI 数据：AI 成本每季度下降约 47%，快于历史上任何变革性技术」（Epoch AI 给了组狠数字：同性能下 AI 成本每季降 47%，比当年电力便宜 54 倍的速度下滑，值得对比下自己行业的降本曲线）.
Papers (3 picks): 「Apple 发布半监督联邦 ASR 训练方案：在线伪标签配合服务器端更新稳定化」（Apple 发了一篇联邦学习论文，讲怎么用伪标签在无标注数据上训 ASR 模型，还给了稳定训练的实操配方，做语音方向的可以看看）; 「DeepSeek-Backed Benchmark Tests LLM Backtest Auditing With Clean Controls」（The non-obvious result is that recall-only evaluations inflate auditor quality: specificity against clean strategies is where models actually diverge, by as much as 79 points. Quants and fintech teams deploying LLMs to audit trading backtests should adopt paired clean-control scoring and clean-aware prompts, since a single-line warning eliminated DeepSeek's false positives in this setup. The benchmark and deterministic scorer offer a reusable template for calibrating any audit-style LLM task）.
How-to (3 picks): 「用 WhisperX 在 SageMaker AI 上实现带说话人标签的语音转写」（AWS 出了个现成的 WhisperX 容器，一键部署到 SageMaker 就能拿到带说话人标注的逐词转写，还讲了扩缩和省钱细节）; 「用 AWS AgentCore Gateway 和 MCP 搭建多账号 AI 智能体」（AWS 出的实操教程，教你用 AgentCore Gateway 加 MCP，让一个智能体安全查询多个账号里的数据，做企业级架构的可以照着搭）.
01
Models
5 items
Chinese contentClaude 在噬菌体中发现类 CRISPR 新酶系统 ART
Anthropic 让 950 个 Claude 智能体跑了 21 小时，真在基因数据库里挖出一个类似 CRISPR 的新酶系统，流程用的还是公开发布的 Claude Science 和 Claude Code。
Xiaomi unveils HySparse2 architecture for MiMo-V3 agent inference
The reveal signals Xiaomi is positioning MiMo-V3 specifically for agent workloads, where attention efficiency over growing histories matters more than raw benchmark scores. Engineers building multi-turn agents can compare HySparse2's KV sharing against Alibaba Qwen and DeepSeek approaches for long-context cost control. The architecture details also hint at what RL training changes shipped in MiMo-V2.6.
IT之家Source
Chinese contentOpenAI 发布 MentalHealthBench：1215 段对话评估 AI 心理健康应对能力
OpenAI 拿出 1215 段心理对话让各家模型答题，GPT-6 Astra 拿 57.3% 第一，Gemini 2.5 Pro 只有 29.5%，数据集已开放下载。
IT之家Source
Chinese contentinferact 开源 Kimi K3 TPU megakernel，单核跑出 709 tokens/s
Kimi K3 在 TPU 上跑到 709 tokens/s，比 GB200 还快，92 个 MoE 层塞进一个 Pallas kernel，做法挺巧，做推理的可以看看。
02
Products
5 itemsHuawei Atlas 960E SuperPoD with 4,096 NPUs and 8 EFLOPS FP8
Optical co-packaging via Hi-ONE NPO points to where large-scale AI cluster interconnects are heading, trading discrete pluggable modules for integrated optics with measurable savings in parts count and power. Hardware engineers planning training clusters and procurement teams evaluating alternatives to NVIDIA-based superpods will find the scale and efficiency figures directly comparable.
Chinese content阿里开源 OpenCodeReview：确定性 pipeline 与 Agent 混合的 AI 代码审查工具
阿里把内部跑了两年的代码审查工具开源了，用确定性 pipeline 管文件选择，token 消耗只有 Claude Code 的九分之一，还能当插件接进 Cursor。
Anthropic Claude Code Cloud Sessions Launch with US$100-250 Credits
Cloud sessions turn Claude Code into an asynchronous worker rather than a live coding partner, closing the gap with cloud-agent offerings from rivals like OpenAI Codex and Google Jules. Developers on Pro or Max plans should claim the credit before 7 Oct 2026 via /claim-credit, as it effectively gives weeks of free remote agent runs. Teams with strict network policies should note that IP whitelists can break authentication on Anthropic-hosted environments.
IT之家Source
Chinese contentOpenRouter 推出 Tool Search：工具定义按需加载，省下输入 token
OpenRouter 出了个 Tool Search，函数标成 defer_loading 就能按需加载工具定义，不用每轮都为几百个工具定义付 token，还不加价。
Chinese contentPerplexity 发布 Fast Search：基于 Rust 引擎 Photon 的搜索 API
Perplexity 把搜索 API 换成了自家 Rust 写的 Photon 引擎，95% 结果 230 毫秒就回来，做搜索相关应用的可以试试延迟够不够快。
03
Industry
4 itemsChinese contentJürgen Schmidhuber 加盟 Sakana AI 出任首席科学顾问
LSTM 和世界模型概念的提出者 Schmidhuber 加入了 Sakana AI 的 RSI Lab，做递归自我改进研究，10 月还要在东京办公开活动。
Chinese contentEpoch AI 数据：AI 成本每季度下降约 47%，快于历史上任何变革性技术
Epoch AI 给了组狠数字：同性能下 AI 成本每季降 47%，比当年电力便宜 54 倍的速度下滑，值得对比下自己行业的降本曲线。
Chinese content谷歌 10 月 1 日发射试验卫星 MVP，在太空处理 AI 查询
谷歌要往天上扔数据中心了：下周发射的 MVP 卫星带 4 枚 TPU，能直接在太空答简单 AI 问题，这是验证第一步。
IT之家Source
OpenEvidence logs 42 million physician queries in August
The interesting implication is that a domain-specific AI product can enumerate user demand faster and more precisely than general-purpose labs, turning query data into a flywheel for new medical knowledge and research funding. Healthcare founders, clinical researchers and med-tech investors can mine this pattern to identify where clinical knowledge gaps exist. It also offers a benchmark for vertical AI products elsewhere: depth of daily use may matter more than total user count.
04
Papers
3 itemsChinese contentApple 发布半监督联邦 ASR 训练方案：在线伪标签配合服务器端更新稳定化
Apple 发了一篇联邦学习论文，讲怎么用伪标签在无标注数据上训 ASR 模型，还给了稳定训练的实操配方，做语音方向的可以看看。
DeepSeek-Backed Benchmark Tests LLM Backtest Auditing With Clean Controls
The non-obvious result is that recall-only evaluations inflate auditor quality: specificity against clean strategies is where models actually diverge, by as much as 79 points. Quants and fintech teams deploying LLMs to audit trading backtests should adopt paired clean-control scoring and clean-aware prompts, since a single-line warning eliminated DeepSeek's false positives in this setup. The benchmark and deterministic scorer offer a reusable template for calibrating any audit-style LLM task.
DeepSeek Details DSec Sandbox Platform for Agentic RL Training
The throughput figures reveal the infrastructure scale behind agentic RL training, which most labs keep undisclosed. Teams building agent training environments or RL pipelines can use the reported concurrency and creation-rate numbers as a reference point when sizing their own sandbox fleets. The arXiv report with Wenfeng Liang as co-author signals this is a company-level engineering priority, not a side project.
05
How-to
3 itemsChinese content用 WhisperX 在 SageMaker AI 上实现带说话人标签的语音转写
AWS 出了个现成的 WhisperX 容器，一键部署到 SageMaker 就能拿到带说话人标注的逐词转写，还讲了扩缩和省钱细节。
Chinese content用 AWS AgentCore Gateway 和 MCP 搭建多账号 AI 智能体
AWS 出的实操教程，教你用 AgentCore Gateway 加 MCP，让一个智能体安全查询多个账号里的数据，做企业级架构的可以照着搭。
Chinese contentDatabricks 发布教程：用 ai_query 在治理数据上运行 Jev 决策模型
Databricks 出了个可直接导入的 notebook，教你把 Jev 决策模型跑在自家 Serverless GPU 上，SQL 里调 ai_query 就能拿分类和概率分数，换上自己数据就能用。
420Stories
78First-party
62New models
136Sources
AITOP · generated by the editorial system