Chinese contentMistral 开放权重模型 ML4 发布，附技术报告《Le Chonk》幕后故事
Mistral 把 ML4 的权重直接开放了，代码智能体和网络安全能力是主打，三周内比 Preview 版瘦身不少，技术报告也一起放出来了。
DailyVOL 2026.11.04438 items
AI 巨头重写代码底座：OpenAI 发布 Pando
Singapore·Wednesday, 4 November 2026·DAILY · 08:00 SGT
2026年11月4日，今日最大焦点是 AI 编码基础设施的重构：OpenAI 发布自研 Git 后端 Pando，推送延迟降近 80%、日均分发近 1PB 数据，仅用 Codex 约 10 周上线；Anthropic 则以 Scion 完全替换 Git，服务器开销降至原来 1/200——agent 写码速度倒逼工具链重写。第二主题是 可解释与对齐研究：J-PACT 无监督识别多智能体合谋达 99%，另有研究显示 RLVR 只是强化模型已有能力（GSM8K 首次准确率 58.0%→61.3%，但覆盖率本就 94.5%）。第三，架构与推理优化密集涌现：AttentionIndex 令长上下文稀疏注意力吞吐最高提升 1.8 倍，VERS 可省 45.8% 生成 token，Zhipu 发布长时程智能体框架 CoworkOS。
01
Models
5 items
Chinese contentKimi K3.1 Pro 发布，权重即将开源，Together AI 将直播解读架构
Kimi K3.1 Pro 权重明天开源，Kimi 团队亲自讲架构决策，做 Agent 部署的话这场直播挺值得蹲。
Tencent WeChat Open-Sources WeVisOne-VL Vision-Language Model
One small model replacing a stack of separate detectors, OCR engines and matting tools simplifies deployment pipelines for practitioners building document or retail vision systems. The model and data pipeline are on GitHub at github.com/Tencent/WeVisOne-VL, so teams can self-host the 2B or 4B weights. Compared with Qwen3-VL or dedicated segmenters, it trades a single specialised leaderboard edge for broad task coverage plus pixel-level outputs at low parameter counts.
Weight Space Programming lets LLMs read and edit neural network weights
Direct weight editing by LLMs reframes model debugging: instead of retraining, practitioners could patch specific failure modes like repetitive outputs in place. The GDPx benchmark and DeepSeek super-agent references suggest early infrastructure is forming around this idea. Researchers and engineers working on post-training or model introspection should track how these methods compare with fine-tuning workflows.
Chinese content腾讯混元发布 Hyra 2.0，AI 智能体可自主研发科学模型与科研工具
腾讯混元这个 Hyra 2.0 让 AI 自己设计模型、写代码、做训练，从蛋白质工具到修香山芯片都干过了，看看三层智能体怎么分工的。
02
Products
5 itemsChinese contentOpenAI 自研 Git 服务器 Pando，Anthropic 则用 Scion 替换 Git 应对 agent 写码潮
OpenAI 和 Anthropic 都被 agent 写代码的速度逼得重做了代码管理系统，一个改造 Git 一个直接换掉，技术细节讲得很透。
Chinese contentCoworkOS 发布：把任务策略与执行管理分离的长时程智能体框架
一个把'指挥'和'干活'拆开跑的智能体框架，人能直接改文件插手，断线还能接着干，成绩追平 Codex 和 DeepSeek Harness。
Chinese contentAnthropic 开源 FlashProve：普通 CPU 可验证回复由哪个模型生成
Anthropic 开源了 FlashProve，普通 CPU 十几秒就能验证回复到底出自哪个模型，比之前的 VerInf 快太多，转售商糊弄不了人了。
Anthropic launches Message Threads beta for Claude Messages API
Threads cut upload overhead and latency for long-running agents, but they are not a cost-saving feature: billable tokens are identical and caching still has to be configured separately. Developers building computer-use or document-heavy agents benefit most, since screenshots and files no longer need re-uploading every turn. Teams with strict data retention or HIPAA requirements should hold off, and batch processing is not yet supported.
DeepSeek Search and Fetch APIs start charging US$0.75 per 1,000 calls
Agent workflows that loop heavily on Search now carry a separate tool bill on top of model token costs — 10,000 successful searches a day adds roughly US$7.50 before inference fees. Developers building on Claude Code or Codex against DeepSeek endpoints should audit their call volumes and consider caching or batching to control spend. The per-account QPS caps and 402-on-low-balance behaviour also matter for anyone running production agents at scale.
03
Industry
5 itemsChinese contentAnthropic 修订 IPO 文件披露：算力承诺总额最高达 1.07 万亿美元
Anthropic 的 IPO 文件披露了最高 1.07 万亿美元的算力承诺，是收入的 32 倍，现金只有 1319 亿，这份账单值得仔细看看。
Chinese contentMarvell 约 1150 亿美元全股票收购 Lumentum，创半导体并购纪录
Marvell 花了 1150 亿美元买下 Lumentum，就是冲着 AI 数据中心的光互连去的，比 AMD 当年收购 Xilinx 还大一倍多，拆给你看这笔账怎么算的。
IT之家Source
Chinese contentNvidia 要求海外云客户 51% 算力服务不流向中国企业
Nvidia 把出口管制打到了云租赁这条灰色通道上，连终端用户是谁都要报，B300 中国报价翻了四倍多，细节很有信息量。
Chinese contentSam Altman 称下一代 GPT 曾运行 1 万个副本证明一个千禧年大奖难题
Altman 亲口说下一代 GPT 用 1 万个并行副本跑了 4 天证明千禧年难题，去中心化协作的方式挺新鲜。
Chinese contentTutor Intelligence 获 7500 万美元 B 轮融资，用远程人工接管训练机器人
这家公司让机器人在美国 20 个州的真实工厂干活，卡住了就由远程人工接管，接管数据还能拿来训练模型，刚拿了 a16z 领投的 7500 万美元。
04
Papers
4 itemsJ-PACT Detects Collusion Between LLM Agents via J-Space
Unsupervised detection matters because supervised probes need labelled collusion data that real deployments rarely have, so J-PACT lowers the barrier to auditing multi-agent systems. Teams running agent swarms — especially in finance, trading or messaging pipelines where hidden coordination is hard to spot — can use J-space signals as an oversight layer. The method worked across four open-weight model families, suggesting the collusion signature is not model-specific.
Chinese content研究显示 RLVR 只是强化模型已有能力，并未带来新推理知识
有人在 OLMo-2-1B 的 13 个训练检查点上做了逐层解剖，用探针数据证明 RLVR 只是让模型挑出本来会的解法，没教会它新东西，数据和结论都挺扎实。
AttentionIndex Speeds Learned Sparse Attention on GLM-5.3 and DeepSeek-V4-Flash
The method cuts indexer cost, which accounts for over half of prefill time at one million tokens, without the recall degradation that two-stage block approaches like HiSA suffer as context grows. Serving teams running long-context inference on models with learned sparse attention, such as GLM-5 or DeepSeek-V4, can use it to raise prefill throughput 1.3-1.8 times while preserving retrieval accuracy. Both target models come from Chinese labs, making it directly relevant to practitioners deploying these open-weight models.
Scaling Theory Links Multi-Epoch Pretraining Overfitting to Ensembling
As datasets saturate, labs will reuse data across epochs, and this work predicts when that overfits and how ensembling shifts the optimal stopping point. Practitioners planning pretraining or distillation budgets get concrete allocation rules rather than intuition. The FineWeb experiments make the theory testable against single-model scaling baselines.
05
How-to
3 itemsBuild a governed semantic layer for agents with EDB Postgres AI
Putting governance, joins and row-level security in the database means agent permissions are enforced by Postgres itself, not by fragile prompt-level safeguards. Developers building agents on Bedrock AgentCore or Strands can follow the tutorial to make data access auditable and scoped per agent. It contrasts with approaches that hard-code business logic into each agent's instructions.
KVCache Meetup Korea Spotlights LLM Serving Optimization
The talks show that LLM serving bottlenecks sit as much in memory architecture — CXL pooling, HBM bandwidth — as in GPU compute, a framing many benchmark writeups skip. Engineers building inference infrastructure can study the roofline analysis and LMCache MP Mode patterns for their own deployments. The event also signals a growing Korean accelerator and serving stack around FuriosaAI's RNGD and SK hynix memory.
NVIDIA Explains PyTorch Performance Tuning with TensorRT 11.1
The post gives a conceptual framework for judging whether a compiled engine is genuinely better, which matters when automated tuning trades accuracy for speed without a clear method to evaluate that trade-off. Engineers deploying PyTorch models with TensorRT can use Global Performance Tuner in the 11.1 release to automate tuning instead of hand-tuning configurations.
438Stories
69First-party
42New models
116Sources
AITOP · generated by the editorial system