模型与技术
权重空间编程. OpenAI、DeepSeek、Anthropic 等公司的 LLM 可直接读取并编辑神经网络权重;复现结果显示,它们能不经传统训练实现定向修复并改变行为。
- GPT-6 Astra Edits Qwen Weights: MATH-500 7.9% to 48.1%(在新标签页中打开)
- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights(在新标签页中打开)
- Like GPT-6 Astra, Qwen 3.8 27B can interpret model weights(在新标签页中打开)
- I tried MIT's "weight-space programming" on my own use case: Claude rewrote 8 neurons of Gemma 4 12B to change how it translates one Italian tense(在新标签页中打开)
- DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming(在新标签页中打开)
- What Qwen is thinking when it gets stuck saying "Wait"(在新标签页中打开)
- Qwen 3.8 27B on my laptop can interpret model weights, too(在新标签页中打开)
- GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights(在新标签页中打开)
- [AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai(在新标签页中打开)
- Tried the weight editing thing from the MIT post: Some cool applications, but nothing crazy(在新标签页中打开)
Hydrogen-1.0. Deep Principle 发布了一个基于 GLM-5.3 的 753B MoE 开放权重模型,面向化学、材料和生命科学,并称在 FrontierScience-Research 上击败了 GPT-6-Astra。
前沿模型开放. 腾讯 Hy4 与 Kimi K3.1 Pro 已上线或开放权重,其中 Hy4 新增 Computer Use 和图像理解能力;百度也发布了 280B 基础 MoE 模型,Qwen 的代码仓库则暗示未来将推出 Qwen4 系列。
- Alibaba's Qwen Code repo just leaked qwen4-max, qwen4-coder-plus, and qwen4-omni-120b(在新标签页中打开)
- Official Statement: Qwen Models Will Stick to Open Source to the End!(在新标签页中打开)
- Kimi-K3.1-Pro is now open-weights(在新标签页中打开)
- Kimi K3.1 is likely benchmaxxed (with some proof)(在新标签页中打开)
- Tencent says Hy4 ranks among the best open models. Its weights aren't out yet(在新标签页中打开)
- Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters(在新标签页中打开)
- Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model(在新标签页中打开)
- Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models(在新标签页中打开)
AI 加速 AI 研究. METR 为期 11 周的评估发现,Anthropic 的模型改进速度约翻了一番,其中约 80% 来自其自身研究中使用的 AI,已越过 Anthropic 的负责任扩展阈值。
智能体与基准测试
主动式编程智能体. DeepSeek 开源了一个主动式 GitHub 编程智能体,可将获批请求转化为并行工作项;在模拟环境中,它完成任务的速度比单会话基线快 1.9 倍。
长程智能体基准. SemiAnalysis 发布了包含 206 项任务的基准测试 GDPx;Claude Opus 5.5 以 12% 的得分居首,GPT-6 Astra 为 9.6%,表明智能体在持续数小时的真实工作中仍显吃力。
智能体架构趋同. LangChain 认为,标准的“LLM 循环 + 工具”智能体架构正在趋同,但长程上下文、渐进式披露和沙箱内工具调用又带来了新挑战。
Codex 级 Git. OpenAI 打造了 Git 后端 Pando,把 Codex 级别的并发推送的中位延迟缩短到原来的近五分之一;同时,它正在收购 Pierre Computer Company,后者是 Code Storage 及多个已支撑 Codex 的工具的开发商。
Meta 收购 Conductor. Meta 收购了 AI 编程初创公司 Conductor。该公司的智能体团队让开发者能在多个并行工作区中快速启动项目;创始人 Charlie Holtz 表示,新客户可以期待明年初看到新东西。
基础设施与本地推理
本地 MoE 提速. 借助 tandem.cpp、Strata 等补丁以及新的量化方案,Qwen Flash-Next 和 DeepSeek V4.1 Flash 可在单 GPU 上通过 CPU/SSD 卸载达到 97–200+ tok/s;8GB 显存的配置如今解码速度也能达到 50 tok/s。
- tandem.cpp: Qwen3.8-Flash-Next at 97 tok/s on a single 4090 + 64 GB RAM, 3.3× llama.cpp, works for any MoE(在新标签页中打开)
- Speedy is shockingly good for qwen3.8 FN(在新标签页中打开)
- 8 agents × 250K context, concurrently, on two 3090s (~600+ tok/s total)(在新标签页中打开)
- Sorry Sam, my "quite terrible" home DRAM is at 100% now: patched Strata to make DeepSeek V4.1 agent sessions start 20× faster on a good old standard gaming GPU, no datacenter required(在新标签页中打开)
- Getting incredible results with a single RTX 3090ti and 128gb of ram. Deepseek-V4.1-Flash-1017 (35 tok/s decode) and Qwen3.8-Flash-Next (130 tok/s decode)(在新标签页中打开)
- Qwen 3.8 Flash Next NVFP4 on 64gb RAM and 5090, Basalt. Ditched windows for it. Very hyped for qwen 4 model family now.(在新标签页中打开)
- We are in the golden age of Local LLM(在新标签页中打开)
- Qwen Flash Next on Strata is unreal.(在新标签页中打开)
- Another benchmark has hit the quant war(在新标签页中打开)
量化之争. Unsloth 与 ISTA-DASLab 围绕 DeepSeek-V4.1-Flash 的量化质量公开争论,Unsloth 称其 GGUF 量化版本在长上下文编程任务上性能高出 10–20%;这场竞争正催生更好的量化方案。
- Unsloth vs GSQ-RCO - 512K context DeepSeek-V4.1 GGUF benchmarks Terminal Bench(在新标签页中打开)
- Kind of enjoying the ISTA (GSQ-RCO) vs Unsloth beef(在新标签页中打开)
- Inconsistencies in Unsloth's LiveCodeBench Results(在新标签页中打开)
- New Qwen3.8-Flash-Next Unsloth Dynamic V3.1 GGUFs(在新标签页中打开)
- Unsloth Qwen3.8-Flash-Next GGUFs V3.1 Update (November 4th 2026)(在新标签页中打开)
- Qwen 3.8 Flash Next GSQ RCO quant as good as BF16(在新标签页中打开)
GPU 资源管理器. Berth 是一个基于 Go 的 GPU 资源管理器,为应用提供租约、优先级和请求队列,支持 CUDA 检查点,并可将模型暂存于内存,从而更快切换 AI 工作负载。
商业与政策
OpenAI 融资 110 亿美元. OpenAI 已收到一轮融资中的首笔 110 亿美元,这轮融资对其估值为 1.25 万亿美元;在科技股下跌后,这低于此前融资 300 亿美元、估值 1.4 万亿美元的预期。
Anthropic IPO 承压. Anthropic 营收增长 12 倍,达到 170 亿美元,但因算力承诺亏损 233 亿美元;投资者对其 1.5 万亿美元的 IPO 估值并不买账,有人甚至建议估值 1 万亿美元。
- Anthropic grows revenue twelvefold but posts $23 billion loss(在新标签页中打开)
- Anthropic warns that it’s very close to RSI capability, aka self-building AI(在新标签页中打开)
- Anthropic is still grappling with the economics of compute(在新标签页中打开)
- Anthropic's $1.5 trillion IPO hits resistance as investors balk at valuation(在新标签页中打开)
- [AINews] Anthropic IPO: Whither Now?(在新标签页中打开)
- Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation(在新标签页中打开)
数据中心遭抵制. 选民批准了 41 项地方数据中心限制措施中的 40 项;OpenAI 承诺不与官员签署保密协议,并承担全部电费,而 Meta 则警告 AI 扩张可能放缓。
- After the midterms, a bigger clash over data centers could be coming(在新标签页中打开)
- Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech”(在新标签页中打开)
- Texas was the new epicenter of the AI data center buildout. That might be kaput.(在新标签页中打开)
- Data centers lost the midterms(在新标签页中打开)
- How a Blue Wave midterm might reshape the AI industry(在新标签页中打开)
- After the midterms, is the AI industry in for a rough two years?(在新标签页中打开)
- Big Tech lost big in the 2026 election(在新标签页中打开)
- All the latest updates on AI data centers(在新标签页中打开)
- After the midterms, things are looking rocky for Elon Musk and data centers(在新标签页中打开)
- OpenAI might rebrand its data centers as ‘AI compute campuses’(在新标签页中打开)
- OpenAI is stepping back from NDAs for data center projects(在新标签页中打开)
- After midterm rout, OpenAI tries to win over communities it needs for AI expansion(在新标签页中打开)
- Meta warns that election backlash against data centers could make AI expansion harder(在新标签页中打开)
- ‘Total repudiation’: The blue wave is about to hit AI(在新标签页中打开)
- OpenAI is already trying to shape the 2028 election(在新标签页中打开)
这是本周回顾 – 下周日见。