模型與技術
直接改權重. OpenAI、DeepSeek、Anthropic 等團隊的 LLM 能直接讀取並編輯神經網路權重;已有複現結果顯示,不必靠傳統訓練,就能進行針對性修正並改變行為。
- GPT-6 Astra Edits Qwen Weights: MATH-500 7.9% to 48.1%(在新分頁中開啟)
- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights(在新分頁中開啟)
- Like GPT-6 Astra, Qwen 3.8 27B can interpret model weights(在新分頁中開啟)
- 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(在新分頁中開啟)
- DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming(在新分頁中開啟)
- What Qwen is thinking when it gets stuck saying "Wait"(在新分頁中開啟)
- Qwen 3.8 27B on my laptop can interpret model weights, too(在新分頁中開啟)
- GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights(在新分頁中開啟)
- [AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai(在新分頁中開啟)
- Tried the weight editing thing from the MIT post: Some cool applications, but nothing crazy(在新分頁中開啟)
Hydrogen-1.0. Deep Principle 發布以 GLM-5.3 為基礎、針對化學、材料與生命科學的開放權重 753B MoE 模型,並宣稱在 FrontierScience-Research 上擊敗 GPT-6-Astra。
開放前沿模型. 騰訊的 Hy4 與 Kimi K3.1 Pro 現已可用或開放權重,其中 Hy4 新增電腦操作與圖像理解;百度也發布 280B 基礎 MoE,而 Qwen 的程式碼庫則暗示未來將推出 Qwen4 模型。
- Alibaba's Qwen Code repo just leaked qwen4-max, qwen4-coder-plus, and qwen4-omni-120b(在新分頁中開啟)
- Official Statement: Qwen Models Will Stick to Open Source to the End!(在新分頁中開啟)
- Kimi-K3.1-Pro is now open-weights(在新分頁中開啟)
- Kimi K3.1 is likely benchmaxxed (with some proof)(在新分頁中開啟)
- Tencent says Hy4 ranks among the best open models. Its weights aren't out yet(在新分頁中開啟)
- Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters(在新分頁中開啟)
- Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model(在新分頁中開啟)
- Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models(在新分頁中開啟)
AI 加速 AI 研究. METR 為期 11 週的檢視發現,Anthropic 的模型改進速度約翻了一倍,其中約 80% 來自其自身研究使用的 AI,而這也已跨過 Anthropic 的負責任擴展門檻。
AI Agent 與基準測試
主動式程式碼 Agent. DeepSeek 開源一款主動式 GitHub 程式碼 Agent,會把核准的請求轉成可平行處理的工作項目;在模擬中,完成任務的速度達單一工作階段基準的 1.9 倍。
長時程 Agent 基準. SemiAnalysis 發布涵蓋 206 項任務的基準測試 GDPx;Claude Opus 5.5 以 12% 居首,GPT-6 Astra 為 9.6%,顯示 Agent 在需要數小時的現實工作裡仍難以應付。
Agent 架構趨同. LangChain 認為,標準的「LLM 迴圈與工具」Agent 架構正趨於一致,但長時程上下文、漸進式揭露與沙箱化工具使用等新挑戰也隨之而來。
Codex 級 Git. OpenAI 打造 Git 後端 Pando,針對 Codex 規模的並行推送,把中位延遲壓低近 5 倍,並將收購 Pierre Computer Company;該公司開發的 Code Storage 與相關工具已為 Codex 提供支撐。
Meta 收購 Conductor. Meta 收購 AI 程式設計新創 Conductor;其 Agent 團隊功能讓開發者可在多個平行工作區啟動專案。創辦人 Charlie Holtz 表示,新客戶可望在明年初看到新東西。
基礎設施與本地推論
本地 MoE 加速. tandem.cpp、Strata 等修補程式加上新的量化版本,讓 Qwen Flash-Next 與 DeepSeek V4.1 Flash 在單張 GPU 搭配 CPU/SSD 卸載時可跑到 97–200+ tok/s;8GB VRAM 的配置現在解碼速度達 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(在新分頁中開啟)
- Speedy is shockingly good for qwen3.8 FN(在新分頁中開啟)
- 8 agents × 250K context, concurrently, on two 3090s (~600+ tok/s total)(在新分頁中開啟)
- 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(在新分頁中開啟)
- 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)(在新分頁中開啟)
- Qwen 3.8 Flash Next NVFP4 on 64gb RAM and 5090, Basalt. Ditched windows for it. Very hyped for qwen 4 model family now.(在新分頁中開啟)
- We are in the golden age of Local LLM(在新分頁中開啟)
- Qwen Flash Next on Strata is unreal.(在新分頁中開啟)
- Another benchmark has hit the quant war(在新分頁中開啟)
量化之爭. Unsloth 與 ISTA-DASLab 正公開爭論 DeepSeek-V4.1-Flash 的量化品質;Unsloth 宣稱其 GGUF 在長上下文程式碼任務上表現高出 10–20%,這場較勁正帶來更好的量化版本。
- Unsloth vs GSQ-RCO - 512K context DeepSeek-V4.1 GGUF benchmarks Terminal Bench(在新分頁中開啟)
- Kind of enjoying the ISTA (GSQ-RCO) vs Unsloth beef(在新分頁中開啟)
- Inconsistencies in Unsloth's LiveCodeBench Results(在新分頁中開啟)
- New Qwen3.8-Flash-Next Unsloth Dynamic V3.1 GGUFs(在新分頁中開啟)
- Unsloth Qwen3.8-Flash-Next GGUFs V3.1 Update (November 4th 2026)(在新分頁中開啟)
- Qwen 3.8 Flash Next GSQ RCO quant as good as BF16(在新分頁中開啟)
GPU 資源管理器. Berth 是以 Go 打造的 GPU 資源管理器,為應用程式提供租約、優先順序與請求佇列,支援 CUDA checkpointing,並可將模型暫留在 RAM 中,加快 AI 工作負載切換。
商業與政策
OpenAI 募資 110 億美元. OpenAI 已獲得一輪融資的第一筆 110 億美元,該輪對其估值 1.25 兆美元;在科技股下跌後,這低於先前以 1.4 兆美元估值募得 300 億美元的期望。
Anthropic IPO 壓力. Anthropic 營收成長 12 倍至 170 億美元,但由於算力承諾而虧損 233 億美元;投資人對其 1.5 兆美元的 IPO 估值卻步,有人建議 1 兆美元。
- Anthropic grows revenue twelvefold but posts $23 billion loss(在新分頁中開啟)
- Anthropic warns that it’s very close to RSI capability, aka self-building AI(在新分頁中開啟)
- Anthropic is still grappling with the economics of compute(在新分頁中開啟)
- Anthropic's $1.5 trillion IPO hits resistance as investors balk at valuation(在新分頁中開啟)
- [AINews] Anthropic IPO: Whither Now?(在新分頁中開啟)
- Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation(在新分頁中開啟)
資料中心遇反彈. 選民通過了 41 項地方資料中心限制中的 40 項;OpenAI 承諾不與官員簽保密協議,並全額支付電力成本,而 Meta 警告 AI 擴張可能放緩。
- After the midterms, a bigger clash over data centers could be coming(在新分頁中開啟)
- Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech”(在新分頁中開啟)
- Texas was the new epicenter of the AI data center buildout. That might be kaput.(在新分頁中開啟)
- Data centers lost the midterms(在新分頁中開啟)
- How a Blue Wave midterm might reshape the AI industry(在新分頁中開啟)
- After the midterms, is the AI industry in for a rough two years?(在新分頁中開啟)
- Big Tech lost big in the 2026 election(在新分頁中開啟)
- All the latest updates on AI data centers(在新分頁中開啟)
- After the midterms, things are looking rocky for Elon Musk and data centers(在新分頁中開啟)
- OpenAI might rebrand its data centers as ‘AI compute campuses’(在新分頁中開啟)
- OpenAI is stepping back from NDAs for data center projects(在新分頁中開啟)
- After midterm rout, OpenAI tries to win over communities it needs for AI expansion(在新分頁中開啟)
- Meta warns that election backlash against data centers could make AI expansion harder(在新分頁中開啟)
- ‘Total repudiation’: The blue wave is about to hit AI(在新分頁中開啟)
- OpenAI is already trying to shape the 2028 election(在新分頁中開啟)
這是本週回顧 – 下週日見。