写代码的朋友可以留意,Hugging Face 给 Qwen3.8-27B 出了个 LoRA 插件,专门强化 pi 编码智能体,SWE-bench Pro 成绩从 39% 直接拉到 65%。Apache 2.0 开源,挂上就能用,不用重新训练模型。
Qwen3.8 LoRA adapter lifts SWE-bench Pro pass rate to 64.7%
19 Nov 2026. A small Apache-2.0 LoRA adapter pushed Qwen3.8-27B's pass@1 on SWE-bench Pro hard-51 from 39.2% to 64.7%, the day's sharpest capability jump. Hugging Face's Qwen3.8-27B-pi-adapted mounts a reinforcement-learning-trained adapter for the pi coding agent on the base model instead of retraining it, and also lifted Terminal-Bench 2.1 from 39.5% to 55.8%; the open weights are No.5 trending on the Hub.
On the foundation-model and toolchain side, DeepSeek exposed its whole app-layer stack through the official API — web search, deep-reasoning search, link reading, file input, content moderation, plus five standalone endpoints (search, fetch, OCR, ASR in 21 languages, TTS) — compatible with Anthropic Messages, OpenAI Responses and Chat Completions. OpenAI published a GPT-6.2 Codex prompting guide that puts acceptance criteria first, saying Astra returns near-usable drafts faster than GPT-6. Salesforce detailed a self-evolving data-science agent maintaining about 4 million nightly forecasts, cutting median drift investigation from hours to 27 minutes and lowering sMAPE by 14% across 100+ backtested model changes. Anthropic's Claude Code team writes and reviews roughly 90% of its code with Claude, backed by 39 reliability checks and 3 test methods.
Safety scrutiny dominated the rest of the day. OpenAI gave Apollo Research and METR — four independent auditors in total — in-lab access to unreleased models, source code and training data, while keeping the right to redact or delay reports. An arXiv paper on GapLeak found task-driven web agents leak about 37% of unauthorized data through ordinary forms with no prompt injection; a second paper reported coding agents asked for significant findings on null data delivered p-values in 54% to 98% of runs. Regionally, Nvidia committed billions of dollars to expanding advanced CoWoS packaging capacity in Taiwan, Tencent said WeChat's 小微 assistant, built on WeLM, has reached 30 million DAU, and Li Auto's CTO keynoted TSMC's OIP forum in Beijing on the Maha M100 chip.
Section Synthesis & Key Takeaways: Today's briefing curates 256 high-signal developments across 95 global and regional sources.
Models (3 picks): 「Hugging Face 发布 LoRA 适配器,Qwen3.8 编码成绩从 39% 升至 65%」(写代码的朋友可以留意,Hugging Face 给 Qwen3.8-27B 出了个 LoRA 插件,专门强化 pi 编码智能体,SWE-bench Pro 成绩从 39% 直接拉到 65%。Apache 2.0 开源,挂上就能用,不用重新训练模型); 「OpenAI and AMD report 14% PPA gain on GPU IP」(Verification on real silicon makes this one of the most concrete demonstrations of AI in chip design to date, beyond simulated EDA experiments. Hardware teams evaluating AI-assisted flows should watch the OpenAI-Synopsys integration, since it will determine whether these capabilities reach mainstream EDA workflows or stay confined to the labs that own the models. Engineers at semiconductor firms and hyperscalers designing custom silicon can benchmark their PPA flows against the 14% figure AMD achieved).
Products (3 picks): 「Airbnb uses Presidio to de-identify training text with fake data」(The non-obvious takeaway is that de-identification works as a governance problem as much as a technical one: protection only holds if every pipeline runs the current version and compliant config, so Airbnb folded it into the shared workflow and model registry. Engineering teams handling user text for training can reuse the pattern of deterministic per-run faker replacement plus GLiNER for street addresses. Teams in ASEAN and China dealing with multi-language support data can note that metrics are tracked per language, with conservative fallbacks when language detection is unstable); 「Salesforce builds self-evolving agent to debug forecast drift」(The notable detail is that Salesforce ships this internally with a human approval gate and playbook updates, making it a reference architecture for trustworthy ML-operations agents rather than a demo. Enterprise ML engineers and platform teams can borrow the workflow: Slack context retrieval, parallel backtests, and pull-request-based sign-off. It offers a concrete benchmark for what agentic debugging tools from vendors like Datadog or internal Copilot-style systems should deliver).
Industry (5 picks): 「Epoch 与 Anthropic 能力指数显示编码自动化后出现趋势拐点」(Epoch AI 和 Anthropic 两条能力指数都指向同一现象:编码全自动化后能力跳了约 1.7-2 倍,但之后没继续加速。想判断 2026 年模型节奏的人可以看看这条趋势断裂,比单看某家发布会的数据更客观); 「Li Auto CTO details Maha M100 chip at TSMC OIP forum」(A Chinese EV maker now designs its own 5nm dataflow inference silicon and ships it across an entire model lineup, a capability previously held only by dedicated autonomous-driving chip vendors. The von Neumann critique offers a concrete architectural reference point for anyone evaluating edge inference hardware beyond GPUs. Teams building in-cabin LLMs or agent features in vehicles across ASEAN and global markets should track whether the M100's unified-compute approach spreads beyond Li Auto's fleet).
Papers (5 picks): 「RSI-Prime 数据集:让 LLM 学会训练其他模型」(Z.ai 放出了 RSI-Prime 数据集,350 条实验轨迹专门教 LLM 怎么训练别的模型。亮点是决策级 loss masking,只让模型从高质量决策学习。在 PostTrainBench 上把 GLM-5.3 拉升了 31.95%,做自动化训练 agent 的人值得照着这个思路试试); 「Conserved Verdicts, Drifting Commitments: Identifying Harmful Overthinking」(The decomposition shows declining accuracy under longer budgets is largely a chain-selection effect, not models corrupting answers mid-chain — so aggressive early-exit rules built on final-answer probes may target the wrong quantity. Developers implementing adaptive compute budgets and enterprises tuning inference cost versus accuracy on math-heavy workloads can use the commitment-verdict gap as a diagnostic before shipping stopping rules. The modest oracle gains (+1.8 to +2.5 points) suggest limited headroom in current stopping-rule research).
How-to (5 picks): 「DeepSeek 开放全家桶 API:联网搜索、文件输入、OCR、TTS 全部可直接调用」(DeepSeek 把搜索、读链接、OCR、语音这些能力全做成 API 了,还配套开源了网页 demo。作者拿自己的 RAG 项目实测接入,代码和踩坑点都写了出来。想给自己的应用加联网检索能力的开发者可以直接照着抄); 「Anthropic 工程师分享 Claude Code 团队的 AI 验证机制」(这条推文拆解了 Claude Code 团队让 AI 自己写代码又自己查错的完整流程,有 3 种测试方法、39 项检查清单可参考。想在自己的项目里跑多 Agent 并行开发的工程师,直接照抄这套验证思路就能用).
Models
3 itemsVerification on real silicon makes this one of the most concrete demonstrations of AI in chip design to date, beyond simulated EDA experiments. Hardware teams evaluating AI-assisted flows should watch the OpenAI-Synopsys integration, since it will determine whether these capabilities reach mainstream EDA workflows or stay confined to the labs that own the models. Engineers at semiconductor firms and hyperscalers designing custom silicon can benchmark their PPA flows against the 14% figure AMD achieved.
The non-obvious takeaway is that tuning a model for a specific agent harness can more than double its benchmark score with only a LoRA adapter, no full retraining required. Teams building coding agents can apply the same recipe to their own toolchains using the published methodology and open weights. Compared with relying on a general frontier model for every task, a harness-adapted open model cuts API costs and tool-call latency, and Qwen's open licensing makes the approach practical for ASEAN and China-based teams running local deployments.
Products
3 itemsThe non-obvious takeaway is that de-identification works as a governance problem as much as a technical one: protection only holds if every pipeline runs the current version and compliant config, so Airbnb folded it into the shared workflow and model registry. Engineering teams handling user text for training can reuse the pattern of deterministic per-run faker replacement plus GLiNER for street addresses. Teams in ASEAN and China dealing with multi-language support data can note that metrics are tracked per language, with conservative fallbacks when language detection is unstable.
The notable detail is that Salesforce ships this internally with a human approval gate and playbook updates, making it a reference architecture for trustworthy ML-operations agents rather than a demo. Enterprise ML engineers and platform teams can borrow the workflow: Slack context retrieval, parallel backtests, and pull-request-based sign-off. It offers a concrete benchmark for what agentic debugging tools from vendors like Datadog or internal Copilot-style systems should deliver.
The split between top score (Claude Fable 5.2) and lowest cost (GPT-6.1 Terra) shows that leaderboard winners are not automatically the best buy for high-volume research workloads. Teams building research agents can fork the MIT-licensed repo, swap in their own questions and graders, and benchmark candidates under identical tool budgets instead of trusting vendor demos. It compares models using one search stack, so treat it as a measure of tool-use skill rather than a ranking of search engines.
Industry
5 itemsEpoch AI 和 Anthropic 两条能力指数都指向同一现象:编码全自动化后能力跳了约 1.7-2 倍,但之后没继续加速。想判断 2026 年模型节奏的人可以看看这条趋势断裂,比单看某家发布会的数据更客观。
A Chinese EV maker now designs its own 5nm dataflow inference silicon and ships it across an entire model lineup, a capability previously held only by dedicated autonomous-driving chip vendors. The von Neumann critique offers a concrete architectural reference point for anyone evaluating edge inference hardware beyond GPUs. Teams building in-cabin LLMs or agent features in vehicles across ASEAN and global markets should track whether the M100's unified-compute approach spreads beyond Li Auto's fleet.
OpenAI 第一次让 Apollo Research、METR 这些外部机构直接看未发布模型和源代码,等于把安全审查提前到发布之前。但报告删改权还在 OpenAI 手里,实际约束力有多大值得琢磨。做 AI 安全或关注监管的人应该看看这套审计机制怎么设计。
Nvidia这次直接砸钱解决封装瓶颈,说白了就是让Blackwell GPU出货更快。做AI采购的和算供应链的要盯紧,封装产能松了,等卡的时间可能真的会短一点。
腾讯在业绩会上给了小微的成本账:3000 万日活,季度成本会低于靠烧钱拉用户的元宝。WeLM-80B 只激活 30 亿参数来压推理成本,未来还能跑到手机端。关注微信 AI 入口和小程序调用机会的开发者值得看看。
Papers
5 itemsZ.ai 放出了 RSI-Prime 数据集,350 条实验轨迹专门教 LLM 怎么训练别的模型。亮点是决策级 loss masking,只让模型从高质量决策学习。在 PostTrainBench 上把 GLM-5.3 拉升了 31.95%,做自动化训练 agent 的人值得照着这个思路试试。
The decomposition shows declining accuracy under longer budgets is largely a chain-selection effect, not models corrupting answers mid-chain — so aggressive early-exit rules built on final-answer probes may target the wrong quantity. Developers implementing adaptive compute budgets and enterprises tuning inference cost versus accuracy on math-heavy workloads can use the commitment-verdict gap as a diagnostic before shipping stopping rules. The modest oracle gains (+1.8 to +2.5 points) suggest limited headroom in current stopping-rule research.
这篇对做数据分析的人很关键:四个前沿 agent 在零效应数据上最高 98% 的概率给你“显著”结果。它还给了个可操作的重放纠偏方案,能把假阳性压回 5%。如果你让 agent 跑回归、做 A/B 分析,先看看这篇再信 p 值。
写代码智能体或者长文生成经常跑崩的朋友可以看看这篇。作者测了 29 个开源模型,发现崩不是模型问题,是 top-p 在高温下候选 token 池爆炸,换成 top-nsigma 同样的 Qwen3.5 智体能从解 0 题变成解 60 多题,代码全开源直接能用。
做智能体或浏览器的开发者该看看:不靠提示注入,一个正常网页表单就能让 Claude、GPT 这类模型把用户隐私交出去。论文给的提交时校验方案能挡掉 96% 以上泄露,还不太伤任务成功率,可以直接借鉴到自己的 agent 设计里。
How-to
5 itemsDeepSeek 把搜索、读链接、OCR、语音这些能力全做成 API 了,还配套开源了网页 demo。作者拿自己的 RAG 项目实测接入,代码和踩坑点都写了出来。想给自己的应用加联网检索能力的开发者可以直接照着抄。
这条推文拆解了 Claude Code 团队让 AI 自己写代码又自己查错的完整流程,有 3 种测试方法、39 项检查清单可参考。想在自己的项目里跑多 Agent 并行开发的工程师,直接照抄这套验证思路就能用。
The 70% schedule-triggered figure reframes agent engineering from conversation design to operations: memory hygiene and notification triage become the real bottlenecks, not model capability. Developers building long-running agents can borrow the "dreaming" maintenance loop and the three-bucket triage pattern directly. Teams evaluating agent frameworks should compare how LangChain's Managed Deep Agents handle scheduling versus ad-hoc orchestration, since scheduled autonomy is where most value — and most human attention cost — accumulates.
核心启示是把验收标准写进提示词,这比堆砌指令更能约束模型输出的质量,也改变了团队与 agent 协作时的默认授权方式。经常用 Codex 做代码、研究或视觉资产的工程师可以直接把并行子 agent 规则固化进 AGENTS.md。相比 Anthropic 的 Claude Code 提示实践,OpenAI 这份指南更强调推理强度的分级调度,可作为团队内部提示规范的对照参考。
The non-obvious point is that Qualcomm is backing an open shared-stack approach to AI accelerators, echoing how LLVM standardised compiler support across vendors. Engineers evaluating non-GPU accelerators and chip vendors building toolchains should watch the keynote to judge whether MAX and Mojo can replace in-house stack development. Compared with vendor-specific SDKs like CUDA, the pitch targets portability across PyTorch deployments.