Sebastian Raschka 发布 8 个推理模型从零构建的编程视频，共 11 小时，用 Python 和 PyTorch 从预训练的 Qwen3 0.6B 模型出发实现评测、推理时扩展、强化学习和蒸馏。课程先构建数学答案验证器并在 MATH-500 上评测，再实现温度采样、top-p 过滤与自一致性，随后用简化版 GRPO 做 RLVR 训练，最后用 DeepSeek R1 生成的答案蒸馏小模型。内容配套其《Build a Reasoning Model (From Scratch)》一书，早期文本生成与评测示例可在 CPU 上运行，后续训练实验建议使用 GPU。
#教程/实践#推理#数据/训练
开发者 Gathin 用 Claude Code 通过 Wi-Fi 接入小米电视，停用 25 个系统组件并重写极简桌面，随后把 prompt、桌面源码和踩坑记录开源为 CleanYourTV，数天获约 1.37 万 Star。另一位 YouTuber Ansub Khan 的 TV Detox 项目实测让电视可用内存从约 370MB 升至 600—650MB。文章指出这类工具只能清理系统层预装包与官方桌面，无法清除流媒体 App 内的贴片广告和促销弹窗，并提到相机圈也在借 AI 恢复被隐藏的固件功能。
#现象/趋势#Agent#开源生态
#教程/实践#Agent#RAG
https://t.co/DfsTmI4zPE
#教程/实践#Agent#MCP/工具调用
Local machines had become a bottleneck for agentic mobile development at @Shopify, so we gave agents their own remote dev environments. They now author 61% of our merged PRs 🧵 https://t.co/UEohGFp9ql
#教程/实践#Agent#部署/工程
Big news: @Kimi_Moonshot is back in the top-4 labs in the Code Arena: WebDev with the release of Kimi K3.1 Pro! This model just landed #8 in the Code Arena: WebDev with 1692 pts, matching Claude Opus-5 (Max) at 70% lower cost. This is a 9 place jump (+37 points) from Kimi K3 (Max) at #17 with 1655 pts. Kimi K3.1 Pro improved from Kimi K3 across all categories. Highlights: - #7 in Brand & Marketing (up from #9) - #9 in Simulations (up from #20) Cost efficiency is another upgrade. Priced at $2/$10 per million input/output tokens, Kimi K3.1 Pro (Max) has a median cost per task of $1.23 (29% less than Kimi K3). See its placement compared to the Pareto frontier in thread. With open weights coming soon, Kimi K3.1 Pro would be the new #1 open model in the Code Arena: WebDev, overtaking DeepSeek-v4.1-Flash. Stay tuned! See @Kimi_Moonshot’s ranking by lab, and 3D generations from Kimi K3.1 Pro below in thread. Huge congrats to the @Kimi_Moonshot team on this release!
#评测/基准#Agent#编码
#评测/基准#安全/对齐#DeepSeek
#评测/基准#Agent#安全/对齐
#评测/基准#Agent#编码
#评测/基准#推理#部署/工程
SemiAnalysis 工程团队在 Trainium3 上实测了原生 PyTorch 训练栈 TorchNeuron 与 TorchTitan，用 Qwen3-32B 在 16 颗 Trainium2 上训练，eager 模式达 11,099 tokens/s、23.89% MFU，compiled 模式达 14,156 tokens/s、30.47% MFU。作者认为 Neuron Explorer 剖析器优于 Nvidia 和 AMD 的方案，但首次编译耗时过长，gpt-oss 训练配方首步编译用了 27 分 17 秒。FlexAttention 测试通过 96 个用例中的 89 个，7 个失败均因编译超时而非数值错误。
#行业动态#数据/训练#部署/工程
AWS 发布在 Amazon SageMaker AI 上做多轮强化学习（MTRL）的实操指南，以 Wordle 为例，把智能体与奖励函数部署到 Amazon Bedrock AgentCore，训练作业通过 GRPO 更新 LoRA 适配器。示例中 gpt-oss-20b 在默认推理强度下对留出单词的解题率从 61% 提升到 84%，训练 token 花费约 339 美元，AgentCore 运行时约 79 美元。文章还给出奖励设计的两处反转错误、并发下 token 失效、工具调用退化为纯文本等失败细节，并建议在最终服务的解码设置下做评测。
#教程/实践#Agent#推理
文章用房价预测公式解释模型本质是一个系数从数据中学来的数学函数，50、0.2、25 这类学来的数值就是参数，其中权重是连接强度、偏置是常数项。作者区分了训练与推理，并对比房价小模型、GPT-3 的 1750 亿参数和 Llama 2 70B 版本的显存占用，指出模型是算答案而非像数据库那样查答案。文中还提醒模型预测的是模式而非事实，预测质量取决于训练数据的质量与多样性。
#教程/实践#推理#数据/训练
#论文/研究#安全/对齐#多模态
Announcing Artificial Analysis Coding Agent Index v2.0: FrontierCode v1.1 from Cognition replaces DeepSWE, SWE-Atlas-QnA grading is revised, and Claude Opus 5.5 in Claude Code leads at 61.7 Today we are launching v2.0 of the Coding Agent Index. The Index evaluates models together with their agent harnesses, now across 290 tasks covering software engineering, terminal use, and repository Q&A, with the three benchmarks weighted equally. Changelog (Index v1.5 → Index v2.0): → FrontierCode v1.1 replaces DeepSWE v1.1. FrontierCode is a benchmark from @cognition with a private test set: 100 software engineering tasks on real open-source repositories, graded on whether a maintainer would merge the change. An attempt that fails a blocking check scores zero, and one that passes them all scores between 70% and 100% depending on the non-blocking checks it meets. We also scan each trajectory with a modified version of Cognition's rule-based scanner, and any attempt flagged for accessing solution material scores zero. → Revised SWE-Atlas-QnA grading, now SWE-Atlas-QnA-AA. Scale AI's 124 repository Q&A tasks are unchanged. We reworded 127 of the 1,306 rubric criteria for more consistent verdicts and moved to GPT-6 Luna (xhigh) as the judge, with five votes per criterion. The dataset and judge configuration are published on Hugging Face. Key results: ➤ Claude Opus 5.5 leads the Index: Claude Code with Claude Opus 5.5 (max) scores 61.7, followed by Claude Code with Claude Sonnet 5.5 (max) at 59.3, Muse Code with Muse Wonder Full (max) at 58.9 and Claude Code with Claude Fable 5.1 (max with fallback) at 57.6. ➤ FrontierCode v1.1 is where Opus 5.5 pulls ahead: Opus 5.5 scores 56.0% on FrontierCode, ahead of Codex with DeepSeek V4.1 Flash (max) at 52.6%, Codex with GPT-6.1 Sol (max) at 49.4% and Codex with GPT-6 Astra (max) at 49.3%. Below the top, FrontierCode scores are tightly packed: 15 of the 24 configurations on the Index score between 43% and 49%. ➤ SWE-Atlas-QnA-AA grading is stricter and more consistent: across the 21 configurations evaluated on both Index versions, the mean score falls 1.7 points, from 58.8% to 57.1%, with 14 of 21 moving down. Individual configurations move by up to 7 points: Muse Code with Muse Spark Open (xhigh) falls 7.0, Kimi K3 in Kimi Code CLI and Codex with GPT-6 Luna (max) 6.2, and Muse Code with Muse Wonder (max) 5.6, while Codex with GPT-6.1 Sol (max) gains 2.1. ➤ Four configurations sit on the cost frontier: Codex with DeepSeek V4.1 Flash (max) scores 47.3 at $0.39 per task, Codex with GPT-6.1 Sol (max) 54.1 at $1.50, Codex with GPT-6 Astra (max) 54.7 at $7.04 and Claude Code with Claude Opus 5.5 (max) 61.7 at $11.42. Costs are pay-per-token API costs, not subscription prices. We will keep updating the Coding Agent Index incrementally, with more evaluations and agents to come.
#评测/基准#Agent#编码
TinyFish 发布两份技术指南，一份定义生产级智能体搜索架构，一份总结深度研究系统的构建经验。架构基于 TinySearch、TinyFetch、TinyBrowser 和 TinyAgent 四个网页原语，围绕基础模型运行七阶段控制循环，涵盖任务定义、检索规划、工具路由、证据评估与失败重试。在固定 gpt-5.6-sol 模型和 100 个任务的 OpenBenchmarks 测试中，TinyFish 搜索加抓取完成率 79.0%，中位任务成本 0.050 美元、耗时 24 秒；纯搜索配置下完成率为 59.3%。
#论文/研究#Agent#搜索
robots.txt 只是一份抓取许可声明，不是访问控制，它只能约束愿意遵守的爬虫抓哪些路径，管不住不遵守的爬虫、已流出的内容和不经过爬虫的获取方式。文章按 RFC 9309 说明 User-agent、Allow、Disallow 和 500 KiB 读取上限等规则，并给出按 UA 屏蔽 GPTBot、CCBot、ClaudeBot、Bytespider 的 robots.txt 写法，以及用 nginx 按 UA 做差异化限速、用日志统计各爬虫请求量、用 curl 验证限速是否生效的做法。作者强调 robots.txt 表达意愿、服务端限速控制成本、日志看清事实、授权协议解决法律问题，四件事各管一段。
#教程/实践#部署/工程#数据/训练
OpenAI 公布两项基准结果：在自有的 COBOL 迁移测试中，GPT-6.1 Sol 在低推理强度下达到 86.6% 准确率，高于 GPT-5.6 Sol 的 5.4%，成本降低 39%，高推理强度下达到 97.9%。另一项 MultiLangBench 测试模型用 20 种语言从零构建兼容 SQLite 的数据库引擎，GPT-6 Astra 平均通过率 87.4%，GPT-5.6 Sol 为 26.8%，低成本版本 GPT-6 Luna 为 40.3%。OpenAI 已在 GitHub 发布 MultiLangBench 的设置与测试套件，但未公开 COBOL 迁移评测的测试用例，外部无法独立核查该结果。
#评测/基准#编码#OpenAI
#评测/基准#推理#DeepSeek
Anthropic 应用 AI 团队在多个企业部署对比中发现，带技能（skills）的单一智能体在质量上常优于子智能体架构，且单任务成本和延迟更低，并给出把子智能体折叠进主循环的迁移方法。文章指出子智能体仍适用于上下文膨胀导致性能下降、需要并行处理独立任务、或工具选择与指令冲突的场景，并建议先把子智能体作为工具调用，再逐个转成技能和 MCP 工具，用端到端评估对比前后表现。文中提到 Claude Code 面向更先进模型的系统提示词缩减了超过 80%，并配套提供 build-eval、prompt-audit、hillclimb、cost-optimize 等 Claude API 技能。
#教程/实践#Agent#MCP/工具调用