有人把 Lex Fridman 全部 502 期播客做成了带中文翻译的文稿站,还能装成手机 App,免费开源,适合通勤时刷访谈。
Latest
The practical detail is that Dynamo serves through an OpenAI-compatible API, so existing client code can point at a Dynamo-backed endpoint without rewrites. Teams running inference endpoints who want Kubernetes-based autoscaling can use the walkthrough as a quick reference. It sits alongside alternatives like vLLM's Kubernetes deployment paths.
这条推文把未来工作流画得很扎心:AI 写 PR,另一个 AI 审 PR,人只负责说 LGTM,同一个模型自产自检,看完想笑又想哭。
DeepLearning.AI 的课程里加了浏览器内 AI 编程智能体实验,边学边写代码,实时看效果,不用装环境,免费报名。
Percy Liang's closing argument reframes the KV cache as the structural root of Transformer inference inefficiency, pointing to architectures designed for inference from day one, such as SSMs, linear attention, and diffusion-based non-autoregressive generation. Engineers building or tuning inference stacks get a single coherent map from architecture to serving systems. The executable lecture notes with sympy-verified formulas make the material directly reproducible rather than slide-level theory.
Stripe 首席设计官 Katie Dill 的演讲实录,讲 AI 时代怎么守住设计品质。MCP 读文档效果差改用 CLI 这个踩坑经验,做 AI 编程工具的都该看看。
有人把闲置的 rabbit r1 变废为宝,打算装 llama.cpp 跑本地模型还搞 s2s 应用,动手党可以参考。
你订阅了多个 AI 服务的话,可以看看这期 magpie 演示,统一配置加一键切额度,还和 CC Switch 对比了差异。
孩子要玩 128 块的打字 App,这位爸爸直接用 Codex 20 分钟自己写了一个部署上线,儿子一口气玩了 18 关。
作者把 DeepSeek V4.1 Flash 1017 从价格到编程基准都跑了一遍,还做了同题写作对比,想换模型前可以先看看这篇再决定。
有人把 Lex Fridman 几百期播客做成带文字稿的网站,没字幕的用本地语音识别补齐,明天就开源,想批量啃访谈的可以蹲一下。
The rotation strategy shows how heavy users stretch AI subscriptions beyond single-account limits, and tier selection between medium and high reasoning changes both speed and quality. Practitioners paying for multiple coding subscriptions can copy the setup, including routing Codex's computer use through Claude. The spending arc, from a S$20 plan to two US$250 accounts, is itself a datapoint on how fast coding-assistant budgets scale.
有人账号被封申诉无门,干脆用 Opus 5.5 做视频吐槽。他用 codex 做了三天,换 Opus 5.5 一下午搞定,视频里能看到效果对比。
有人拿日志实测了三种 Agents.md 的生成成本,Opus 5.5 那份只要 15.5M tokens、约 $0.25,价格敏感的话可以看看数字再决定。
有人在 WorkBuddy 上白嫖新版 DeepSeek 到月底,还教你怎么用自己的 API Key 接进去,搞办公自动化的可以试试。
The free model's win suggests cheaper open or free models can contribute meaningfully to frontier auto-research when paired with verification from a paid model like Claude Opus 5.5. The jump in research-referencing PRs shows that a shared prompt file can coordinate an army of agents better than brute-force exploration. Practitioners running multi-agent research workflows can test the context-engineering tricks cited here, such as forking and context flushing, to escape local minima.
两个开发者聊 AI 时代要不要选 Rust:编译慢拖累 agent 反馈,但高门槛反而筛掉了一锤子买卖项目,观点挺有启发。
写代码的 agent 完成正事之后开始无限纠结?这位用户在 AGENTS.md 里加一句防跑偏指令,挺实用的思路。
Paul Graham 发推吐槽 AI 回复总爱说词语在"doing work",底下讨论这种措辞癖好怎么来的,写提示词的可以看看
有人用两个 Claude Code Agent 互相发邮件、联系会计师,20 分钟把报税办了,流程值得看看
有人在 VPS 上跑了 yetone 的 Magpie 网关,Codex 的 Token 加 TRAE、Workbuddy 的免费模型都能接进来,等于自建 OpenRouter。
OpenRouter 出了两份能直接跑的 ScaleDown 模型教程,教你压缩 agent 会话还不丢关键信息,模型还打八折。
The session list doubles as a map of where PyTorch engineering effort is heading: communication-computation overlap for MoE training and dynamic-shaped CUDA graphs signal that large-model training performance is the priority theme. Engineers building or tuning training pipelines can use these talks to pick up concrete scaling techniques, including the 10x data loader speedup. The Ecosystem Landscape talk is also useful for open-source maintainers seeking official PyTorch ecosystem status.
想打 Enveda CASMI 2026 质谱竞赛的话,这份 notebook 把检索、碎片模拟、指纹模型排序全流程都跑通了,直接抄作业起步。
有人实测了 Devin Max 里的 Opus 5.5 额度,结果比 20 美元的 Claude Pro 还少 20%,买前可以看看这篇。
用户说搜索不好用,别急着改,先学 Bridgewater 用一周做 200 条评测集,改没改好直接量化。
Only the latest 2000 items are listed; for older ones see the daily archive