Latest

Every published item, in order · 5279 items
Nov 14
How-toBuildersOfficialSelected15:06
Gemini 4 Carbon Distills 30 AI Insights from Next Token

The distilled list doubles as a practical mental model for building AI products: own your context, treat tokens as a cost of goods, and design for agent-first architecture. Developers and indie builders deciding what to ship next can use the 30 points as a checklist against their own roadmaps. Readers who follow Chinese AI discourse will recognise the named contributors and can trace each insight back to its original newsletter issue.

Event
Primary source歸藏(guizang.ai)@op7418Source
How-toBuildersSelected12:26
Dotey argues agent UX needs hybrid of main chat and sessions

The framing suggests the next round of differentiation in agent tooling will come from orchestration interfaces, not model quality — a hybrid of routed messaging and task drill-down. Developers designing agent platforms and teams comparing Claude Code with Dots or Grokbot should use this two-axis lens when choosing or building their own coordination layer. Multi-agent workflow teams in ASEAN and China face the same trade-off and can adopt the coordinator-plus-drill-down pattern directly.

宝玉@doteySource
Chinese contentClaude 团队发布智能体架构说明:收件箱、计时器、记忆、权限、发件箱五部件拆解

Anthropic 把智能体架构拆成五个部件讲透了,最点睛的一句是:始终在线是体验设计,不是让模型一直挂着烧 token。你用队列和定时任务拼过的那些东西,这份文档帮你把职责边界理清楚了。自己搭 agent 的工程师都该看一眼。

shao__meng@shao__mengSource
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
How-toBuildersMulti-sourceSelected09:52
Anthropic Explains Five Building Blocks of Always-On AI Agents

The event-driven pattern means an assistant can sit dormant instead of polling a model continuously, cutting cost without cutting capability. Builders of proactive assistants, notification bots, and workflow automation should study how inbox, timer, memory, permissions, and an outbound queue combine into a safe loop. The advice to start with low autonomy and human approval on important actions is a practical guardrail for anyone deploying agents against real user accounts.

Event
宝玉@doteySource
Chinese contentAgent 自主决策如何纠错:四条降低无效行动成本的设计思路

这条推文把主动服务型 Agent 最头疼的问题——错误决策拖出大量无效行动——拆成四条可落地的设计方案,从早期决策、分支推演到自验证和人机分工。做智能体产品的开发者能直接拿这套框架对照自己的系统,尤其适合正在做 planning 和工具调用调优的团队。

小胡子哥@Barret_ChinaSource
How-toBuildersOfficial06:37
Chinese content用 NVIDIA NeMo Switchyard 在 Amazon Bedrock 上按任务路由编程 Agent 请求

编程 Agent 跑任务,真正该看的指标是每解决一个任务花多少钱,不是每 token 单价。这篇教程手把手用 AWS CDK 部署 LiteLLM 加 NVIDIA NeMo Switchyard,把请求路由到 Bedrock 上的不同模型,再和单模型基线比成本。自己搭 Agent 团队、想在强模型和便宜模型之间省钱的开发者可以直接照着抄。

Event
Primary sourceAWS Machine Learning Blog@Aswath Ram A SrinivasanSource
Chinese contentRipple-65B Preview 架构拆解:从零构建小型编程模型

一场关于从零训练编程模型的技术直播,主讲人会讲清楚为什么不用 MoE、哪些消融实验失败了,还有 RL 训练里踩过的坑。自己做模型微调或 RL 训练的工程师值得看,这类失败经验比论文里的成功结果更少出现。

Official accountTogether AI@togethercomputeSource
How-toBuildersOfficial06:10
Chinese contentAmazon Bedrock AgentCore V2 快照机制与四种会话冷启动优化模式

AWS 把 AgentCore V2 的快照启动机制讲透了,还给了四种把初始化工作挪出首个请求的模式。部署智能体总被冷启动拖慢的团队可以直接照着改架构,比翻文档自己摸索快得多。

Primary sourceAWS Machine Learning Blog@Evandro FrancoSource
How-toBuildersOfficialSelected03:39
AssemblyAI cuts speech-to-text inference costs 70% on AWS Spot

The SQS retry-on-interruption pattern is the transferable piece: it shows how to make latency-sensitive GPU inference safe on Spot rather than batching only. Teams running speech, audio or vision inference on AWS can replicate the ECS-plus-SQS design to cut GPU spend sharply. The multi-Region European deployment detail also matters for anyone planning SLA-backed inference outside the US. Practical reading for platform engineers weighing Spot against on-demand GPU fleets.

Primary sourceAWS Machine Learning Blog@Belair AlbahagwiSource
How-toBuildersOfficial02:26
Chinese content用 Amazon SageMaker HyperPod Data Capture 搭建模型改进飞轮

AWS 把线上推理数据捕获和 HyperPod 微调打通了,客服路由这个例子从发现问题到量化提升全程走完。如果你每天在维护生产推理服务、苦于没有系统的 bad case 收集流程,这篇的落地方案可以直接抄。

Primary sourceAWS Machine Learning Blog@Shantanu TripathiSource
Chinese contentStanford 教授 Marco Pavone 自主机器人教材与课程资料免费公开

Stanford 教授 Marco Pavone 把自主机器人的整本教材加课程资料全放网上了,一分钱不收。想入行机器人或自动驾驶的人不用再到处拼凑资料,跟着 Stanford 正课的讲义走一遍就行,比大多数零散视频教程系统得多。

Stanford HAI@StanfordHAISource
How-toBuildersMulti-source01:02
Chinese contentsmolmaxxing:让1G内存的旧电视盒子和iPad也能本地跑大模型

家里吃灰的电视盒子、旧iPad现在真能跑本地大模型了,1G内存都行,实测token速度都贴出来了。用的还是Qwen3.5、Ling-mini-2.0这些模型,源码在GitHub上,想白嫖旧硬件玩本地推理的可以照着改。

Event
Gorden Sun@Gorden_SunSource
How-toBuildersSelected01:00
SparseRead Cuts Agent Reading Tokens by Up to 92.9%

The shift from post-hoc context compression to pre-read filtering changes how agent harnesses should be designed, and the reported 92.9% token saving suggests the approach scales beyond toy demos. Builders working on coding agents, RAG pipelines, or long-horizon automation can adapt the intent-locate-budget loop directly into their own tooling. Compared with simply adding search tools or summarisation layers, SparseRead treats reading itself as a budgeted, task-driven control problem.

AIGCLINK@aigclinkSource
Chinese contentCamThink 在 NeoMind 平台内置 MiniCPM-V 4.7 1B,实现边缘视觉场景理解

适合做边缘视觉的工程团队看:CamThink 把 MiniCPM-V 4.7 1B 塞进 NeoMind 平台,相机拍照直接触发 Agent 数人头、查违规、读表,全程本地推理不出内网。相比上云方案省了隐私顾虑,流程和提示词配置思路可以直接抄。

OpenBMB@OpenBMBSource
How-toBuildersSelected00:47
Claude Code adds /doctor prompt-audit command

The shift here is treating prompts as maintenance debt: rules that once tamed older models actively hurt newer ones, so instruction files need periodic audits like dependencies. Developers and teams running long-lived Claude Code setups should run /doctor prompt-audit (on 2.1.283+) and review flagged stale paths, renamed commands and conflicting global versus project rules. The matching prompt-audit inside the claude-api skill extends the same check to API prompts and tool descriptions for custom agent builders.

Event
berryxia@berryxiaSource
Chinese content深度体验 EnConvo:一个对话框覆盖编码、转写、视频下载多场景

做效率工具的朋友可以看看这个 EnConvo,一个框就能调 Claude、ChatGPT、DeepSeek、Grok 干活,编码、转写、下视频都在里面搞定,省得装一堆软件来回切换。想在 mac 上统一管理 AI 模型的话可以先看看这篇体验再决定要不要装。

Event
berryxia@berryxiaSource
How-toBuildersMulti-source00:39
Chinese contentlevelsio 用 Three.js 光照探针重构行走视频,考虑转型整个产品

levelsio 用 Three.js 光照探针把 AI 生成的行走视频重构出了不错的画面质感,全程只花 5 分钟跑在一台 VPS 上。他说效果比普通的图进图出换风格强得多,甚至打算把 fly.pieter.com 整个产品转向这个玩法。做 3D 场景生成或独立开发的人可以看看他这套流程。

Event
Pieter Levels@levelsioSource
How-toBuildersMulti-source00:35
Chinese content开发者分享半年使用体验:Claude Code + Opus 5.5 成为日常主力,Codex 在计算机操作上仍占优

这是一线开发者的真实对比:Claude Code 加 Opus 5.5 写代码更强,Codex 则在 computer use 上更稳,两个新玩具 ChatGPT Dot 和 Claude Duet 也各有看点。正在纠结选哪套编程工具的人可以参考。

Event
kimmonismus@kimmonismusSource
Chinese content在手机和笔记本上远程管理 Mac Mini 里的常驻智能体,这些工具哪个好用?

作者想用 Mac Mini 跑常驻智能体,手机远程控制,把 Raft、Teamily、Kylon、AgentHive 四个工具挨个试了一遍,优缺点说得挺细。想在家搭 always-on 智能体的可以照着挑,也欢迎在评论区补充别的方案。

Xianbao Qian@Xianbao_QIANSource
Sep 17
AITOP16:09
Anthropic拟10月在新加坡设亚太第五个办事处 开...

Anthropic拟10月在新加坡设亚太第五个办事处 开放本地招聘 数据显示,按人口比例计算,新加坡的Claude.ai使用水平在121个国家中排名第二。 我们观察到,人工智能公司Anthropic计划于10月在新加坡设立其亚太地区的第五…

AITOPRead
Aug 13
Aug 7
AITOP17:04
出口结构剧变:中国卖给世界的,不再是产品而是解法

中国出口"新新三样":从打火机到爬墙机器人,世界不再只是买便宜货 我们观察到一组清晰的数据信号:中国出口的"新新三样"已经从过去的服装、家具、家电,换成机器人、AI模型和创新药物。这不是产业升级的修辞游戏,而是发生在澳大利亚和巴西的实证样…

AITOPRead
Aug 2

Only the latest 2000 items are listed; for older ones see the daily archive