Latest

Every published item, in order · 29229 items
Nov 10
ModelsBuildersMulti-source13:35
Chinese contentClaude 发布 Fable 5.2 与 Mythos 5.2,Agent 能力再升级

Claude 这次一口气发了 Fable 5.2 和 Mythos 5.2 两个版本,主打 Agent 实操能力。最亮眼的是 Fable 5.2 能直接生成一个可交互的完整操作系统 Demo,不是截图那种糊弄的。做 Agent 开发或自动化的朋友可以看看这 10 个案例,对比下自己手上的工具。

Event
AI Will@FinanceYF5Source
How-toBuildersSelected13:23
Grok Bot vs ChatGPT: a product-atom analysis of agent design

The framework gives builders and product designers a vocabulary for comparing agent interfaces beyond feature lists: session, project and bot each shift where conversational continuity lives. Teams designing multi-agent products get a concrete design principle — Simon's near-decomposability — for splitting bot responsibilities. Practitioners choosing between ChatGPT, Grok and similar tools can use the atom lens to predict where their workflow context will accumulate before committing.

李继刚@lijigang_comSource
How-toBuildersMulti-source12:38
Chinese content当任务只需 5 行代码时,你的 subagent 会开始干什么

写代码时谁没见过 subagent 把 5 行的小事搞成大工程?这条视频吐槽的就是这个现象,评论区都在聊怎么控制 agent 的任务拆分粒度。如果你在用 Claude Code 或类似的多智能体工作流,看了会很有共鸣。

Event
Justine Moore@venturetwinsSource
Chinese content时尚品牌 Heliot Emil 用 Kling 4.0 制作 Core Uniform 系列宣传影片

快手旗下 Kling 4.0 这次不是发功能,而是直接交作品:帮哥本哈根品牌 Heliot Emil 做了整支时尚大片。能看出它的光影和织物质感到了什么水平,做视频生成或品牌内容的朋友值得看看成片。

Event
Official account快手可灵 Kling@Kling_AISource
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
ProductsBuildersSelected11:49
Browser Use 1.0.0 rewrites agent on raw CDP with Python REPL

The counterintuitive result is that removing structured action spaces and abstraction layers improved agent quality, cost and speed simultaneously — a design lesson that applies well beyond this one library. Developers maintaining Browser Use automations need to plan migrations now since 90% of the API surface is gone and pinning with '==' is advised. Compared with heavier agent frameworks like Playwright-based orchestration stacks, this raw-CDP approach sets a new baseline for what minimal web-agent scaffolding can achieve.

Browser Use@browser_useSource
ProductsBuildersSelected10:44
Meta's JiTTests generates failing tests to catch bugs pre-merge

Rather than accelerating code review, the system shifts AI verification to proving behavior changes with executable tests, a mechanism competitors like GitHub Copilot and CodeRabbit do not yet automate end-to-end. Teams running large-scale code review pipelines can adapt the weak-to-strong capture filtering and LLM judge panels to control false-positive noise. Meta's figures are self-reported without independent controls, so treat the 60% true-positive rate and 20% incident reduction as directional benchmarks. Engineering leads evaluating pre-merge defect detection should study the two-workflow design.

shao__meng@shao__mengSource
How-toBuildersMulti-sourceSelected10:44
Anthropic guides developers on choosing between Claude Opus 5.5 and Fable 5.2

The counterintuitive finding is that effort level, not model choice, is often the cheaper lever — a low-then-high retry strategy beat always-high on both cost and accuracy in Anthropic's own benchmarks. Teams building with Claude Code or the API should adopt the start-with-Opus, downgrade-for-crated-tasks, upgrade-only-after-effort workflow and re-measure on their own workloads. Cost per completed task, not per token, is the metric that matters when budgeting agent pipelines.

Event
shao__meng@shao__mengSource
How-toBuildersSelected10:43
Airbnb fine-tuned Gemma4-E2B classifier matches frontier quality at 1% cost

The counterintuitive result is that longer prompts cannot close the quality gap once a task has a latency budget — task-specific fine-tuning beats prompt engineering for enforcement workloads. Teams running high-volume moderation, compliance or routing classifiers should treat sub-billion-parameter fine-tuned models as the default baseline rather than defaulting to APIs. The published break-even threshold (15–30 calls per minute) gives engineering teams a concrete decision rule for build-versus-buy, and the latency and cost methodology is directly reusable.

shao__meng@shao__mengSource
Chinese contentvikingmute:AI 让独立开发更容易,但赚钱更难了

独立开发者 vikingmute 提出了个扎心观察:AI 让做软件空前容易,但供给指数级上升后付费的人反而在缩小。他想说的不是技术,是 AI 时代普通开发者的赚钱困局和阶层分化。如果你在做独立产品或打算 vibe coding 创业,这篇值得看看再决定要不要辞职。

Viking@vikingmuteSource
ModelsUS vs ChinaMulti-source09:39
Chinese contentClaude 发布 Fable 5.2 与 Mythos 5.2,部分任务成本仅 Opus 5.5 一半

Claude 一次发了 Fable 5.2 和 Mythos 5.2 两款模型,官方说编码、推理、视觉都比 Opus 5.5 强,部分任务成本还便宜一半。经常用 Claude 写代码、跑技术项目的朋友可以直接对比下成绩单再决定升不升级。

Event
AI Will@FinanceYF5Source
ModelsBuildersMulti-source09:38
Chinese contentKevin Ngo 展示用 Claude Fable 5.2 制作的浏览器可玩船游 Demo

有开发者用号称 Claude Fable 5.2 的模型做了一个乘船逛 8 个梦境世界的网页小游戏,打开就能玩。不过这个模型名我在官方渠道没查到,大家先当社区演示看。想对比各家模型做游戏 Demo 能力的可以留意。

Event
AI Will@FinanceYF5Source
Chinese content从 IMO 竞赛圈到 AI 创业核心:Alexandr Wang 与 Scott Wu 的人才网络

Alexandr Wang 和 Scott Wu 这批 IMO、IOI 出身的竞赛选手,如今分别带着 Scale AI 和 Cognition 站在 AI 行业前排。这条回顾说明了顶级竞赛圈层如何在创业前就埋下人才网络,对关心 AI 人才流动和创业生态的人很有参考价值。

AI Will@FinanceYF5Source
ModelsUS vs ChinaMulti-source09:36
Chinese contentAnthropic 发布 Claude Fable 5.2 与 Mythos 5.2,提速 30% 并下调百万 token 价格

Anthropic 这波把百万 token 价格砍到 $8/$40,还号称提速 30%。8 天写操作系统、30 小时证费马大定理这两个演示听着很猛,但都是官方自述没第三方验证。做长任务的团队可以对比下 Opus 系列的价格差再决定。

Event
小互@imxiaohuSource
ModelsUS vs ChinaMulti-source09:26
Chinese contentClaude Fable 5.2 开放所有付费套餐,价格降至 8/40 美元每百万 Token

如果你在用 Claude API,这个更新直接省钱:同样的调用价格降了两成。做文档处理、代码生成这类大 Token 消耗的团队收益最明显,建议测一下 5.2 在自己任务上的表现再决定要不要切换。

Event
AI Will@FinanceYF5Source
Chinese contentAlexandr Wang 回忆 2017 年与 Scott Wu 的午餐:竞赛圈群体如今遍布 AI 领域

Scale AI 的 Alexandr Wang 发了条回忆帖:2017 年他和 Scott Wu 吃饭时聊竞赛圈能不能复制 PayPal 黑帮,十年后这帮 IMO/IOI 队友创办了 Cognition、Pika 等公司。想了解这批 AI 创始人共同的数学竞赛成长路径,这条帖子很有意思。

AI Will@FinanceYF5Source
How-toBuildersMulti-source09:22
Chinese contentRiley Brown 演示用 Claude Opus 驱动 Blender 做 AI 视频后期工作流

Riley Brown 用 Claude Opus 加 Blender,一句提示词 10 分钟内做完了以前要 Screen Studio 加人工剪辑才能出的产品演示视频。和 Veo、Sora 那种直接生成画面不同,它是让模型操控剪辑软件做后期,结果可改可控。经常要做 App 演示视频的开发者可以照着搭一套。

Event
AI Will@FinanceYF5Source
Chinese contentnuance_ai 发布可实时对话的 AI 虚拟形象,延迟与拟真度显著提升

nuance_ai 刚上线一个能实时语音对话、带肢体动作的 AI 虚拟形象,演示里说话节奏和表情已经接近真人。做虚拟主播、数字人客服的团队可以看看这条演示,对比自家方案的延迟和拟真度。

Event
Justine Moore@venturetwinsSource
Chinese contentRobbyant 发布开源触觉基础模型 LingBot-Touch,预训练数据近 5000 小时

Robbyant 开源了一个触觉基础模型,用近 5000 小时真实触觉数据预训练,走的是 JEPA 那套潜空间预测路线,论文里还和 MAE、VAE 对比过。做机器人灵巧操作或具身智能的团队可以直接拿来用,跨传感器和跨灵巧手的迁移能力是卖点。

elvis@omarsar0Source
ModelsUS vs ChinaMulti-source91°08:34
Chinese contentAnthropic 发布 Claude Fable 5.2 与 Mythos 5.2,编程推理视觉全面升级

Anthropic 这次发了两个同底层的新模型:Fable 5.2 开 Xhigh 档在 FrontierCode 2.0 拿了 49.1% 的最高分,成本还只有 Opus 5.5 的一半;Mythos 则要申请才能用。做复杂长任务的团队值得看看,日常活官方还是推荐 Opus 5.5。

Event
shao__meng@shao__mengSource
ModelsUS vs ChinaMulti-source08:19
Chinese contentClaude Fable 5.2 (xHigh) 登顶 Code Arena: WebDev 榜首

Anthropic 又发新模型了,Claude Fable 5.2 在 Code Arena 的 WebDev 评测拿了 1832 分,比自家的 Opus 5.5 和 Google 的 Gemini 4 Carbon 都高。关键是价格还降了 20%,速度更快 30%。平时用 Claude 写代码、做网站的朋友可以直接换新版试试。

Event
lmarena.ai@lmarena_aiSource

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