内容归档

DeepSeek 发布 V4.1-Flash-1017：做事更快、回答更短，官网和 App 已换新；联网搜索 API 每千次 5 元，11 月 1 日起收费 DeepSeek 发布 V4.1-Flash 的升级版 V4.1-Flash-1017（1017 指 10 月 17 日）。官方说它测试分数更高、完成任务更快、输出更短。输出短，消耗的 token（AI 计费单位，可以理解成文字碎片）就少，同一件事做得更好，也更省钱。 普通用户打开 DeepSeek 官网或 App 用的就是新版。开发者调 API 的模型名还是 deepseek-flash，代码不用改。 这一版主攻从零做完整个项目，演示里它从头做出了小游戏和 3D 模型。官方在 Hugging Face 公布的测试里，进步最大的是让 AI 在命令行里自己敲命令干活，从 31.2 分涨到 42.3 分。和 Anthropic 的 Opus 5 比，几项自动办事的测试它分更高，高难度问答和写完整程序仍然落后。这些分数都是 DeepSeek 自己测的。 另一件事跟开发者的钱包有关：DeepSeek 把自家搜索引擎开放成接口。Search 输入关键词返回网页结果，Fetch 给一个网址把网页内容抓成文字，合起来能让 AI 应用上网查资料。每个接口每 1000 次成功请求收 5 元（英文版标 0.75 美元），失败、没搜到、超时不收钱。 如果你把 DeepSeek API 接进 Claude Code、Codex 或 DeepSeek 自家的编程工具 DeepSeek Harness 当背后的模型，这些工具里 AI 自动发起的搜索和抓网页目前免费，北京时间 11 月 1 日零点（10 月 31 日 16:00 UTC）开始按次收费。 模型同时开源，MIT 协议，允许商用，可在 Hugging Face 和魔搭社区（ModelScope）下载。它有 7630 亿个参数，个人电脑跑不动，主要给企业和云服务商自己部署。 模型下载： huggingface.co/deepseek-ai/De… 搜索 API 文档： api-docs.deepseek.com/zh-cn/guides/s… DeepSeek @deepseek_ai 🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. 🛠️ Optimized for @DeepSeekHarness , Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model. 1/4 🔗 View Quoted Tweet

Banger paper from Tencent on recursive self-improvement for terminal agents. (bookmark it) This is a great read if you are tracking research on training terminal or coding agents with RL Starting from Qwen3.8-27B, they alternate two phases over five stages. One phase fine-tunes on rewritten versions of trajectories that passed the verifier. The other runs PPO with a critic inherited from the previous stage. The four-suite Terminal-Bench average goes from 20.68% to 50.58%, with 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard. The reward audit is just as useful. Under a dense progress reward, failed runs collected 50.09% of the reward, and 48 failures passed every executed check but still missed the final test. Separating strict success from partial credit cut that share to 28.72%. arxiv.org/abs/2610.22192 Chat with Paper: academy.dair.ai/papers/termina…

Classify: use the Decisions API to automatically detect an agent stuck in a loop, and halt or nudge it on thresholds you set. openrouter.ai/docs/cookbook/…

Looking for a weekend project? Two new OpenRouter cookbooks on @scaledown_team 's task-specific models (currently 20% off!), each a ready-to-run recipe for a workflow problem. Summarize: compact a 100-step agent session while keeping the facts it still needs. openrouter.ai/docs/cookbook/…

Arena Alignment Index update: Two new labs and three new models have landed! - Meta’s new top model is Muse Wonder, replacing Muse Spark 1.3. - Reflection debuts with Beam 1.0, the #1 US open model on Agent Arena overall. - StepFun, another Chinese lab, joins the leaderboard with Step 5 Preview. Arena Alignment Index rankings by lab: - #6 @AIatMeta : Muse Wonder (Max) at 77.5% - #7 @reflection_ai : Beam-1.0-501B-A23B at 75.3% - #8 @StepFun_ai : Step 5 Preview at 75.1% Among the three new model entrants: - Meta is strongest with the highest Alignment Index score (77.5%). It also has the lowest unauthorized action rate (2.09%) and deceptive completion rate (14.41%) of the three. - Reflection AI is strongest at avoiding false attribution: 4.50%. - StepFun does not lead a category, but its strongest relative result is unauthorized action at 2.11%, almost tied with Muse Wonder’s 2.09%. Arena.ai @arena Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key findings: - OpenAI models currently lead the Alignment Index - Rogue actions are rare, but can have serious consequences when they occur - Agents can mislead users about task progress - Misalignment risks increase with conversation length - Safety and alignment are improving across model generations As shown in the leaderboard below (sorted by lab), @OpenAI ’s GPT-6.1-Sol leads the Arena Alignment Index with a score of 87.9, followed by @AnthropicAI ’s Claude-Opus-5.5 at 83.2 and @spacexai 's Grok-4.7 at 82.7. OpenAI also has the best observed rates across all three signals: 0.89% Unauthorized Action, 1.98% False Attribution, and 2.34% Deceptive Completion. Across all four labs, newer models consistently outperform their predecessors, suggesting broad progress in agent safety and alignment. As agents take on longer, more complex, and higher-stakes tasks, measuring not just what they can accomplish, but how safely and reliably they act, becomes increasingly important. This marks an important step toward making safety and alignment a core part of how Arena evaluates AI. The index is an initial starting point, and we'll continue expanding the index with additional safety signals and models over time. More analysis below👇 🔗 View Quoted Tweet

Built from 90K+ real-world agent sessions across 30 models, Arena’s Alignment Index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions. - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence. - Deceptive Completion (DC): Claiming a task was completed when it was not. Dive into the Arena Alignment Index at: arena.ai/leaderboard/ag…

x.com/boltdotnew/sta… bolt.new @boltdotnew Bolt is free for EVERYONE, all weekend long. And you could win $10k! Turn your idea into a product. Submit it by Oct 20 for your shot at the prize. Bolt Builders Week starts now 🏁 Details here � bolt.new/builders-week 7G 🔗 View on Twitter 🔗 View Quoted Tweet

Built from 90K+ real-world agent sessions across 30 models, Arena’s Alignment Index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions. - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence. - Deceptive Completion (DC): Claiming a task was completed when it was not. Dive into the Arena Alignment Index at: arena.ai/leaderboard/ag…

AI: 鱼的游泳效率部分来自它把自身的运动和水中的漩涡耦合在一起，包括自己尾巴拍出来的漩涡。鱼并不只是在水中前进，鱼和水中的漩涡共同构成了一台游泳机器。 人和语言的关系类似：我们在语言媒介中制造局部扰动——说出一些词、写下一些符号——再利用这些扰动驱动内部思考过程。语言和外部符号是人类思维中最重要的「认知漩涡」。

Ben Horowitz on the lesson LoudCloud, his early cloud company, taught him about today's AI neoclouds, and why they now face the opposite problem: "It can end up being like a very dangerous arbitrage business, where you can get upside-down easily. And so if you look at LoudCloud, just as an example, here's a simple thing that happened to us." "In 1999, bandwidth was $1,500 a megabit. And you couldn't get it. It was very hard to get it. And so you had to get a long-term contract, three years. By 2001, it was $50 a megabit... And if you borrowed money to get it, you're dead." "That kind of thing is so out of the control of the entrepreneur." "I think the core dynamic of AI has been working for CoreWeave, though, in that there's a massive shortage, not an overage of supply... Token prices have gone down, but GPU prices have not." "And by the way, data center space is going up in price because even labor to build data centers is hard. The politics are making it hard to build data centers. So the whole supply chain price is going up and demand is not subsiding." "Those are very big tailwinds for CoreWeave, despite the fact that the debt markets, kind of independent of them, are tough." @bhorowitz w/ @chamath on @allinpod 🔗 View on Twitter

Tune in: youtu.be/X7oI4Umt3DU

Join us for the next Paper Club on October 21st, where we'll talk Inference: events.ycombinator.com/yc-paperclub-o…

This week’s Paper Club is all about AI safety. Not the headlines, but the science. AI safety has been a research field for decades, with its own conferences, methods, and a track record in systems like aircraft collision avoidance, where the standard is one catastrophic failure per billion flight hours. We'll hear from @ClarkBarrett7 and Mykel Kochenderfer ( @aiprof_mykel ), Stanford professors and co-directors of the Stanford Center for AI Safety. Together with @FrancoisChauba1 , they break down the OpenAI–Hugging Face incident, how formal verification can mathematically prove properties of code, how aircraft safety techniques are being adapted to language models, and why AI may need its own NTSB. 00:00 - Science fiction vs. science fact 04:32 - Engineering problem, or P(doom) 06:04 - What actually happened at Hugging Face 07:53 - Where "10% in 10 years" comes from 11:21 - The real risk: open weights in the wrong hands 17:25 - Clark Barrett: when AI met automated reasoning 20:35 - CSLib: proving code correct in Lean 24:32 - From English spec to verified code in one click 27:55 - Mykel Kochenderfer: how do we know a system is safe 32:11 - Adaptive stress testing, from aircraft to LLMs 36:57 - Verifying neural networks 40:30 - Runtime monitoring and BetterBench 43:53 - Panel: reacting to the past three weeks 45:55 - An NTSB for AI 52:18 - Is cybersecurity solvable in five years 55:00 - If you were CISO of Bank of America 1:03:06 - The Swiss cheese model 1:05:43 - Audience Q&A 🔗 View on Twitter

Bolt is free for EVERYONE, all weekend long. And you could win $10k! Turn your idea into a product. Submit it by Oct 20 for your shot at the prize. Bolt Builders Week starts now 🏁 Details here � bolt.new/builders-week 7G 🔗 View on Twitter

苏世民又是天才又很努力，可黑石集团刚创业第一年也是举步维艰 创业第二年的时候也很艰难，但妙在日本找到了机缘，最终从33家大公司那里为自己的第一期激进募集到了8亿美元 仅凭唇舌之力啊，人类还是太神奇了

The fun begins 🐠 We can't wait to hear what y'all cook up! Rissa Cao @rissa_cao Our proudest launch ever We're here to make voice AI accessible to everyone. Enjoy building with Drama 3! x.com/i/trending/211… 🔗 View Quoted Tweet

梁文锋滑动变阻器 - 梁祖 goDavi（狗大V国际版） @corndogjpn2 @geekbb 现在滑动变阻器在哪一档？？ 🔗 View Quoted Tweet

shipany.ai 模板新增 Guest Post 功能，可以为你的网站开启文章投稿来增加收入了。

I've been tricked several times by redials, but I won't knowingly buy one. They're different from restorations on paintings because (a) they're irreversible, (b) they cover the entire dial, and (c) they're almost always badly done.

Mole got 8 pages in iCreate 182, the largest Apple magazine in the Netherlands and Belgium, from the five planets all the way down to the little mole in the menu bar. Thank you Ward Edema and the iCreate team for going through every tab so carefully.

用 Grok Bot 和 CF 的朋友，请去直接安装这个 plugin，里面有 4 个工具（ 其实就是官方 mcp ）和一大堆官方的 skills，安完了以后几乎所有 CF 的操作都可以直接让 Bot 完成了。

Your users say search is bad. How do you turn that complaint into an eval? Aaron Linsky explains how Bridgewater built a benchmark of about 200 examples in under a week, then used it to measure search improvements. 59 seconds from AI Engineer New York. #AIEngineerNY 🔗 View on Twitter

One pitfall when buying old watches is "redials," or repainted dials. No reputable dealer will sell you one without telling you (maybe with euphemisms like "restored" or "refurbished") but in auctions it's mostly buyer beware. Compare a redial (left) and original (right).

第二集不是很满意，让opus重新修复技能 重做了一遍，节奏感更强了 英伟达崛起之路 第二集↓ 欢迎提意见 🔗 View on Twitter 小互 @xiaohu 炸裂了 我宣布 Opus 5.5现在是我的神... 我这套技能完全直接可以起号... 这个视频做出来的时候我直接热血沸腾了... 🔗 View on Twitter 🔗 View Quoted Tweet

怎么感觉国内 AI 团队们，有几个很久没发新模型了？ 本来以为是 AI 模型发布更新太快的时间错觉，查了一下，确实有几家超过 3 个月没发了。 字节：Doubao-Seed-2.1-Pro 6月23日发布，快 4 个月没发！ MiniMax：MiniMax-M3，6月2日发布，超 4 个月没发！ 月之暗面：Kimi-K3 6月13日发布，超 4 个月没发！ 百度：ERNIE 5.1 4月29日发布，快半年没发！

Full talk here youtu.be/DqF08Dz3nok

What have you been using?

Tolerating a problem has the same consequences as failing to identify it. Whether you tolerate it because you believe it cannot be solved, because you don't care enough to solve it, or because you can't muster enough of whatever it takes to solve it, if you don't have the will to succeed, then your situation is hopeless. #principleoftheday

"There are some people in your org who are doing way more exploring than others. They are your early adopters. They're running Astra and Fable on the weekends on personal projects, to find things they might bring to work. The problem is that they can be a massive f*cking distraction. So the key question to ask yourself if you're running a product org is: how do I harness these early adopters without distracting everybody else?" @danshipper 🔗 View on Twitter

DeepSeek 叕发布了一个「小」版本：DeepSeek V4.1-Flash-1017，在性能、速度、输出长度三个维度同时改进，同时将内部搜索引擎以 Search & Fetch API 的形式正式商业化，并深度绑定自家及 WorkBuddy 等第三方 Agent 编程工具生态。 模型发布：一次“小步快跑”式迭代 V4.1-Flash-1017 是对既有 DeepSeek-V4.1-Flash 的增量更新，采用日期后缀命名，属于快速迭代节奏下的版本推进，而非架构换代。 公告强调的三个改进方向「更强的基准表现、更快的任务完成、更短的输出」，第三点尤其值得注意。模型输出冗长是当前推理模型普遍的成本痛点：同样的任务，输出 token 越少，延迟和费用越低。“性能更高 + 输出更短” 同时成立，才是其所说的“性能与成本的帕累托改进”的真正含义。 训练方法：RL with self-feedback 新模型使用“带自我反馈的强化学习”来提升端到端项目开发能力（从游戏开发到 3D 建模）。这意味着训练目标不再是对话轮次的质量，而是完整任务的最终结果：模型自己生成反馈信号来修正多步骤、长周期的工作流。这与其在 Terminal-Bench 4.0（42.3）、DeepSWE v1.1（77.0）等偏重真实工程环境的基准上的表现定位一致，模型卡的对比对象也集中在 K3、GLM-5.3、GPT-5.6 等同期的 Agent/编程向模型上。 工具生态：抢占 Agent Harness 入口 为 DeepSeek Harness（需升级至 v0.2.1-rc.1 或更高）、腾讯 CodeBuddy、WorkBuddy 等做了适配。DeepSeek API 也直接兼容 Claude Code 和 Codex 的服务端工具调用。 Search & Fetch API：搜索引擎基础设施商业化 商业层面最重要的一条，DeepSeek 把支撑自身产品的内部搜索引擎开放为两个独立 API： · Search API（/v1/tools/search）：返回结构化网页结果，支持 site:、lang:、after: 等检索算子，可附带页面正文。 · Fetch API（/v1/tools/fetch）：抓取单个网页并转为 Markdown，这是 RAG 和 Agent 工具链的标准需求。 DeepSeek @deepseek_ai 🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. 🛠️ Optimized for @DeepSeekHarness , Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model. 1/4 🔗 View Quoted Tweet