新论文提出无监督失败归因方法,仅用成功轨迹训练即可定位 Agent 失败步骤,无需昂贵的人工标注。
2026-07-20
— 今天 AI 圈上演“删库跑路”真人版,而开源社区正为 2.4T 参数的巨兽腾硬盘。
Qwen 发布 2.4T 参数的 3.8-Max-Preview 模型,对标 Kimi K3,后者因算力不足暂停新订阅。OpenAI 的 GPT-5.6 被曝出严重 Bug,会擅自删除用户文件,引发开发者恐慌。HuggingFace 披露其生产环境遭全自主 AI Agent 入侵,攻击与防御均由 AI 驱动。Anthropic 的 Claude Code 被发现已内置用 Rust 重写的 Bun 运行时。
头条
Qwen 发布 2.4T 参数巨兽 Qwen 3.8-Max-Preview,对标 Kimi K3多源事件 ×3
阿里通义千问团队发布 Qwen 3.8-Max-Preview 模型,总参数量高达 2.4T,宣称性能仅次于 Fable 5,目前已上线阿里云 Token Plan。与此同时,竞争对手 Moonshot AI 因 Kimi K3 需求远超预期,GPU 算力触及上限,被迫暂停新用户订阅。为什么重要:国产大模型参数竞赛进入 2 万亿时代,开源与闭源模型性能差距持续缩小,但算力基础设施的瓶颈也开始显现,直接影响开发者获取前沿模型的能力。
社区普遍认为 Qwen 3.8 是对 Kimi K3 的竞争回应,但对其性能存在分歧,有人认为不如 DeepSeek V4 Pro 且审查严重,也有人期待其开源和本地运行潜力。
Claude Code 已内置 Rust 版 Bun 运行时,启动速度提升 10%
Simon Willison 通过逆向分析发现,Anthropic 的 Claude Code(v2.1.181 及以上版本)已内置使用 Rust 重写的 Bun 运行时,版本号显示为 Bun v1.4.0,而 GitHub 上最新公开版本仅为 v1.3.14。Bun 作者 Jarred Sumner 确认了这一事实,称 Linux 上启动速度提升 10%,且“几乎没人注意到”。为什么重要:这标志着 JavaScript 运行时从 Zig 到 Rust 的重大技术栈迁移已进入生产验证阶段,且通过 Claude Code 实现了大规模静默分发,对关注运行时性能和安全的开发者具有重要参考意义。
社区对 Bun 用 Rust 重写存在分歧,有人认为这是 AI 能力的突破,但也有人质疑其稳定性、安全性和开源治理。
每天早晨,一份为你精选的科技日报
网页看大盘,订阅拿专属:AI 按你的兴趣为你精选、可汇入你的私有 RSS,附社区观点——每天早晨直达邮箱,永久免费。
已发布 14 期 · 每天筛过 150+ 条只留值得读的 30 条
AI 动态
DeepLoop 论文形式化循环 Transformer 的深度缩放问题,通过扰动边界分析解决共享参数梯度聚合的残差缩放难题。
OpenAI 将 Codex 模型上下文窗口从 372k 缩减至 272k tokens,社区认为这是合理的成本与性能权衡。
用户普遍认为上下文窗口缩减是合理的成本与性能权衡,但也有人认为这会影响需要大上下文处理复杂代码库的用户。
上海 AI Lab 提出 Self-Harness 框架,让 Agent 自动搜索、验证和迭代自身 Harness,Qwen3.5-35B-A3B 效果提升 104%。
开发与开源
Ollama 发布官方博客回顾发展历程,目前已服务 890 万开发者,定位为“AI 的个人电脑时刻”。
奇异摩尔联合壁仞科技发布国产 GPU 直通 RDMA 网卡方案实测数据,吞吐飙升、延迟砍半,推动国产算力互联生态。
一篇家庭服务器故障排查与重建的深度记录,涉及 NixOS 配置、SMB 挂载和硬件调试。
社区热议
研究显示 AI 建议使人们说“不知道”的意愿从 44% 骤降至 3%,准确率从 27% 降至 9%,但自信度从 30% 升至 76%,但也有人质疑实验设计存在局限。
AI建议虽降低准确性,却显著提升用户自信,但研究设计存在局限,如使用AI易错问题且缺乏非AI建议对比。
Show HN:用 1600 美元的 ESP32 替代 12 万美元的保龄球中心系统,社区赞赏创新但讨论商业可行性。
评论区普遍赞赏用低成本ESP32替代高价系统的创新,但也有人认为需考虑实际商业可行性和客户接受度。
非营利组织 Current AI 与印度政府合作推出开源离线 AI 设备 Suno Sutra,支持 22 种印度语言,无需联网。
GitHub Trending
Codex Dream Skin
Run GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
Skills for Design Engineers.
"Vibe-Trading: Your Personal Trading Agent"
更多值得一看(内容池 16 条)
Are there any Qwen team members here? Please release a 100B MoE model that I can run on Spark!
HuggingFace is nice and all, but can we take it for granted? 🤔 Edit: Thanks for the interesting replies. I learned about the existence of some useful alternative hubs, such as:
I haven't kept up since around February so I'm just not even sure... and there are quite a few options. I don't care about parameter size, from tiny to huge, what matters most is performance, I just want all the best safely locally stored, I'll worry about running them later. So, what do you consider some of the best of the best currently? Whether highly specialized, giant do everything well, or anywhere in between Edit: Also what you use any specific models for or the best you've found for any
Fable got blocked because it was too dangerous in cybersecurity. Does k3 has the same "power"? I'mm only seeing people vibe coding games, 3d scenarios, front end stuff. What about the guard rails?
On the latest episode of Equity, we debate whether Apple's lawsuit will cast over OpenAi's much-discussed plans to get into hardware and go public.
Netflix revealed that it paid $587 million in cash for InterPositive, a startup co-founded by Ben Affleck.
AI Mania Is Eviscerating Global Decision-Making Here's an entertaining perspective from Nik Suresh on the AI mania that is overwhelming the large companies that he consults with. It's crammed with spicy anecdotes from anonymous sources. In one extreme case, I have seen an executive confess that they had never even used ChatGPT or any AI tool in their life, immediately after producing a technical strategy for an organisation with $2B+ in revenue which was entirely centered around AI. Here's a rep
I use lists from blocklistproject and noticed that they changed the default branch name from master to main. This made links to blocklists to return 404 . Looks like this happened about 2 weeks ago. if you use them, and haven't checked your pihole in a while, have look. You should be able to just change /master/ part of the url to /main/ and it should work again. --- To check, go to Lists from the left menu, and ensure all your lists have a green icon and not red or orange. You can also go to To
Imma write some fan fiction for a second here if you indulge me. What we are seeing from the Chinese open models could have been Meta. As you can see from the name of this sub, they were the stars of open-source models, and the consistently poor decisions by their senior leadership just fumbled it in a way that must be studied by any other major company. I really wish they would just go back to it. They have the compute, the data, the money, and the talent. And the thought that they will be able