— GPT-5.6 刷屏，Llama 社区百花齐放，国产算力迈入十万卡时代。
GPT-5.6 一小时解开 50 年数学猜想，医生评价其回答错误少于人类。OpenAI 全量上线 GPT-Live 并翻倍语音配额，同时进军家庭场景。中国首个十万卡国产算力集群落成。PgBouncer 通过 so_reuseport 将吞吐量提升 4 倍。CISA 披露自己都未准备应急手册。
OpenAI 研究院 Ethan Knight 宣布，GPT-5.6 Sol Ultra 在不到一小时内完成了存在半个世纪的循环双覆盖猜想证明，通过 64 个子 agent 并行将原本可能一整天的任务压缩到一小时。同时，Sam Altman 引用医生研究称，医生认为 GPT-5.6 回答的错误比医生手写回答更少。GPT-Live 已对所有 ChatGPT 用户全量发布，本周末语音配额翻倍。
Reddit 社区对数学证明反响热烈，多位研究者称新 Prompt 技巧（不替模型规定解法、钉死验收标准）值得学习。也有用户调侃 Sam Altman 用“Elon 又迷恋我”来证明模型最强。
OpenAI 正招聘面向家庭、看护者及老年人的产品经理，ChatGPT 用户中 35 岁及以上占比从 26% 升至 31%。同时，arxiv 论文《Persona Cartography》提出用 OCEAN 框架在权重空间中测量和控制 LLM 的个性特质（开放性、尽责性等），训练低秩适配器来抑制或放大特定人格。
Hacker News 评论认为 AI 个性控制对安全性重要，但也担心人格操纵伦理风险。
美国网络安全局 CISA 在 7 月 10 日的复盘报告中承认，5 月一名承包商将敏感密钥暴露在公开 GitHub 仓库后，CISA 没有预设应急手册，团队在事件初期“不得不花时间编写”。报告强调应提前准备所有预期场景的应急手册。
Hacker News 评论批评联邦机构安全演练不足，称“连 CISA 自己都做不到合规”。
Qwen3.6 35B-A3B 在 Q8_0 CPU 模式下生成飞行模拟器 HTML，用户称“远超出它的体量”。
Qwen3 30B A3B 在 RTX 5060 Ti 上以 Float8 跑出 50 tok/s，比 llama.cpp 快约 50%。
Tencent Hy3（295B MoE）和 NVIDIA Nemotron-Labs-Audex-30B-A3B 的 GGUF 量化版本发布，含完整基准数据。
Infinity-Parser2 发布：可控数据合成管线+多任务强化学习，实现端到端文档解析，开源 500 万样本双语语料。
Agentic Neural Architecture Search：LLM 生成种子架构并分解为“槽位架构”，自动定义任务特定搜索空间。
ShipThatCode：通过从头构建 Redis、Git、数据库等项目学习系统编程，支持 9 门语言。
Ant：自研 JavaScript 引擎的轻量级运行时，9MB 二进制实现 near-V8 速度，兼容 npm 生态。
推荐 SQLite 默认启用 STRICT 表：避免类型混乱，多数 Hacker News 评论支持默认严格。
多数评论支持SQLite默认启用严格表模式，认为动态类型易导致数据错误，但也有人认为非严格模式在嵌入式场景下更灵活。
《从第一原理理解网络与互联网》长文，以信号转换视角解释通信全过程。
评论区多数人认为文章内容优质、解释清晰，但也有人认为其风格或动画可能受AI影响。
MIT 学生开源单文件 LLM 服务仪表盘，无框架依赖，支持 llama.cpp 和 vLLM。
Politico 报道美国 tech 行业对中国开源 AI 模型崛起担忧，讨论特朗普政府是否再发行政令。
《谁来管理 Agent？》文章引发热议：多数评论认为 AI 代理需要人类管理，但管理权不应被少数精英垄断。
评论区普遍认为AI代理需要人类管理，但管理权不应被少数精英垄断；也有人担忧企业控制会取代开放共享。
Apple 起诉 OpenAI 挖角硬件高管并涉嫌窃取电池/SIP 等零件细节，社区称“细节太野”。
用户成功将 Qwen3.5 35B A3B 跑在 4GB VRAM 的 RTX 2050 上（仅 1.2 tok/s），寻求加速建议。
用户用四块 RTX 5060 Ti 跑 Qwen3.6-27B 做代码生成，称效果非常好。
Reddit 经验帖推荐高信噪比技术博客：Brandur (Go/Ruby/Postgres)、Stripe 博客等。
Skills for Design Engineers.
OpenWiki is a CLI that writes and maintains agent documentation for your codebase.
AI-powered job application framework built on Claude Code. Fork it, fill in your profile, and let Claude evaluate jobs, tailor CVs, write cover letters, and prepare you for interviews.
The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents.
I ran Grok Build CLI (v0.2.93) through mitmproxy. It uploads your entire repo as a git bundle (full history) to xAI's Google Cloud — independent of what you open. With the prompt literally "do not read or open any files," a file I planted came back verbatim when I git clone -d the captured upload. Separately, files it reads (incl. a .env with API_KEY / DB_PASSWORD ) go to cli-chat-proxy.grok.com verbatim. Turning off "Improve the model" doesn't stop it — that toggle governs training, not upload.
Ant Group's Robbyant has released the LingBot-VA 2.0 technical report — a Physical AI video-action foundation model built from scratch for embodiment rather than fine-tuned from a video generator. It predicts future states ahead of execution through Foresight Reasoning, re-grounds on every real observation, and reaches 225 Hz asynchronous control. We break down the causal DiT, the sparse-MoE video stream, the semantic visual-action tokenizer, and where the paper's own numbers don't line up. The
E.g "Qwen 3.5 122B is just 10B active, so it's no where close to the dense 27B model" That is the main sentiment around here and it puzzles me. If a 122B is just worth 10B, then why does model providers bother creating an MoE model when they could've just released a dense 10B model? Heck the 10B dense would run faster than the 122B MoE (no routing overhead), which negates the supposed ( only advantage of MoE is speed ) argument. It sure is not that simple. I mean yes it's only 10B active at a ti
Here is the upper limit of what can be done with $100 bucks worth of video cards. You can have 3 concurrent users with plenty of context, better speeds or close enough speeds than a bunch of cards that provide less VRAM and cost 4+ times. 0.00.008.388 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg) 0.00.008.391 I device_info: 0.00.089.439 I - CUDA0 : NVIDIA P102-100 (10144 MiB, 10013 MiB free) 0.00.197.645 I - CUDA1 : NVIDIA P102-100 (10144 MiB, 10013 MiB free) 0.00.197.656 I - CPU :
Soooo! Since Anthropic share their Jacobian-Lens a few days ago I went on and made a tool based on it which adds the possibilité to export a model which will have the same behavior after tweaking it's J-Space. This means manually alter the behavior and abliterate by using a human brain. I'm still working on it but couldn't wait to produce something first. SO After finally getting a working codebase I immediatly jumped and tried to make pretty pervy model PURELY in the name of science. Let me int
Built around a recent sd.cpp release, aims to expose most of what the backend can do (generate, edit, video paths, models, hardware options), Windows + Linux builds GitHub:
arXiv:2607.07766v1 Announce Type: new Abstract: Large language models (LLMs) have become significant providers of mental health support, yet they remain products of an attention economy whose operational and commercial targets favour sustained engagement over the friction that effective psychological support often requires. Developers' safety responses have been largely reactive, addressing the most visible and acute harms while subtler, longer-term patterns of risk (e.g., dependency, boundary e
arXiv:2607.07760v1 Announce Type: new Abstract: We outline an adversarial social epistemology (ASE) for densely interactive communicative landscapes in which public assertions are scaffolded by chains of testimony, inference, institutional certification, and tacit trust. In such landscapes, agents have incentives and affordances to distort, color, omit, fabricate, or strategically under-specify information for private, reputational, rhetorical, or material gains. We argue that these phenomena ar
"Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way," the company said in a blog post. "We've heard the feedback that this feature missed the mark, so it's no longer available."
Cant believe how much this DDR4 can handle
HyperTexting's new app aims to make the open web feel more like social media by turning websites, blogs, newsletters, and podcasts into a scrollable feed, while also making it easier to post to your own website.