AI 资讯
来自 Twitter · Hacker News · GitHub 的 AI 行业动态
应用层比模型层更有结构性优势，Agent 开发者可以跑赢大厂
swyx 认为 Agent 实验室和开源开发者相比模型大厂有两大优势：可以在所有模型中取最优（argmax），且无需受安全审查约束自由探索能力边界。
@swyx
阅读 →
OpenAI：GPT-5.2 在粒子物理学中发现新公式，推翻数十年理论假设
OpenAI 与物理学家合作发表预印本论文，GPT-5.2 简化了胶子相互作用的复杂表达式并推导出通用公式，推翻了粒子物理中「单负振幅为零」的长期假设。另一内部模型独立推导出相同结论。
@OpenAI
阅读 →
Browserbase + Cerebras：极速浏览器 Agent 模板开源
Browserbase 联合 Cerebras 推出浏览器 Agent 模板，使用最新开源模型批量启动浏览器爬取文档并验证与代码库的一致性，数分钟内完成。
@browserbase
阅读 →
Latent Space：Jeff Dean 深度访谈，聊 Gemini Ultra 与 AI 工程关键数字
swyx 在 Latent Space 播客发布对 Google DeepMind 负责人 Jeff Dean 的深度采访，涉及 Gemini Deep Think、Gemini Ultra 去向以及「AI 工程师必知数字」等话题。
@swyx
阅读 →
Replit：用户反馈一键变功能，Agent 自动实现需求
Replit 推出反馈组件功能，用户在已发布应用中提建议后，Agent 可自动将其转化为已上线的功能，实现需求闭环。
@Replit
阅读 →
Forge：可扩展的 Agent 强化学习框架发布
新开源项目 Forge 发布，提供可扩展的 Agent 强化学习框架和算法，为构建 AI Agent 提供训练基础设施。
@_akhaliq
阅读 →
MiniMax-M2.5 模型上线 Hugging Face 开源社区
MiniMax-M2.5 模型权重已发布到 Hugging Face，同时提供 API 服务，开发者可直接下载使用。
@_akhaliq
阅读 →
LMSys：MiniMax-M2.5 模型发布，编程与 Agent 能力达 SOTA
MiniMax 发布 M2.5 模型，在编码、Agent 工具调用和办公场景中达到 SOTA 水平。该模型通过大规模真实环境 RL 训练，具备架构级编程能力和高效搜索推理，SGLang 已提供 Day-0 支持。
@lmsysorg
阅读 →
AMA with MiniMax — Ask Us Anything!
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1r3t775/ama_with_minimax_ask_us_anything/"> <img...
[D] Teaching AI to Reason With Just 13 Parameters
<!-- SC_OFF --><div class="md"><p><em>Made with</em> <a href="https://paperglide.net/"><em>Paperglide</em></a> <em>✨ — digest research papers faster</em></p> <p><strong>TL;DR:</strong>...
[D] How do your control video resolution and fps for a R(2+1)D model?
<!-- SC_OFF --><div class="md"><p>So I am using a R(2+1)D with kinetics 400 weights to train a classifier on two sets of videos. The problem is that one of the two classes has all videos of the same resolution and fps, forcing the model to learn those features instead of...
[D] Has anyone received their ICML papers to review yet?
<!-- SC_OFF --><div class="md"><p>I thought the reviewing period should have started yesterday, but it still says &quot;You have no assigned papers. Please check again after the paper assignment process is complete.&quot; </p> </div><!-- SC_ON...
[P] SoproTTS v1.5: A 135M zero-shot voice cloning TTS model trained for ~$100 on 1 GPU, running...
<!-- SC_OFF --><div class="md"><p>I released a new version of my side project: SoproTTS</p> <p>A 135M parameter TTS model trained for ~$100 on 1 GPU, running ~20× real-time on a base MacBook M3 CPU.</p> <p>v1.5 highlights (on CPU):</p>...
[R] Has anyone experimented with MHC on traditional autoencoders/convolutional architectures?
<!-- SC_OFF --><div class="md"><p>I&#39;m currently making a baseline autoencoder for this super freaking huge hyperspectral image dataset I have. It&#39;s a really big pain to work with and to get decent results, and I had to basically pull all stops including...
[D] Benchmarking Deep RL Stability Capable of Running on Edge Devices
<!-- SC_OFF --><div class="md"><p>This post details my exploration for a &quot;stable stack&quot; for streaming deep RL (ObGD, SparseInit, LayerNorm, and online normalization) using 433,000 observations of real, non-stationary SSH attack traffic.</p>...
[R] Higher effort settings reduce deep research accuracy for GPT-5 and Gemini Flash 3
<!-- SC_OFF --><div class="md"><p>We evaluated 22 model configurations across different effort/thinking levels on Deep Research Bench (169 web research tasks, human-verified answers). For two of the most capable models, higher effort settings scored worse. </p>...
[D] ICML: every paper in my review batch contains prompt-injection text embedded in the PDF
<!-- SC_OFF --><div class="md"><p>I’m reviewing for ICML (Policy A, where LLM use is not allowed) and noticed that in my assigned batch, if you copy/paste the full PDF text into a text editor, every single paper contains prompt-injection style instructions embedded...
Show HN: Skill that lets Claude Code/Codex spin up VMs and GPUs
I've been working on CloudRouter, a skill + CLI that gives coding agents like Claude Code and Codex the ability to start cloud VMs and GPUs.<p>When an agent writes code, it usually needs to start a dev server, run tests, open a browser to verify its work. Today that all happens on your local...
I'm not worried about AI job loss
I'm not worried about AI job loss
The "AI agent hit piece" situation clarifies how dumb we are acting
Previously:<p><i>An AI agent published a hit piece on me</i> - <a href="https://news.ycombinator.com/item?id=46990729">https://news.ycombinator.com/item?id=46990729</a> - Feb 2026 (916 comments)<p><i>AI agent opens a PR write a blogpost to shames the maintainer who...
Open source is not about you (2018)
Open source is not about you (2018)
OpenAI has deleted the word 'safely' from its mission
OpenAI has deleted the word 'safely' from its mission
CBP signs Clearview AI deal to use face recognition for 'tactical targeting'
CBP signs Clearview AI deal to use face recognition for 'tactical targeting'
Zed editor switching graphics lib from blade to wgpu
Zed editor switching graphics lib from blade to wgpu
MinIO repository is no longer maintained
MinIO repository is no longer maintained
Custom Kernels for All from Codex and Claude
Custom Kernels for All from Codex and Claude
Scaling social science research
GABRIEL is a new open-source toolkit from OpenAI that uses GPT to turn qualitative text and images into quantitative data, helping social scientists analyze research at scale.
Beyond rate limits: scaling access to Codex and Sora
How OpenAI built a real-time access system combining rate limits, usage tracking, and credits to power continuous access to Sora and Codex.
Introducing Lockdown Mode and Elevated Risk labels in ChatGPT
Introducing Lockdown Mode and Elevated Risk labels in ChatGPT to help organizations defend against prompt injection and AI-driven data exfiltration.
GPT-5.2 derives a new result in theoretical physics
A new preprint shows GPT-5.2 proposing a new formula for a gluon amplitude, later formally proved and verified by OpenAI and academic collaborators.