📡 2026-05-19 情报日报

亚利桑那州学生在毕业典礼上向谷歌前CEO埃里克·施密特喝倒彩

谷歌前CEO埃里克·施密特因AI言论在毕业典礼上被嘘

参议员Adam Schiff提议法案要求数据中心自行承担电力费用

必胜客AI系统造成「连锁」问题并导致1亿美元损失，加盟商在诉讼中指控

中国称「世界首个」海上风电水下数据中心已全面运营

亚利桑那大学学生在毕业典礼上向埃里克·施密特的AI煽动言论喝倒彩

被Kevin O'Leary点名的两位犹他州女性反击，发布嘲讽视频

PlayStation总裁称单机游戏不会继续登陆PC

FBI想购买全国车牌读取器访问权限

针对OpenAI诉讼的所有诉求在联邦审判中被驳回

Mozilla对英国监管机构：VPN是重要的隐私和安全工具

Meta内部，一名工程师抗议笔记本监控的帖子正在疯传

揭秘：Facebook页面用AI推广关于政治家的假新闻

荷兰当局称Meta打击网购诈骗做得远远不够

用现代C++在康威生命游戏中模拟无限

免费Cloudflare的乐趣

没人反驳：为什么工程师总是沉默直到太迟

对象存储事务性使用协议

为乐趣和利益混搭建模技术

[P] Sub-JEPA：LeCun团队的LeWorldModel一个简单修复，持续提升性能

[D] AI/ML伦理学家

[P] Witchcraft：基于SQLite的快速本地语义搜索

[D] 自5月9日以来TMLR没有新论文在审？

[N] MLRC 2026开放提交——NeurIPS 2026官方track

[P] 用CUDA内核重写模型推理：瓶颈不只是GEMM

[D] 有人收到ICML 2026 GlobalSouthML workshop的决定了吗？

[D] 候补名单的人会收到邮件吗？EEML 26

[D] 编码代理的未来是JEPA吗？

陪审团裁定 Elon Musk 诉OpenAI案败诉，称其诉讼时效已过

Linus Torvalds评论Linux维护者「无法管理」的AI漏洞报告

Cloudflare发布对50多个自有仓库运行Anthropic的Mythos Preview后发现的问题

Elon Musk：将向第九巡回法院上诉

今日讽刺：我们小创作者不能用AI，大公司却能用同样的AI封禁我们

Microsoft Copilot Cowork现已推出——AI从聊天走向实际工作执行

太空数据中心是真实的还是只是炒作？

想知道这有什么应用场景吗？

为AI编码代理打造的本地优先上下文引擎——符号图+语义搜索，无云

LLM对你最有用的不是写作或编码的事是什么？

Leonard Frankenstein OS

「只要加更多算力」这个AI推理论点越来越让人疲惫

Anthropic 9000亿美元融资——是好入场点吗？

能用AI把minecraft mods从1.21.11更新到26.1.2吗？

单模型AI图像检测在生产中失败。六模型集成实际上长什么样？

OpenSpec + Claude Code规范驱动原型设计在线免费课程

[ Removed by Reddit ]

工具：这是技术胜利，还是价格战胜利？

我用5种语言在6个AI系统上运行相同的研究提示词。结果不一样

为什么GitHub在PR审查方面不提供更好的指标？

学生津贴通常需要多久？如果学校没注册呢

关于使用量计费准备的预算——Github Enterprise Cloud

GitHub通知邮件的text/plain版本最近变得很糟糕

你的发布流程是什么？自动还是手动？

你有多广泛地使用install-* actions？

我构建了SprintHub：一个VS Code扩展，把GitHub Projects v2看板和高信号PR仪表盘直接带入编辑器

想找人评估我的仓库结构，这是我在github的第一个仓库，收到了很多负面反馈甚至没人看文件

roast我的PR总结格式：我试图把AI生成的PR压缩成60秒风险评估

别再手动管理仓库了！看这里如何在Google Antigravity让AI代理帮你做

网络好像挂了！

Python最好的现代DB层是什么？AI友好、简单、始终可用原生SQL转义？

Qwen迫不及待要发布3.7模型了

Qwen 3.7已在Qwen Chat上线

还是为你们高兴

如果免费LLM不再发布，本地LLM会怎样？

llama.cpp MTP支持已合并——Qwen3.6 27B在Strix Halo上2.44倍、在RTX 3090主机上2.17倍

PSA：如果你的Llama.cpp几天没更新且MTP表现不好，更新llamacpp

CPU上 Kokoro 82M vs Supertonic 3 TTS基准测试

MTP (Multi-Token Prediction)：AMD Strix Halo & Radeon 9700 AI Pro上2倍Token生成速度

Qwen 35b a3b 让我惊喜

21张GPU基准测试运行小型TTS模型 (显存峰值：5GB)

Lemonade v10.5.1：Strix Halo的MTP + ROCm 7.13快速入门

MLX引擎比较……oMLX是首选

Qwen3.6-35b-a3b (12Gb VRAM) 配置

llama.cpp nixos flake坏了吗？

对明天Google IO的期待？👀

三星会议记录显示存储器工人获得高达60.7万美元的607%奖金，而逻辑芯片员工只有50%

我构建了一个小CLI来检查仓库中的危险AI代理指令

我构建了一个小CLI来检查仓库中的危险AI代理指令

HF下载器实用工具 tampermonkey

内存专家认为由于中国大力投资，2027年下半年RAM价格将下跌

Built this over the past few weeks as part of a multilingual research project. Figured I'd share it here. Check it out! \~9.8M web documents across 11 languages — hi, bn, ta, te, mr, gu, kn, ml, pa, ur, en. \~8.4B tokens. CC0 license. 🤗 [https://huggingface.co/datasets/AM0908/indic-hplt-v1](https://huggingface.co/datasets/AM0908/indic-hplt-v1)

AI systems are environmentally and socially embedded. They cannot thrive in a degraded human ecosystem. Therefore, the measurement and protection of human health (data integrity, environmental stability, and economic agency) is the primary engineering requirement for the next generation of AI. Slightly rephrased, AI systems are only as good as the human data, institutions, and economic conditions they’re trained on and deployed into. Curious what others think — is this already being treated a

Have you completed the non code Postgraduate Program in AI Agents and Generative AI for Business Applications? I’m considering enrollment later in the year and would like to speak with someone outside of the school who has completed a similar program or is currently enrolled in it or similar and has a background in non tech professional roles.

# Weekly Wednesday Thread: Advanced Questions 🐍 Dive deep into Python with our Advanced Questions thread! This space is reserved for questions about more advanced Python topics, frameworks, and best practices. ## How it Works: 1. **Ask Away**: Post your advanced Python questions here. 2. **Expert Insights**: Get answers from experienced developers. 3. **Resource Pool**: Share or discover tutorials, articles, and tips. ## Guidelines: * This thread is for **advanced questions only**. Beginner

So far, I’ve tried Codex CLI, Claude Code, Gemini CLI, OpenCode, and recently, Pi with local models. Pi is the leanest of them all, with just four tools: read, write, edit, and bash. Its system prompt is only under 2K tokens, and it's perfect for local models. I've been trying out Qwen 27B-MXFP8 with it, and it's much better than I expected! It doesn't have fancy bells and whistles like multi agents, but the only thing I’m missing is searching the web for documentation. I’m sure you can get i

Hey everyone, I’m building a backend that analyzes long YouTube videos using an LLM. Currently, my flow is a slow waterfall: `Download full audio -> Whisper -> LLM -> Return results`. For a 30-minute video, the user waits forever. I want to pipeline this for real-time SSE streaming: `[Chunk Audio on the fly] -> [Whisper] -> [LLM] -> [Stream to UI]` My questions for the data/backend engineers: 1. **Chunking & VAD:** What's the best way to chunk YouTube audio streams (e.g

I posted earlier about RTX 5060 Ti local LLM testing, and I have cleaned the repo up quite a bit since then. The project is now a more structured benchmark/recipe repo rather than scattered notes. It has a static results explorer, schema-validated benchmark JSON, clearer llama.cpp/vLLM notes, single-card and dual-card RTX 5060 Ti recipes, a model-agnostic download helper, and better labels for generation speed, prompt eval speed, MTP/no-MTP, and thinking mode. Repo: https://github.com/5p00kyy/

We kept running into the same problem every time we rented a GPU to run Ollama + OpenWebUI or ComfyUI, we'd spend the first 45 minutes reinstalling everything. Custom nodes, models, configs, all of it. Docker images went stale fast, different providers had different base images, and nothing was truly portable. We got sick of it and built swm. Here's what it does for ComfyUI users specifically: swm gpus -g a100 --max-price 2.00 --sort price shows you the cheapest available GPU across RunPod, Va

I've had 3 successive cases of theft of a Claude API key over the past few weeks. I'm trying to localize the source of the leak, and one possibility is my private repository on GitHub - which is an intermediate link in the CI/CD chain prior to publishing a website on Azure. Curiously, I just got a popup on the GitHub repository saying something to the effect of "We just noticed you're trusting credentials from [**alive.github.com**](http://alive.github.com) and maybe you don't want this" OK.

I don't have a formal degree in IT, but I've been diving deep into fields like RPA, AI agents, LLM fine-tuning, and Machine Learning. According to reports (like UiPath's), Python is pretty much the backbone for all of this. If you were looking to land a Python Developer role starting from scratch today, would you prioritize certifications like the Python Institute’s PCAP, or would you take a different route? I’d love to hear your personal stories and what worked for you!

At work I get unfettered access to gpt 5.4 and sonnet, so I'm quite used to spawning sub-agents to go crazy on a repo and split up tasks. At home I am VRAM poor and like to run the models locally for my own enjoyment. Almost every single sub-agent extension/implementation does not account for any of the restrictions imposed by having 10gb of VRAM and a single slot for a KV cache (thats already quantized). I already work as a developer, so I qwen3.6-35b-a3b tagged teamed a partially vibe-code

[View Poll](https://www.reddit.com/poll/1th83jl)

I need a small model for processing conversation transcripts from larger models, so need usable context window out to at least 200k tokens. I know some models claim to support this, but I don’t know which are actually good at this in practice. Also desirable: low hallucination rate, not super verbose.

Hey everyone, I’m an undergrad from India and I just found out I had two papers accepted at the ICML 2026 GlobalSouthML workshop! I am super excited since this is my first time getting accepted into a major conference venue, but I’m also kind of panicking right now because I absolutely cannot afford a trip to Seoul. Since I've never done this before, I’m hoping some experienced folks can help answer a few questions about how the post-acceptance process works: 1. I saw that the main conference

Hi everyone, I'm starting a research project on financial time-series forecasting using LSTM and Transformer models for predicting S&P 500 market direction. Right now, I'm struggling with obtaining reliable long-term historical data. I tried Yahoo Finance, but downloads are inconsistent/failing for me, and most Kaggle datasets I found only contain around 5–10 years of data. I specifically need: * Around 30 years of historical S&P 500 data * Preferably daily OHLCV data * Reliable and

I’ve been exploring how AI tools and AI agents can actually reduce manual SEO work beyond just basic content generation. Curious to know from people actively working in SEO: * Which SEO tasks are you automating right now? * What workflows are giving you the biggest time savings? * Are you using simple AI tools, custom GPTs, Claude workflows, Zapier/Make automations, or fully autonomous agents? * Which tasks still need heavy human involvement? Some areas I’m personally thinking about: * Keywo

Hey all — I’ve been trying to get a better sense of what people are actually running locally these days. Curious about your setup: GPU (or CPU if you’re brave ) RAM / VRAM Models you use the most Main use case (coding, chat, agents, etc.) Also — what’s the biggest bottleneck you’re hitting right now? I hope to gather more use cases to gain a fuller understanding of GPU performance. Thank you everyone for sharing.

Well I ordered a 3090 today. I plan on pairing it with a 3060 I have for 32gb combined VRAM. Up until now I’ve just been using a 6GB card on my laptop. I’ve been using Gwen 3.5 4B so far. Where should I start? Gwen 3.6 27B? I’m interested in coding applications, mainly to teach myself more about coding so I can understand it a little better. I’ll be using Ubuntu and Llama.cpp, neither of which I’ve set up before so that will be a great learning experience. This is mostly a “I’ve saved up and a

`atool` maintainer account got hacked, and attacker pushed 631 malicious versions across 314 packages in 22 minutes. another day and another attack. it steals everything like AWS keys, GitHub tokens, npm creds, SSH keys, database strings, docker configs, kubernetes tokens. If you have docker socket exposed, it escapes the container with privileged access.

for years, I have been writing down and tricks for people wanting to get better at technical speaking writing. I have finally put them together into a single document. Hope someone finds this useful.

The biggest issue with the peer review system is reciprocal reviewing, which incentivizes reviewers to unfairly reject good papers to increase their own papers' chances of acceptance. My proposed solution is that the conference should divide the authors/papers into 2 halves (A and B). If you are an author in half A, then you will only be a reviewer in half B. All papers by the same author, their coauthors, and coauthors of coauthors should be in the same half. Each AC/SAC can only serve in o

I am a new student in a Private university in India My college will start in 3-4 months and I dont have any student email rn How do i claim the student developer pack from github?

Just stumbled across this insanely intresting anime streaming repo on GitHub and had to share. Not gonna pretend it’s fully stable or production-ready yet, there are definitely rough edges and bugs here and there — but the UI and overall vibe are genuinely impressive for an open-source project. The design is super clean, animations look great, and the codebase is interesting to explore if you're into: * anime apps * streaming UIs * React/Next.js projects * frontend inspiration * open-source s

What's good everybody, I probably have the fastest possible setup on these AMD Radeon RDNA2 GPUs for one reason only. A custom binary that bypasses some assert statement causing a crash in today’s stock releases. This binary bypasses that assert and enables flash attention. Works for rocm lamma cpp build with qwen3.6 35B. tldr; vulkan tok/s 30. stock rocm tok/s: Doesnt run. This build: 70-80 tok/s try it yourself. https://github.com/Minerest/llama.cpp\_RDNA2\_FlashAttnEnabled/releases/tag/m

Decided to build the open source version of the ChatGPT finance features that were released this past week. First release would love any feedback!

No more memory leak (native code)

Hello everyone. I am keeping my identity anonymous today to protect my professional career. I am a junior researcher in Computer Vision, and I am sharing this story because I have hit a devastating deadlock with IEEE T-PAMI and the IEEE Ethics Office. # Our Situation: https://preview.redd.it/v0w62gzmn02h1.png?width=2000&format=png&auto=webp&s=a2d75a1e3a388debdf5b163cb9593c1f7f1c49d5 In the decision letter, we actually received three highly positive reviews (Two EXCELLENT, One

Claude Design can make great animations, but getting to a final video is a bit hard. The audio is missing. Even if you use a TTS model, it does not align. Here is the process I used to get the video above 1. Get Claude to write a good script 2. Feed the script to a Text to Speech (TTS) model to get the audio 3. Feed the audio to a Speech to Text (STT) model to get key timestampes 4. Use the script and the STT output to Claude Design to get a video that's aligned with your audio 5. Use Claude V

These days, I can't open X anymore without seeing some supply chain attacks on PyPI or NPM. Things are really getting out of hand. One very simple yet effective approach to mitigate them is to use a dependency cooldown. That means that you don't install anything that's too new - e.g., every dependency needs to be at least a week old. Why does this work? Because the community usually intercepts them in hours to days. Both uv and poetry support the definition of the cooldown period inside their c

I am waiting for 122B and new 27B

Hi. Besides the fact that an xByA MoE models runs as fast as a yA models but produces better results, what are other benefits of pursuing an MoE architecture and not a dense one with e.g. x/2 (or x/3) parameters? Given that we need enough RAM for xB parameter anyway, aren't MoEs at a disadvantage when RAM is scarce, like the current situation? And thinking of limit cases, is there a limit on x/y, so that it doesn't make sense e.g. to train a 100B1A MoE model? Thanks.

I never see this type of model talked about. Are there many open models in the category? I do a lot of audio cleanup and end up using auphonic but would like to be using a local model. Edit: e.g like voice recovery, reverb removal, auto-EQ type stuff

Hi Everyone, I’m trying to understand whether GitHub Free version supports SAML federation and SSO integration with Microsoft Entra ID (Azure AD). My requirement is: * Federate GitHub with Entra ID * Enable SAML-based SSO * Allow users to authenticate via Entra ID I know GitHub Enterprise Cloud supports this, but I want to know if the same setup is possible with GitHub Free or GitHub Team plans without purchasing Enterprise licenses. Has anyone tested this recently? If not fully supported

Hey folks, I built an open-source tool for people who use MobSF Community Edition. MobSF is great, but the exported reports can be noisy. This tool lets you drop a MobSF JSON, PDF, or HTML report into the browser, runs local WebLLM/WebGPU triage, and labels findings as likely false positive, needs review, or likely real. Repo: https://github.com/moonpiesheldon1337/mobsf-fail-app Live demo: https://moonpiesheldon1337.github.io/mobsf-fail-app/ Why I made it: - MobSF reports often have 100-300

Suggestion: We should call python 3.14 pithon. For those who don't understand. Pi (the math thing) is 3.14

coding agents are everywhere right now but i'm more interested in models that actually take actions autonomously. we built a small vlm for desktop gui automation. i mostly use it for moving data between apps that don't have apis, saves me a lot of copy pasting. still kinda janky on complex UIs though. would be cool to see more people sharing non-coding use cases for local models

Hello, Does anyone with apple silicon had success with it? I tried both the froggeric and the unsloth 27B models I have an m2 max 96GB, and I can't get past 9/10 t/s, it is actually worse than without MTP where I have around 12 t/s... I tried 2,3 and 6 spec-draft-n-max ... I have a pretty high acceptance rate too, > 70%, so where is the problem ? Here's my parameters `gpu-layers = all` `temp = 1.0` `top-p = 0.95` `top-k = 20` `min-p = 0.0` `presence-penalty = 1.5` `flash-attn =