Okay LLM + PyTorch people, trunc_normal_, what the fuck! Many LLM inits use it w/ default cutoffs. It's either not doing anything or it's quite broken due 2 issues. 1. The a/b cutoffs in PyTorch are not in std-devs, they are absolute. So w/ a std=0.02, and -2/2 (default arg) cutoffs that's 100σ!! That is a normal distribution, trun isn't doing anything. 2. There are numerical issues. Even in float32, the truncation produces a handful of -2 (lower cutoff) values, 100σ!! That's incomprehensibly improbable. I doubt a float32 or even float64 algo could even produce it, but clamping a bad float value does. Olmo (@allenai codebases) appear to be one of the few that uses trunc_normal_ and bothered to set the cutoffs properly. It'd be nice to see more train code opened up as a default. We so often only end up with a sanitized version of the inference/fine-tune friendly model these days and may lose details like original init. I've known about #1 for ages, I have an alternate trunc_normal_tf_ implementation in timm for that reason. But I saw those -2's last week when I was debugging something and was a little surprised.

Falcon 9’s fleet leader booster completes its 34th launch and landing https://t.co/xtWRgriDSO

by @aemofo https://t.co/7raV9LPFnW

by @aemofo https://t.co/7raV9LPFnW

OpenAI 官方发布了一个 Claude Code 插件 codex-plugin-cc，让开发者可以直接在 Claude Code 里调用 Codex 做代码审查、对抗性审查，甚至把任务整个移交给 Codex 执行。 这件事有意思的地方在于：这是 OpenAI 主动把自己的工具送进竞争对手 Anthropic 的地盘。Claude Code 有自己的插件生态，OpenAI 这次正式以官方身份入场，把 Codex 包装成 Claude Code 工作流里的一个"随叫随到的第二意见"。 插件提供三个核心命令：/codex:review 跑一遍标准的只读代码审查；/codex:adversarial-review 做对抗性审查，专门挑战现有实现的隐藏假设，适合迁移、鉴权变更、基础设施脚本这类高风险操作；/codex:rescue 则直接把任务交给 Codex 接管，用于线程卡住或需要换个智能体重新来过的场景。 三个命令都支持后台运行，配合 /codex:status 和 /codex:result 管理。还有个可选的 review gate 功能，能让 Claude Code 在 Codex 审查完成前不退出，不过 Srivastav 提醒这可能导致两个智能体循环调用，快速烧掉使用额度。 技术上，插件通过本地 Codex CLI 和 app server 中转，复用已有的认证、配置和 MCP 设置，不额外起运行时。使用前提是有 ChatGPT 订阅（包括免费版）或 OpenAI API key，加上 Node.js 18.18 以上。

In @lawfare, Mihir Khirsagar and I break down Judge Rakoff's recent ruling on AI and privilege in United States v. Heppner. While Rakoff reached the right outcome here, he went further than he needed to with massive implications for how AI can be used in legal services...🧵

PixVerse Power-Up Week Day 1: V6 is live. More control. Better performance. Film-ready output. 15s 1080P audiovisual, generated in seconds. RT+Follow+Reply=300Creds(72H ONLY)

the new @OpenAI banner is pretty fire https://t.co/5X1G1Bd5mt

the new @OpenAI banner is pretty fire https://t.co/5X1G1Bd5mt

🚀 Qwen3.5-Omni is here! Scaling up to a native omni-modal AGI. Meet the next generation of Qwen, designed for native text, image, audio, and video understanding, with major advances in both intelligence and real-time interaction. A standout feature: 'Audio-Visual Vibe Coding'. Describe your vision to the camera, and Qwen3.5-Omni-Plus instantly builds a functional website or game for you. Offline Highlights: 🎬 Script-Level Captioning: Generate detailed video scripts with timestamps, scene cuts & speaker mapping. 🏆 SOTA Performance: Outperform Gemini-3.1 Pro in audio and matches its audio-visual understanding. 🧠 Massive Capacity: Natively handle up to 10h of audio or 400s of 720p video, trained on 100M+ hours of data. 🌍 Global Reach: Recognize 113 languages (speech) & speaks 36. Real-time Features: 🎙️ Fine-Grained Voice Control: Adjust emotion, pace, and volume in real-time. 🔍 Built-in Web Search & complex function calling. 👤 Voice Cloning: Customize your AI's voice from a short sample, with engineering rollout coming soon. 💬 Human-like Conversation: Smart turn-taking that understands real intent and ignores noise. The Qwen3.5-Omni family includes Plus, Flash, and Light variants. Try it out: Blog: https://t.co/yuSAz3DuO8 Realtime Interaction: click the VoiceChat/VideoChat button (bottom-right): https://t.co/nnAW9ZfRet HF-Demo: https://t.co/rLsqejKgCG HF-VoiceOnline-Demo: https://t.co/LIGtmITeSw API-Offline: https://t.co/lNE7fH5YUt API-Realtime: https://t.co/9A3lopXGwV

+3 more

Somebody just mapped the ENTIRE Hermes Agent ecosystem 40+ skills, tools, integrations, and resources for @NousResearch's agent, all in one curated list highlights: ▫️ skills that upgrade themselves after every run ▫️ 734 security playbooks, ready to deploy ▫️ visual UI for memory, sessions, and skills ▫️ remote control your Android from your agent ▫️ Claude Code -> Hermes task handoffs ▫️ full OpenClaw migration in one command ▫️ AI council that debates before executing every entry has a real editorial description, not just a link the ecosystem most people don't know exists yet https://t.co/5dUWWRh99x [ BOOKMARK ]

The Trump Vance FTC was ordered to stop "burdening" AI companies. So Clarifai trained its models on 3 million stolen dating profile pics and gets to keep everything. The cops are working for the robbers. https://t.co/nnK3xZpgMx

Bluesky just launched AI that lets users program their own feed. Not an algorithm that decides what you see. A tool that lets YOU decide what you see. One platform says, "trust our algorithm." The other says, "build your own." This is the split happening across all of AI right now. Who controls the model? Who owns the data? Who decides what it does? The companies betting on user control are winning. @bluesky . @AnthropicAI . And that's the bet we're making at @Uare_ai. AI that works for you. Not the other way around.

130+ skills are now built into your Superagent. Some are ready to use, and some can be created based on what you need. Add a skill once, and your Superagent can use it as part of your workflows. Stack skills, connect tools, and build flows that run end-to-end. https://t.co/B6ntwyRkWm

This seems useful https://t.co/rjrX4jct8U

One of the things that is useful about the ChatGPT GPT-5.4 Pro (and also Thinking) harness is that it is quite good at understanding how to read scientific papers, not just relying on text, but also figuring out which figures are key and inspecting those visually. https://t.co/3jHFNWoieP

Exa is launching a Singapore office focused on web-scale infrastructure! The Google/Bing engineers who built the first generation of web infra from scratch haven't done it for decades. We're doing it again for the next generation of search. This will need to be at bigger scale -- exabytes of data, higher standard for quality, built from scratch for AI -- and it will be a global effort.

Security doesn't need to be intimidating. In just 5 minutes (or the time it takes to make your ☕️), you'll know the basics of securing your projects and keeping them safe with GitHub Advanced Security. The new episode of GitHub for Beginners is up. https://t.co/5HQxEGVejI

X after pushing Japanese posts https://t.co/ZOtdJz7jCQ

Solo dev reverse-engineered Google's billion-dollar algorithm in 7 days Google published the paper that crashed memory stocks worldwide. Then shipped zero code. Tom Turney read the math, opened his terminal, and built the whole thing with Claude - then made it faster than Google promised. Day 1-3: Core algorithms, 141 tests, Python prototype Day 3-5: C port into llama.cpp, Metal GPU kernels Day 5-7: Speed optimization from 739 to 2747 tok/s That's a 3.7x speedup through pure engineering: > fp32 → fp16 WHT > half4 vectorized butterfly ops > graph-side rotation > block-32 storage layout Then he added his own research on top: > Sparse V: skip 90% of value decompressions at long context > Asymmetric K/V: keep keys precise, compress values harder > Temporal decay: old tokens get lower precision automatically Result: 35B model running on a MacBook with 4.6x compressed cache. 613 GitHub stars in a week. Google still hasn't released their own code.

Thanks to @AI21Labs for tracking down a silent uint32 overflow in vLLM's Mamba-1 CUDA kernel and contributing the fix. Root cause: `uint32_t` stride × cache_index overflows silently at scale. Fix merged in #35275. The debugging story is worth a read. 🔗 https://t.co/S4XBnEn1uv

Here comes AutoClaw. We offer a new solution to run OpenClaw locally on your own machine. - Download and start immediately. No API key required. - Bring any model you like, or use GLM-5-Turbo, optimized for tool calling and multi-step tasks. - Fully local. Your data never leaves your machine. We're giving data control back to Claw users. Meet AutoClaw → https://t.co/mI3nne0nz0 Join the conversation → https://t.co/i9MRHOJZKj

I like Will's takes here. The first proposal is pretty close to home for me, and is something I'd really love to see happen: https://t.co/K2YNEdnZBa

@noritunamayo https://t.co/pIm22hVgZB

MIT’s inFORM is a shape shifting interface that turns digital data into physical forms, letting users interact with objects remotely. https://t.co/ZHFjHO367n

South Korea–based WIRobotics has introduced a humanoid robot called Allex. Allex features a lightweight arm with a 15-joint hand that can sense tiny forces and lift objects up to 30 kilograms. The company plans to develop it into a safe, flexible, general-purpose humanoid by 2030.

🚨Another update to our Generative AI US adoption time series results from our paper “The Labor Market Effects of Generative Artificial Intelligence”: we find LLM adoption at work in the US fell over the past quarter (while still up substantially from a couple years ago). https://t.co/Lo1v0V2zq0

I wholeheartedly endorse this piece. It annoys the hell out of me that there's an entire cottage industry of business "journalism" that consists of taking some outlandish claim made by a tech CEO -- often in a tweet or offhand comment -- and then crafting https://t.co/gjctqdvfUU

@emollick it’s certainly not “HUGE” by conventional interpretations; in traditional psychology (Jacob Cohen’s d) .15 is small; even in the context of education I am skeptical, and see quote below from Interpreting Effect Sizes of Education Interventions Matthew A. Kraft), which would still count it as medium, not large and certainly not all caps HUGE. But th thing about effect sizes of 0.15 in my experience is that there is usually enough of a confidence interval around it that you can’t be sure how real it is. That was Cohen’s point. (source for the 6-9 months?)

Microsoft built a voice AI so powerful they had to take it down. Deepfakes. Disinformation. Too dangerous. So they added watermarks, safety controls, and re-released it. For FREE. It's called VibeVoice. Here's what it does: → Clone any voice from 10 seconds of audio → Generate 90 minutes of multi-speaker conversation in one pass → Real-time TTS, first audio in ~200ms → Speech-to-text: 60 min of audio in a single pass, with speaker labels → 50+ languages, 4 speakers, natural turn-taking ElevenLabs: $99/month. Playht: $39/month. VibeVoice: Free. Local. MIT license. 28.5K stars. Backed by Microsoft Research. The fact they had to pull it once tells you everything about how good it is.