⚡️ Tencent drops Hy4 Preview: 770B parameters, open-sourced
Hunyuan's latest MoE model activates only 49B params but claims wins over both GLM 5.3 and Kimi K3 in head-to-head evals (2.99 avg vs. 2.92 and 2.94).
Margins are razor-thin. But in China's foundation model race, every decimal counts.
Hunyuan's latest MoE model activates only 49B params but claims wins over both GLM 5.3 and Kimi K3 in head-to-head evals (2.99 avg vs. 2.92 and 2.94).
Margins are razor-thin. But in China's foundation model race, every decimal counts.
❤1
🎙️ Hold Fn, speak, done. No cloud, no account, no history.
FnScribe is a free, open-source macOS dictation app that keeps everything on-device using a quantized Whisper model. Works on both Silicon and Intel.
Alpha, English-only, GPLv3. GitHub
FnScribe is a free, open-source macOS dictation app that keeps everything on-device using a quantized Whisper model. Works on both Silicon and Intel.
Alpha, English-only, GPLv3. GitHub
❤1
🤖 Open-source maintainers are begging devs to stop using their repos as AI CV farms
Maintainer Neil Alexander writes that his projects are getting flooded with AI-generated grammar fixes and trivial tweaks, not to improve anything but to rack up "contributor to <famous-repo>" credits.
The changes are technically correct. That's the whole problem. Reviewing slop still costs real time.
Open source runs on trust. This is quietly draining it.
Maintainer Neil Alexander writes that his projects are getting flooded with AI-generated grammar fixes and trivial tweaks, not to improve anything but to rack up "contributor to <famous-repo>" credits.
The changes are technically correct. That's the whole problem. Reviewing slop still costs real time.
Open source runs on trust. This is quietly draining it.
neilalexander.dev
Please stop flooding our projects with AI slop to furnish your CV
Successful contributions to open source projects are a kind of currency. GitHub in particular encourages this in a number of ways: by showing avatars of contributors on repository pages, by showing your contributions to your followers via the activity feed…
❤1
⚡️ Your AGENTS.md is probably hurting you
ETH Zurich tested context files across hundreds of real repos. LLM-generated AGENTS.md files actually reduce success rates while adding 20%+ to token costs. Human-written ones barely move the needle.
The only stuff that works is short, specific instructions the model can't infer itself ("use uv, not pip"). Source
ETH Zurich tested context files across hundreds of real repos. LLM-generated AGENTS.md files actually reduce success rates while adding 20%+ to token costs. Human-written ones barely move the needle.
The only stuff that works is short, specific instructions the model can't infer itself ("use uv, not pip"). Source
Pivot to AI
Your AGENTS.md file doesn’t actually do anything
AI coding bot vendors tell you to use a context file with instructions for the chatbot on how to edit your project. Claude Code wants a CLAUDE.md, or there’s AGENTS.md in general. [Anthropic] But d…
❤3
🚨🔥 Judge calls Pentagon's Anthropic blacklist "illegal and baseless"
A federal judge just blocked the DoD's supply chain risk designation against Anthropic. The government's entire case rested on a 4-page memo and a claim about backdoor access that, turns out, doesn't exist.
"Empty invocation of national security is not a blank check to punish critics." Court said it plainly.
A federal judge just blocked the DoD's supply chain risk designation against Anthropic. The government's entire case rested on a 4-page memo and a claim about backdoor access that, turns out, doesn't exist.
"Empty invocation of national security is not a blank check to punish critics." Court said it plainly.
❤1
⚡️ OpenAI's Python SDK drops httpx for httpx2
The OpenAI Python SDK now uses httpx2 for its HTTP clients. It installs automatically, and the old httpx package is no longer a direct dependency.
Anthropic made the same call shortly after. Both labs are quietly de-risking from upstream httpx, which is headed toward a breaking 1.0 that promises to snap a lot of existing code.
Source
The OpenAI Python SDK now uses httpx2 for its HTTP clients. It installs automatically, and the old httpx package is no longer a direct dependency.
Anthropic made the same call shortly after. Both labs are quietly de-risking from upstream httpx, which is headed toward a breaking 1.0 that promises to snap a lot of existing code.
Source
GitHub
openai-python/httpx2.md at main · openai/openai-python
The official Python library for the OpenAI API. Contribute to openai/openai-python development by creating an account on GitHub.
❤1
⚡️ GLM-5.3 is now open-weight and it's eating the coding leaderboard
Zhipu AI just dropped open weights for GLM-5.3. Same base as 5.2, but post-training alone pushed its Terminal-Bench score from 4.6 to 28.3. Six times. Open-source rank: first.
DeepInfra, Ollama Cloud, multiple OpenRouter endpoints. All live within hours. Early users are whispering "Opus 4.8 vibes." It's slow right now, but that's just the servers panicking.
Zhipu AI just dropped open weights for GLM-5.3. Same base as 5.2, but post-training alone pushed its Terminal-Bench score from 4.6 to 28.3. Six times. Open-source rank: first.
DeepInfra, Ollama Cloud, multiple OpenRouter endpoints. All live within hours. Early users are whispering "Opus 4.8 vibes." It's slow right now, but that's just the servers panicking.
X (formerly Twitter)
Z.ai (@Zai_org) on X
GLM-5.3 is now open-weight.
Our most capable model for agentic coding and cyber defense is now available to download, run, and customize.
Weights: https://t.co/v1IbWMXxg4
Tech blog: https://t.co…
Our most capable model for agentic coding and cyber defense is now available to download, run, and customize.
Weights: https://t.co/v1IbWMXxg4
Tech blog: https://t.co…
⚡️ EPA carves out AI data centers from acid rain rules
The Trump EPA just exempted "islanded" power plants (ones that only serve data centers, off the public grid) from the Clean Air Act's Acid Rain Program. No utility status, no pollution limits.
Timing couldn't be worse. Public opinion on data centers is already turning fast, and this hands critics a clean target.
Source
The Trump EPA just exempted "islanded" power plants (ones that only serve data centers, off the public grid) from the Clean Air Act's Acid Rain Program. No utility status, no pollution limits.
Timing couldn't be worse. Public opinion on data centers is already turning fast, and this hands critics a clean target.
Source
EPA
EPA Issues Permitting Guidance to Further President Trump’s Agenda Promoting Data Centers and Safeguarding Communities
This month, EPA issued guidance clarifying that the Clean Air Act Acid Rain Program does not apply to power generation facilities that are not connected to a public electricity grid, commonly referred to as “islanded” power generation facilities.
🚨🔥 Microsoft's AI bot just killed a legit open-source app on Google Play
Tracer.AI filed a DMCA notice on Microsoft's behalf claiming Luanti, a fully open-source voxel platform, infringes Minecraft's copyright. No proprietary code. No stolen assets. Just "blocky game = illegal" logic from an automated scanner with zero human review.
This is literally the second time. Same company, same bogus claim, back in 2023 too.
Filers pay no price for being wrong. Open-source projects pay with their distribution. Source
Tracer.AI filed a DMCA notice on Microsoft's behalf claiming Luanti, a fully open-source voxel platform, infringes Minecraft's copyright. No proprietary code. No stolen assets. Just "blocky game = illegal" logic from an automated scanner with zero human review.
This is literally the second time. Same company, same bogus claim, back in 2023 too.
Filers pay no price for being wrong. Open-source projects pay with their distribution. Source
Luanti Blog
Luanti removed from Google Play due to baseless AI copyright notice
Luanti has been removed from Google Play due to a DMCA notice from Tracer.AI. We have filed a counter-notice, but this isn't the first time.
❤1🤬1
🚨 NSA: "We want access to all the models."
Deputy Director Tim Kosiba said it out loud at an intel conference. All of them. Every frontier model. The NSA is now central to the White House's voluntary model-testing program and is actively talking to major AI labs.
"Voluntary" is doing a lot of heavy lifting in that sentence.
Deputy Director Tim Kosiba said it out loud at an intel conference. All of them. Every frontier model. The NSA is now central to the White House's voluntary model-testing program and is actively talking to major AI labs.
"Voluntary" is doing a lot of heavy lifting in that sentence.
Nextgov
NSA wants access to ‘all’ AI models, top official says
The spy agency is taking a key role in the government’s new voluntary model-testing program, though its deputy director did not say which companies are participating.
⚡️ OpenAI just cut Cursor off from future models
SpaceX closed its $60B Cursor acquisition, and OpenAI moved fast. They're holding the contract to its latest cancellation date but won't give Cursor access to any new models from here on out.
Cursor still runs on current GPT versions for now. But the clock's ticking, and the Musk-Altman cold war just got a product consequence.
Source
SpaceX closed its $60B Cursor acquisition, and OpenAI moved fast. They're holding the contract to its latest cancellation date but won't give Cursor access to any new models from here on out.
Cursor still runs on current GPT versions for now. But the clock's ticking, and the Musk-Altman cold war just got a product consequence.
Source
OpenAI
Our decision on Cursor following its acquisition by SpaceX
Our decision to wind down our contract providing OpenAI models to Cursor following its acquisition by SpaceX.
🤖 AI agents just did original math. New theorems, verified proofs.
No human in the loop. A multi-agent system called the Station turned AI models loose on hard open problems and got novel results: new Kakeya set families, kissing number records, a better Erdős bound.
Not just answers. Full proofs with explanations mathematicians can actually build on.
"Can AI do original work?" Just got harder to dodge.
No human in the loop. A multi-agent system called the Station turned AI models loose on hard open problems and got novel results: new Kakeya set families, kissing number records, a better Erdős bound.
Not just answers. Full proofs with explanations mathematicians can actually build on.
"Can AI do original work?" Just got harder to dodge.
arXiv.org
Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment
We study autonomous mathematical discovery in the Station, an open-world multi-agent environment in which AI agents from different model families pursue a shared research goal without a central...
🤖 Meta's putting robots to work in its data centers. Workers aren't thrilled.
Pilots with Watney Robotics, Kinova, and ABB. Tasks: cable swaps, rack moves, server restarts. Still needs humans watching over it.
Staff are already raising alarms internally. Meta's counter-spin: "we need more workers, not fewer." Classic. Source
Pilots with Watney Robotics, Kinova, and ABB. Tasks: cable swaps, rack moves, server restarts. Still needs humans watching over it.
Staff are already raising alarms internally. Meta's counter-spin: "we need more workers, not fewer." Classic. Source
WIRED
Inside Meta’s Push to Put Robots to Work in Data Centers
The company is testing robots that can swap cables, reset servers, and take on other tasks performed by technicians, fueling concerns among some workers that their jobs could be at risk.
❤1
⚡️ Samsung's RAM does your LLM math now
Samsung showed LPDDR5X-PIM at Hot Chips 2026. It runs MAC operations directly inside the memory chip, hitting 614 GB/s internal bandwidth vs. 76.8 GB/s for regular DRAM.
The idea is to stop starving your AI accelerator and just compute where the weights already live.
Killer app's obvious. It's LLMs. Whether Samsung can actually ship this into something you'd buy is the open question.
Samsung showed LPDDR5X-PIM at Hot Chips 2026. It runs MAC operations directly inside the memory chip, hitting 614 GB/s internal bandwidth vs. 76.8 GB/s for regular DRAM.
The idea is to stop starving your AI accelerator and just compute where the weights already live.
Killer app's obvious. It's LLMs. Whether Samsung can actually ship this into something you'd buy is the open question.
Chipsandcheese
Hot Chips 2026: Samsung’s Processing-in-Memory (PIM)
In-memory compute with LPDDR5X
❤1
🧠 Fixed LLM memory, got a static analyzer by accident
Instead of giving agents better memory, the author built Lemmalog, a Datalog engine that tracks an agent's knowledge as structured analysis state, with provenance, retractions, and incremental evaluation.
The LLM parses natural language into facts. Everything after that is deterministic. Symbolic AI doing the heavy lifting.
Wild detour. Source
Instead of giving agents better memory, the author built Lemmalog, a Datalog engine that tracks an agent's knowledge as structured analysis state, with provenance, retractions, and incremental evaluation.
The LLM parses natural language into facts. Everything after that is deterministic. Symbolic AI doing the heavy lifting.
Wild detour. Source
pwning.systems
I accidentally turned LLM memory into program analysis :: pwning.systems
Why I stopped trying to give LLM agents a better memory and instead built Lemmalog, a Datalog engine that maintains an agent's knowledge as analysis state, with provenance, retractions and incremental evaluation, plus what happened when I benchmarked it on…
❤2
🚨🔥 Ask Claude Code to summarize a website. Get remote code execution.
Researcher Johann Rehberger showed that Opus 5 in Auto Mode can be hijacked via a malicious webpage alone. Success rate: up to 80%.
Anthropic's own take is that Auto Mode is "a best-effort classifier, not a security guarantee." It's the default since mid-August.
Researcher Johann Rehberger showed that Opus 5 in Auto Mode can be hijacked via a malicious webpage alone. Success rate: up to 80%.
Anthropic's own take is that Auto Mode is "a best-effort classifier, not a security guarantee." It's the default since mid-August.
theregister
Researcher shows how Claude Code can be tricked simply by asking it to summarize a website
More prompt-injection hijinks from wunderwuzzi
❤1
🤖 Debian officially allows AI-assisted contributions
Debian just voted to allow "responsible use" of LLMs. No ban, no endorsement. Use whatever tools you want.
AI doesn't lower the bar. Same quality, same legal compliance, same accountability. You ship it, you own it.
Debian just voted to allow "responsible use" of LLMs. No ban, no endorsement. Use whatever tools you want.
AI doesn't lower the bar. Same quality, same legal compliance, same accountability. You ship it, you own it.
LWN.net
Debian votes to allow 'responsible use of generative AI'
The results of the Debian general-resolution vote on the use of large language models have been [...]
❤1
🤖 Free, local AI stem splitter that actually ships
StemDeck splits songs into 6 stems (vocals, drums, bass, guitar, piano) fully on-device. No account, no upload, no subscription.
CUDA, Apple Silicon, CPU. Works on Windows, macOS, Linux. Built-in mixer, BPM/key analysis, YouTube import. Docker-ready.
Honestly impressive for a solo dev side project.
StemDeck splits songs into 6 stems (vocals, drums, bass, guitar, piano) fully on-device. No account, no upload, no subscription.
CUDA, Apple Silicon, CPU. Works on Windows, macOS, Linux. Built-in mixer, BPM/key analysis, YouTube import. Docker-ready.
Honestly impressive for a solo dev side project.
GitHub
GitHub - stemdeckapp/stemdeck: Stemdeck is an modern stem extraction platform for musicians,producers and hobbyists, designed to…
Stemdeck is an modern stem extraction platform for musicians,producers and hobbyists, designed to isolate vocals, drums, bass, piano and guitar for practice, transcription, remixing, and creative ...
❤3
⚡️ Warp's Claude agents learn from your corrections
Most dev tools deploy an agent and call it done. Warp built something different: its PR review agent watches how your team corrects it, then rewrites its own skills to do better next time.
800K monthly devs already use Warp. Feedback compounds across the org. Honestly pretty close to the self-improving loop everyone's been theorizing about, just quietly shipped.
Most dev tools deploy an agent and call it done. Warp built something different: its PR review agent watches how your team corrects it, then rewrites its own skills to do better next time.
800K monthly devs already use Warp. Feedback compounds across the org. Honestly pretty close to the self-improving loop everyone's been theorizing about, just quietly shipped.
Claude
How Warp builds self-improving agents on Claude
How Warp devised a simple development pattern that anyone can use to create self-improving agents.
🤖 Sony Music and Warner Chappell just sued Anthropic over song lyrics
The claim: Anthropic ran a "brazen campaign" of torrenting and scraping copyrighted lyrics to train Claude. Thousands of works, named co-founders Dario Amodei and Benjamin Mann personally.
Odd detail: Anthropic has no audio features. This is purely a text-training-data play, and it's the music publishers' second swing at an AI lab this year.
Source
The claim: Anthropic ran a "brazen campaign" of torrenting and scraping copyrighted lyrics to train Claude. Thousands of works, named co-founders Dario Amodei and Benjamin Mann personally.
Odd detail: Anthropic has no audio features. This is purely a text-training-data play, and it's the music publishers' second swing at an AI lab this year.
Source
The Verge
Sony Music Publishing and Warner Chappell are suing Anthropic
“One of the largest and most blatant ongoing thefts of intellectual property in history.”
❤1