The creator of Linux just publicly called out the AI hype. Word for word.
Linus Torvalds took the stage at Open Source Summit 2026 and said this:
"When I see people saying 99% of our code is written by AI, I literally get angry. Because those same people — I can pretty much guarantee — 100% of their code is written by compilers. But they never say that."
He is not anti AI. The Linux kernel saw a 20% jump in submissions this release because of AI tools. He uses it. He gets it.
His point is something most people are too afraid to say.
AI is a productivity tool exactly like compilers were. Compilers boosted programming by 1000x. AI adds another 10x on top. Enormous. But nobody says "the compiler wrote my code." So why are we saying AI wrote it?
He also flagged something nobody is talking about.
AI is flooding small open source projects with drive-by bug reports. Someone runs a prompt, files a report and disappears when asked for a patch. Maintainers with one or two people are drowning trying to keep up.
"Sometimes AI reports a bug and when you ask for more information the person has done that drive-by and does not even answer your question. That is the real burnout issue."
And his final warning was the sharpest of all.
"People who do not understand the complexity of systems will prompt systems and write processes that will fail."
The AI hype crowd is very loud right now.
Linus has been building real systems for 35 years. When he talks, engineers listen.
Full interview here:
https://t.co/LmXJtvKc4O
If you see someone doesn’t have friends and your first thought is they probably deserve it vs wow let me try and be their friend, I wouldn’t want to be your friend.
Gemma 4 Quantization-Aware Training (QAT) weights are now available on Ollama!
They reduce memory requirements while maintaining model quality.
E2B:
ollama run gemma4:e2b-it-qat
E4B:
ollama run gemma4:e4b-it-qat
12B:
ollama run gemma4:12b-it-qat
26B:
ollama run gemma4:26b-a4b-it-qat
31B:
ollama run gemma4:31b-it-qat
Try them with ollama launch integrations to use with your favorite tools 👇👇👇
So here's how I capture audio from any app on Mac with Perspective Transcribe 2.0. In Settings, Quick Capture, pick the app and set a shortcut. I use it to grab beta feedback on Discord, transcribe it, and paste it into AI. My favorite feature. TestFlight soon.
Privacy AI 2.0.25 is live!
FEATURES
- Video Understanding From Any Link
Paste a link and the assistant pulls page text, title, and captions. Missing captions trigger on-device transcription. Social links merge post text with attached-video captions, and links with long tracking parameters no longer truncate. Works with most major video and social platforms.
- Real-Time Flight Search
New search_flight tool returns status, delays, gates, and baggage with fallback across FlightAware, AirLabs, AviationStack, AeroDataBox.
IMPROVEMENTS
- Faster first reply in a new chat (skips the relevance scorer).
- Keyboard opens automatically in empty chats.
- Input bar no longer collapses while a draft is in the field.
- Quick Ask, LLM Council, Siri, sub-agents no longer disturb an open OpenAI Responses chat.
- llama.cpp b9279: HunyuanOCR/HunyuanVL unified, DeepSeek-OCR resize fix, faster Metal, NvFP4 LM head, sliding-window crash fix.
- MLX-Swift tag-20260522: Gemma 4 ~24 percent faster, three Gemma 4 crashes fixed, Idefics3 prefill in one pass, Qwen3.5/Qwen3VL/GlmOcr M-RoPE stable.
- GGUF tool calling: Gemma 4, Qwen, LFM2 reliably invoke tools via OpenAI function-calling schema.
- Faster first reply from local GGUF: Metal kernels pre-compile at load.
- Novel Writer on small local models: outline parser recovers from missing tags, character names stay stable, stray markers no longer leak in.
- Banner on small local models showing context size and output cap.
- Reasoning models (GLM-5.1, DeepSeek, OpenAI o-series) now get budget for hidden reasoning plus full structured output.
- Chapter correction pass extends chapters under 70 percent or condenses over 130 percent.
- Token Usage charts now count thinking tokens for reasoning models.
- Duplicate bottom chat action bar removed; all actions in the top-right menu.
BUG FIXES
- Web search falls through Tavily, Brave, LangSearch when first returns nothing.
- macOS screens stuck in light-mode colors now follow Dark Mode immediately.
- Clear Chat on OpenAI Responses drops the server reference and wipes the per-chat workspace.
- Fixed Gemma4-E2B/E4B crash on 32 KB Metal threadgroup devices (incl. iPhone 16 Pro Max).
- Fixed GGUF metadata read crash on external drives or iCloud I/O errors.
- Fixed MLX long-prompt crash on iOS; prompts fed in 512-token chunks on iOS/Mac Catalyst.
- Fixed MLX vision models ignoring an image after a text-only reply.
- Fixed duplicate "Recovered From..." entries on cold launch.
- Custom assistant generation uses a fresh random seed (no more identical results).
- Verbose VoiceOver reads the full response instead of only "Response complete".
- Collapsed thinking preview no longer reads stray letters under VoiceOver.
- Bullet markers decorative; VoiceOver reads item text directly.
- LaTeX/math blocks expose a "Math expression" label for VoiceOver.
I watched someone get fired for performance issues that had never once been raised in any of her reviews. When she asked for documentation HR said it was "observational feedback" and not the kind that gets written down. She got a lawyer.
Positive thinking can’t cure disability
You can’t “try harder” to be better
You aren’t “failing” because you got sick
People who believe they will never become disabled or that they will “try harder” than us are denying reality
Disability can and does come for almost everyone
Happy news: a new jam to add to your calendar! 🗓️
This July, the Games For Blind Gamers jam community is hosting its first Midyear Event, called: “Making Games Accessible Jam.”
People often assume Python or TypeScript orchestration performance doesn’t matter because the bottleneck is the LLM.
That’s true for small systems, but at scale, AI Backends are orchestrating huge volumes of concurrent LLM calls, tools, APIs, workflows, retries, and agents.
Benchmarking @AgentField_ai with other popular agent frameworks, we are hyper optimized and efficient with almost 8.2 millions operations per second on a single thread, all thanks to Go.
Excited to share that Impulse Response is now officially on its way towards release! Impulse Response is a totally blind accessible Adventure/Psychological Horror Puzzle-game. The game will be releasing on PC with Mac and Lennox coming at a later date.
So I'm looking at the Apple Foundation Models adapter training toolkit and I've found something interesting. I think the model in its purest form could do image recognition. I've found reference to <image_start> and <image_end> in the tokenizer-config.json file. That is really interesting.
I've connected my Hermes Agent to its own Mastodon account so it can read and post on Mastodon only with my permission of course. This stuff is amazing. It had to work a lot to figure out how to get it to work though.
Humanware are proud to share that BrailleNote evolve has been voted Best Product at this year’s Feel Vision Award during SightCity.
This recognition is especially meaningful because the Feel Vision Award is decided entirely by public vote. It reflects the voices of the people who use, support, recommend, and believe in assistive technology.
Among many outstanding innovations presented at the event, BrailleNote evolve earned the top prize.
Thank you to everyone who voted, visited us at SightCity GmbH, shared their feedback, and continues to support the future of braille technology. This recognition means a great deal to all of us at HumanWare - Linkedin post - https://t.co/FKy1E9nQN6