Local AI Weekly #7: Bubble or Not?
The Reuters proclaimed that the AI bubble has already burst in China. Chinese AI and chip stocks have lost about $480 billion in value. In the West, Amazon has put most of its Nova models on “keep the lights on” mode.
And yet, chip maker SK Hynix posted record profits of $29.7 billion in Q3. Their CEO says the memory shortage runs till 2030. So the bubble is popping for some, but not for the folks selling you RAM. Looks like we have to wait some more for the prices to come down for our homelab expansion.
This edition of Local AI Weekly is supported by PikaPods. You can self-host open source software in under a minute with PikaPods. New users get $5 free credit to try it out.
🔎 Discover AI tools
LokalBot records and transcribes meetings on your computer. No bot joins the call hete and you also get dictation and search, all with local models through llama.cpp on Vulkan or the CPU.
OpenNoteTaker does the same job inside a browser tab. Whisper runs as WebAssembly and your audio is never uploaded. Recording, transcripts and exports are free, no account needed. However, the AI-written minutes and translation are paid extras, though, and they send text to the developer’s backend. It’s Chrome and Edge only for now.
Subtitld 26.10 is the biggest release yet of this GPL-3.0 subtitle editor. You can go from raw video to timed, translated and dubbed subtitles in one window. It uses whisper.cpp or Vosk for speech, and Piper, Kokoro and others for voices. Many of these run fully offline.
🤖 The agents keep shipping
Gregory Kurtzer, the man behind CentOS and Rocky Linux, built Ginux in 11 days. It’s a new kernel written mostly in Rust, and it’s not Linux.
A Codex agent wrote about 139,000 lines while Kurtzer set the architecture and checked the work. The “G” stands for GPT, by the way, not GNU. Consider it as a research prototype.
Former GitHub CEO Thomas Dohmke launched Entire, a Git forge for humans and agents working together. The pull requests become “Trails” here with eval scores that agents update on every push. The memory plugins are open source and local-first. The forge itself is hosted, so don’t expect to self-host it.
🗞️ Open model news
Twitter’s original founder, Jack Dorsey’s Block showed off Dessa, a speech model that listens and talks at the same time. It was trained only on human speech, whereas most AI models are trained on text. Block says it runs on a MacBook Pro with M4 Max or better. There’s no word on weights or a license yet, just a waitlist.
Mistral Large 4 is due on 31 October. Just in time for Halloween.
👀 Big Tech watch
China’s STAR 50 index of AI and chip stocks fell 40% in under four months, as per Reuters. GPU makers MetaX and Moore Threads have each lost about half their value since July. Traders blame sky-high valuations and the threat of bans on Chinese systems in foreign data centers.
Amazon has moved most Nova models to “keep the lights on” mode, as Futurism reported. The money now goes into one frontier model.
SK Hynix posted record numbers: 83.3 trillion won (South Korean currency) in operating profit on 104.6 trillion won revenue. That’s an 80% margin. The company expects demand to beat supply through 2030, and big contracts already cover more than half its 2027 and 2028 output. Consumer RAM gets whatever is left, so don’t wait for cheap DDR5 😭
🗂 AI jargon: Open weights
You would have come across this term multiple times in this newlsetter itself. Open weights means the trained model file is public. You can download it and run it on your own hardware. That doesn’t automatically make it open source. The training data and code usually stay private, and the license decides what you’re allowed to do with it.
However, the good thing is that once the weights are on your disk, no company can “retire” them on you.
😆 Meme
⚡ Quick tip: Back up the models you rely on
If you use Ollama for local AI, run ollama list to see your models and their IDs. You can then copy the models folder to your backup drive for safekeeping. It’s usually ~/.ollama/models for a user install, or /usr/share/ollama/.ollama/models if Ollama runs as a system service.
Keep exploring AI locally. See you next week.
