Sam Altman, the OpenAI CEO, said it's "quite terrible" that so much of the world's limited DRAM sits in personal computers that use it maybe 5% of the time. In his view, it would do more work in a data center.
My take? Idle RAM is the price of owning your computer. Nobody can rent it out from under you, raise its price or switch it off. That ownership is the reason this newsletter exists.
This is a short edition, though. I've been busy lately and couldn't explore much. But there's still enough here to keep your RAM busy for more than 5% of the time. Take that, Altman.
🔍 Discover AI tools
Superwhisper is now on Linux. The popular Mac dictation app now has a native Linux app for system-wide voice input. You can transcribe locally on your own CPU or GPU, or send audio to the cloud for speed. It is NOT open source software.
voiceio is for anyone who'd rather stay open source. This MIT-licensed offline dictation app streams text into whatever app you're using, on both Wayland and X11. It can also fine-tune its model on your own recordings while your computer is idle, and switches to the new model only if it scores better.
OpenKruise Agents is for running AI agent sandboxes on your own Kubernetes cluster. Interesting part is that your API keys never enter the sandbox. An egress proxy pulls them from Kubernetes Secrets and adds them to each request on its way out. It also adds L4/L7 egress rules and an allowlist for MCP tools, so a prompt-injected agent can't wander into your internal network. Overkill for a laptop. Worth a look if you self-host agents for a team.
🐧 AI in the kernel
Linux 7.4-rc1 is out, and its merge window was the biggest on record. 19,671 commits, against 16,267 for 7.3.
Torvalds credits a wave of fixes for bugs that AI found. He'd rather get them during the merge window than as late fixes in the RC stage, and called that a much healthier model. He also admits it may not be all AI. 7.4 is expected to be the next LTS kernel, and a third of the patch is AMD GPU register headers.
👀 Big Tech Watch
Now you know why Altman is eyeing your RAM.
Nikkei Asia reports that off-balance-sheet AI obligations at Alphabet, Microsoft, Amazon, Meta and Oracle grew 90% in six months, to about $3.2 trillion. All those signed data center leases that haven't started yet, plus orders for chips and servers. If they show it in the balance sheet, their shares will tank even further.
Trillions of dollars in data centers need a lot of RAM to fill them. Yours included, apparently (as per Sam Altman).
🗂 AI Jargon: KV cache
When a model generates text, it stores some working data for every token in the conversation so it doesn't have to recalculate it. That store is the KV (key-value) cache.
It grows with context length. That's why a model that runs fine in a short chat can run out of memory once you paste in a long document.
😂 Meme
Plenty more of the memes on our Instagram page.
What are you exploring on Local AI these days?
If this helped you, consider supporting It's FOSS
It's FOSS has been helping people use Linux for the past 14 years. Help us stay independent from big tech. Become a Plus member, enjoy ad-free reading and get 5 eBooks.
Skip the AI slop and SEO spam in your Google results
Tell Google you prefer It's FOSS, and it will show our Linux news and tutorials, created by experienced Linux users, more often when you search. It takes just a click.
Set as preferred sourceGet FOSS Weekly in your inbox
