Ladies and gentlemen... the weekend. (also: you are important and are not alone 🧡)

I recently acquired a copy of the Osborne Computer Corporation bankruptcy filings. In late 1983, several Osborne employees filed a reorganization plan in an ultimately failed attempt to save the company. It's a fascinating look into one of the first media-hyped tech companies to crash and burn in spectacular fashion.

I've uploaded the full 128-page document to the Internet Archive.

https://archive.org/details/in-re-osborne-computer-corporation

The Fediverse is cool

NGC 7253 — a pair of spiral galaxies, four million light years apart, caught in a lovely celestial dance.

Captured by the 8.1m Gemini North telescope on Maunakea, Hawai‘i.

The name of the image translates to “The Twin Spirits Pulling Together Creating Life,” and was chosen by the 2026 Project Hōkūlani interns.

https://noirlab.edu/public/images/noirlab2621a/
More info on the name at https://noirlab.edu/public/news/noirlab2621/
1/n

Shifting costs

Ownership is still the problem, nothing new

Training on content you own. You bought a book? Great, a model you own should be able to use that.

How I imagine a stable version of this: small, specialized models than run on cheap, power-efficient NPUs

How about local AI?

Well they sell Nvidia Spark for that. 😅

And kidding aside, real costs shift to home’s energy usage if you self-host

Also, lots of fear-mongering here too: what if the model is poisoned, etc etc

And this year, because LLM inference products like Copilot made enterprise customers pay for the real costs… Companies end up rationing tokens for employees. So habits that were built during the subsidized API usage create pain points; probably pushing people toward per-consumer subscriptions (which is still subsidized)

So what used to be an enterprise-only income stream, and the biggest income source for Amazon and Microsoft, is now also available for consumers. That’s the business shift in this “AI era”.

The consumers pay for what used to be something that is only for enterprise.

Consuming LLM inference via API usage is the same, except the real costs are bigger (way bigger) and GPU heavy

That’s a different department 😅

I guess if we’re paying a business for ‘cloud compute’ it’s not really their interest to make it easy of fast or simple for us

I’m incredibly grateful to have this opportunity to build something new (green field) for work. More than the potential of the product/s, I get to experience a lot of the friction engineers have to go through

At this point I’m gonna become an expert in Azure pipelines. Might as well

Mindblown by nitro v3

Vitest 5 is here! 🎉

- Big performance improvements (vm pools up to 53% faster, ~18% boost across the board including Browser Mode)
- New builtin Trace View in Browser Mode
- Nested projects support
- New benchmarking API
- New `vi.when` API
- Better defaults
- And a lot of bug fixes!

https://vitest.dev/blog/vitest-5.html

The same repo readme in sr.ht is fine: https://git.sr.ht/~ayoayco/threads#readme