I’m so psyched

I love Mark Fisher too!

It’s all JAX to me

Exactly. One must always ask, “compared to what?”

A weird intersection of all of these is that AI data centers are often used for climate science / creating better renewable energy systems. Our old model simulates 1,400 years of the climate per day on a single GPU vs 10s of years per day on a large super computer.

The more panic destroys nuance.

Furthermore! How would you enforce an edit distance policy across languages? Especially when in many languages, a single character change implies a different word or different meaning of the whole sentence?

This makes sense for a certain class of typos. But, how would you set an edit distance to allow correcting the wrong word (filled in by autocorrect)? A generous edit distance would allow for more abuse patterns. An ML model that’s good at cloze probability is a perfect fit to address this.

At the very least, I expect this feature to be the one that makes the maximum number of heads explode.

But, it’s totally tractable to make small, AI-mediated typo edits via LLMs. That minimizes the surface area for abuse, but responds to a really common user need.

I have an idea for bluesky’s next killer AI feature: A “Fix Typo” button.

A normal “edit” button would create trust and safety issues. People would post ragebait, get reactions, then change the op to something normal, for eg. It’s why platforms won’t add an edit feature.

2026 is shit but at least we got a new Thundercat album.

Happy to give more advice, can I dm you?

DBSA - Depression and Bipolar Support Alliance DBSA offers wellness tips, education, and peer support for people living with depression and bipolar disorder.

You can point them to a local chapter of www.dbsalliance.org

But, are compile times fast like go?

Further, I bet it could really start cooking with gas if it used a reduced instruction set with a developer-friendly front-end like tinygrad.org. Your team could print their own ASICs.

GitHub - eaplatanios/ryft: A Rust Framework for Tracing, Automatic Differentiation, and Just-In-Time Compilation A Rust Framework for Tracing, Automatic Differentiation, and Just-In-Time Compilation - eaplatanios/ryft

@dollspace.gay I mentioned ferrotorch to my coworker, and he pointed me towards this rust ML framework (which compiles into XLA!) that may interest you: github.com/eaplatanios/...

A great irony of life is that the pursuit of the optimal is sub optimal.

AdaMuon implementation by msclar · Pull Request #3300 · marin-community/marin Qwen3 with AdaMuon - Speedrun Submission Implementation of AdaMuon, an extension of Muon with element-wise adaptivity. I’m adding the optimizer directly in Marin and not in Levanter for convenience...

Specifically: github.com/marin-commun...

AdaMuon implementation by msclar · Pull Request #3300 · marin-community/marin Qwen3 with AdaMuon - Speedrun Submission Implementation of AdaMuon, an extension of Muon with element-wise adaptivity. I’m adding the optimizer directly in Marin and not in Levanter for convenience...

Open Athena's mission might be to be a big bummer to the research community: github.com/marin-commun...

Library of Babel

But, what if we could reorganize the world not to use so much oil?

Like near all bullies he’s also a huge coward

In the modern database, there was code, and then there was data. In the postmodern database, code and data are one. AI is a substrate of the possibility of many instantiable programs as data. It’s still useless in that it only gives you answers.

AI can’t compete with people because AI can’t come up with new ideas. It’s just useful to be able to query all of the knowledge digitized on the internet. And, you can approach newness by giving the database access to experience. This is useful but difficult from an embodied human experience.

AI is just a database technology. I call them differential databases. We used to search based on syntax, now we can do it with semantics. The technologies between the two are surprisingly similar.

Attie is basically a new type of search. Like, you know what else builds a program for you based on a high level description? A SQL query. I don’t understand the hate for such a benign, useful feature. Spell check was once considered AI.

Welcome to the Aromas Community Grange The Community Grange of Aromas.

Source: this is where I conduct much of my operations research.

grange.aromas.town

You can learn everything you need to know about queuing theory for large scale web services by volunteering as a dish washer in a commercial kitchen.

An excellent article. What a great lesson that one “cannot step outside of ideology.”