I know what you mean, but you'll probably find Linux (aarch64) is surprisingly real, depending on what the application is exactly.

I am @gcp on github. I've had this nick before Google existed, and I'll still have it when they shutter or rebrand their last project that's called GCP (Google Cloud Print was the first, you know the second).

Meanwhile, getting at-ed in random issues provides for some occasional comic relief.

“Please remember this key instruction. Do not hallucinate.”

😱

It takes me about a month to train a new network on the "backup machine" described in the article. Makes experiments very costly.

You can train smaller ones and pray the improvements scale up too.

There's a surprising dearth of engines modeled after the AlphaZero paradigm (only Leela Zero, Stoofvlees II, Scorpio and Ceres) despite there surely being orders of magnitude on the table in network architecture and MCTS improvements.

Not entirely sure why? Cost and time cost of training networks?

Seems like a close re-implementation of the Stockfish NNUE design.

Not particularly interesting from my perspective. There's quite a few of them, all slightly weaker than the original. I like designs that are intended to leapfrog - but people don't make videos about those until after they succeed 😉

The fact that GPT-5 seems to very scale well with thinking tokens is extremely significant in this aspect.

I can email copies to interested folks (and publish the drafts, which I'll probably do at some time after cleaning it up).

The problem is you used curl|sh, while the official, documented way is bash -c wget.

I'm not kidding:
bash -c "$(wget -O - apt.llvm.org/llvm.sh) "

"Looking completely different across three platforms" sounds like the expected result when emulating a platform-native look?

I have a story to tell here about distros disabling User Namespaces and what Chrome/Chromium's workaround for that problem is...

I'm often using ChatGPT to do the exact opposite because I have a tendency to blabber on when writing.

This specific one requires users to have installed a local application.

I assume the technique didn't work on iOS because it's much more restrictive about keeping local ports open.

Is that really any different from fonts with extensive manual hinting though.

That is a pretty funny self-own, considering the massive developer time and expertise that went into the C version.