Barak Michener

@barakmich.bsky.social

TIL Stockfish in 16.1 dropped all classical evaluation heuristics in favour of NNUE (aka, a machine learned evaluation function). I get that it performs better, probably faster, definitely less code, but some serious explainability was lost.

Watching deepseek-r1:671b run locally on my AI box (the 1.58 quant version, I’m not a billionaire) is like catting a file over 900 baud. At 4 tokens per 3 words on average — and I’m getting about 4 tok/s — it’s really, back of napkin, ~50 baud

An interesting emergent property of switching to jujutsu for version control is I have more active ‘branches’ and even empty-but-described commits with “what I intend to do here” Because the rebases are so powerful and cheap, I can get code & thoughts out of my head and worry about merging later.

TFW you use a right-wing jerkwad’s useful blog post for the opposite (in this case rename git master->main instead of the reverse (because ‘main’ is apparently offensive to dumbasses))

Also, hello BlueSky! This is going to be my barking-about-technology-into-the-ether place now, because (A) fuck the muskrat running Twitter, and (B) I tried my hand at Mastodon but also have no time anymore to self-host... and if I'm gonna pick somewhere, might as well where the people I know are.