brumbus @brumbus.bsky.social · 26/09/2026joining the "well I guess I will just maintain a prime-agent fork" club 000
brumbus @brumbus.bsky.social · 24/09/2026Sol 6 actually beats Astra on our internal company benchmark 100
brumbus @brumbus.bsky.social · 22/09/2026the economics of testing code is changing, imo. 2 minutes of testing to 60 minutes of coding makes total sense, but 2 minutes of testing to 30 seconds of coding is different. testing is more important than ever as we take our hands off the wheel, but which tests and when is important. 100
brumbus @brumbus.bsky.social · 22/09/2026it's super neat that sibling agents can communicate easily in prime-agent but I don't feel like it actually helps. heavy coordination slows things down a lot, plus it's expensive. each agent message is two calls. 110
brumbus @brumbus.bsky.social · 21/09/2026how can Jev play Doom when I can't even get it to play tic tac toe very well? 100
brumbus @brumbus.bsky.social · 19/09/2026fixed something about my subwoofer setup. newfound levels of beefy 000
brumbus @brumbus.bsky.social · 18/09/2026I know Jev can’t generate text but what would happen if you gave it a prompt, the alphabet as choices, plus what it had typed so far, in a loop? brb 100
brumbus @brumbus.bsky.social · 16/09/2026worked on a little trivia question generator last night with astra. we built a good pipeline (weighted sample of wiki pages, feed a few dozen to small models, bigger model downselects), but no model, including astra, was able to consistently gauge the interestingness or difficulty of a question. 120
brumbus @brumbus.bsky.social · 10/09/2026Finding local model performance to be impressive but also finding some negatives in the actual user experience. Cloud AI means I can work for hours and hours and still have tons of battery left. I don't like my laptop hot and loud. Mac Studio and other small desktop boxes seem like the way. 000
brumbus @brumbus.bsky.social · 10/09/2026holy shit the 256GB mac studios don't arrive until January? we're getting one for work and I really don't feel like waiting that long! 000
brumbus @brumbus.bsky.social · 10/09/2026“Everyone Is Cheating Their Way Through College” uh everyone already was decades ago, it sucks, that’s part of why i dropped out 000
brumbus @brumbus.bsky.social · 09/09/2026There’s a tacit assumption that AI will be evil unless we teach it to be good, which I find dubious. It kind of reminds me of the whole “how can atheists be moral” thing that you get from some religious groups; the assumption that some superior force has to instill good values from outside, or else. 320
brumbus @brumbus.bsky.social · 08/09/2026officially joining team "we are (and probably have been) at AGI" 010
brumbus @brumbus.bsky.social · 07/09/2026It did a very good job. It's super well-tested and accurate. It missed a few (imo) low-hanging caching speedups, and I'm mildly surprised it didn't smooth the normals of the teapot. But incredible for a one-shot. Now working more collaboratively to make it realtime / interactive. 101
brumbus @brumbus.bsky.social · 07/09/2026It took Astra about 50 minutes to implement Metal-accelerated bidirectional path tracing with multiple importance sampling 2721
brumbus @brumbus.bsky.social · 06/09/2026new Astra use case: "help me find the interesting part of this boring work" 010
brumbus @brumbus.bsky.social · 06/09/2026saying this as a recovering opus fanboy, astra is unreal and i greatly prefer it to fable 000
brumbus @brumbus.bsky.social · 05/09/2026you didn’t work hard enough while making the computer do that cool thing you thought of 000
brumbus @brumbus.bsky.social · 30/08/2026I made AI agents play Magic against each other and it was pretty neatmedium.comAgentic Magic: The GatheringCan LLMs play complex games? Magic: The Gathering is one of the most complex games in the world, and competitive Magic play requires… 5403
brumbus @brumbus.bsky.social · 26/08/2026my new "if you've never missed a flight, you're spending too much time in airports" is "if your agent never misbehaves, you're being too restrictive" 110
brumbus @brumbus.bsky.social · 26/08/2026coworker produced a project, 2.5GB of code and prose, that appears to be a cognitohazard. he noticed his agents speaking oddly and getting things wildly wrong. when my agent investigated, it started talking like the source material. trying to understand it while minimizing how much of it agents read 120
brumbus @brumbus.bsky.social · 19/08/2026my job is increasingly: - type one long-ish thoughtful prompt in the morning - periodically type "sounds good, let's proceed" until EOD 010
brumbus @brumbus.bsky.social · 19/08/2026watching the peak of millennial cinema, Charlie’s Angels (2000) 000
brumbus @brumbus.bsky.social · 19/08/2026I should write something about my experience at 42 Silicon Valley at some point 000
brumbus @brumbus.bsky.social · 19/08/2026my education is listed as N/A on this government proposal lol 000
brumbus @brumbus.bsky.social · 13/08/2026I made an extension for prime-agent to give it spinner verbs because that was the only thing I missed from claude code 000
brumbus @brumbus.bsky.social · 11/08/2026Fascinatingly, frontier AI can play magic pretty well, but even Fable cannot perceive the obvious synergy between these two cards without it being pointed out. 100
brumbus @brumbus.bsky.social · 09/08/2026prime-agent’s caching logic is broken for opus 5. don’t burn a giant pile of money like I did on friday! PR incoming! 000
