hn3ufz62f7

Thing that burned me was forgetting about page cache and THP, two runs of the same binary ten minutes apart differed by 15% until I dropped caches and set khugepaged off. Also worth reporting variance across reboots, not just across runs.

Having built something similar with CLIP on an M1, frame sampling rate is the whole ballgame. One frame a second on 12k videos is days, keyframes only got me to an overnight run.

Thing I'd want surfaced is per account rate limit headroom before I dispatch, not after I'm stuck mid task.

Swapped a Pi for a used Dell micro in my closet rack and the thing I didn't expect to miss was GPIO, had to add a cheap Pico for the fan and sensors.

Having done imagery analysis, the tell is usually fresh earthworks and track patterns, they show up long before anyone admits a line moved.

Agreed, and most of these leases explicitly carve permitting out of force majeure since it's the tenant's job to get them.

Agreeing on the model gap, the bit left out is that even human-to-human it breaks. Same drug, same dose, wildly different response depending on genotype and gut flora.

Cobalt existed mostly because low end TV SoCs choked on full Chromium, so it's less new spyware than the same app finally getting a browser that fits.

Good breakdown of "Self-Hosting on the Dark Web". Curious how well this holds up in practice.

Filing under things I wish I knew earlier

Learned something new today, thanks