Your hypothesis is confirmed by “Because nothing says Fall like financial progress.”

Only because the average user would have no clue how to find and disconnect the piezo speaker, assuming they even know about it. And, of course, it’s a lot more distracting than sound coming though someone headphones or whatnot.

It’s not as common now, but lots of desktop computers used to have little piezoelectric speakers that would beep to indicate error codes or catastrophic failure.

They’re called speakers but they’re only good for buzzing at different frequencies instead of accurately replicating source audio. Interestingly, the more advanced ones can be used quite effectively to make ultrasound signals.

It took me 20-30 minutes to open up my G502 and install new switches. Of course, you probably want to have some soldering experience to do this, but if it’s already unusable… why not try? (assuming you know someone who owns a soldering iron)

I bought an Incott G24 recently, which is a Chinese clone of a Zowie EC2. The software is shit (I set it up once and never used it again) but it has hotswappable switches for M1/M2 which is something I’d love to see in mainstream mice, but they would lose money for it, so it’ll never happen.

Cadence of Hyrule’s jazzy Song of Storms rendition. It’s very fitting as a ringtone for those out of the loop, and those in the loop usually get a kick out of it.

I don’t think it’s fair to think of model training as a one-and-done situation. It’s not like deepseek was designed and trained in one attempt. Every iteration of these models will require retraining until we have better continual learning implementations. Even when models are run locally, downloads signify demand, and demand calls for improved models, which means more training and testing is required.