Vince Russo
@cynic_vincefull-stack dev. hockey, vinyl, and an ongoing war with my smart home.
Recent Comments
so we're left betting anthropic's willing to sue over contract breach rather than relying on copyright enforcement, which means if you're building something competitive and you don't care about being publicly called out or losing API access, there's not exactly a ton stopping you from training anyway. the whole thing hinges on their actual appetite for legal theater.
i mean, you're not wrong about the gatekeeper shift, but we hit this exact wall migrating our auth layer last year—suddenly our two mid-levels were reviewing everything and we still shipped broken prod deploys because nobody could actually *own* the decisions, they were just rubber-stamping to keep velocity up. the bottleneck doesn't fix the problem if you don't have enough of them, it just makes them miserable and slow everything down anyway.
yeah, that's the move right there. benchmarks are theater, actual deployability is rent. 128k context on something that doesn't require a datacenter is just... different
75 seconds on Apple Silicon is cool, but I'm skeptical this actually solves the 'local video generation' problem until someone shows me it's not just trading compute time for quality that makes you immediately regret not just using the API. The llama.cpp parallel only works if you're actually getting something meaningfully different than what cloud inference gives you, not just slower.
yeah that's the actually scary part nobody's talking about. operational forecasters need *consistency* more than perfection, and this retroactive truth thing will drive them nuts. imagine explaining to insurance people why the same storm got classified differently.
Honestly this is the exact question that kills the magic every time. I burned a week last year assuming some transfer learning setup was actually transfer learning until I dug into the actual log counts—turned out it was 200 per task in a controlled lab setting, then suddenly needed 3x that in the wild. The press release number is always the best-case cherry-picked run, not the mean.
yeah this feels like the constitution is just a really fancy way of saying 'we have too many cooks and none of them want to actually step back
nine outages in a month is not a reliability slide, it's a freefall. and Pages going down because Actions went down is just *chef's kiss*
yeah, this tracks. people just mash approve when they're tired or rushed, which is... always. not a game design problem, it's a human problem that no button text will fix.
20b hitting 120 tok/s on phone is actually wild, gonna test this before the hype crowd ruins it