There must be 500 cars per traffic light. Imagine if we put FSD hardware in smart traffic lights, it would save more time and have more safety impact than the same hardware in a small number of cars.
You have to imagine that each conversation has gigabytes of user context, and "utilization" of a big GPU means generating lots of tokens. It does not mean loading all this state just to ask a three word question. But that might be what we (humans) need an AI to do, sometimes.
AI through a social lens: 10 years ago we got algorithms that prevented us from hanging out with real people. Now we have new AIs that prevent us from working with real people. I wonder if a society that places less value on individualism would have better defenses against these things.
Stop it, making kids get up an hour earlier in the dark is a bad idea.
buchanan.house.gov/2026/05/21/b...
It's a little weird commenting on Reddit now, because in the back of your mind, you're thinking, I'm kind of pre-training the next LLM.
I think there is an appetite for using AI to make you smarter, but that the cloud models are structurally unable to do it - for them to do a Socratic Dialogue would be a punishingly bad use of a cloud GPU, endless swapping of user state for only a tiny number of billable tokens.
Wonder if the “obviously AI” text is deliberately written that way precisely so future AIs can exclude it from training
A major cultural problem with AI coding harnesses is around change control. Just before release, human engineers make the smallest changes possible to fix a bug. They go slowly on purpose. AI harnesses don't have a built-in sense of this - a 2000-line refactor is fine at any time!
