Post

The discourse isn’t simply about whether they are useful. The main criticism is AI that wastes an inordinate amount of natural resources, not just money, compared to the use we get out of them.

People thought I was damned weird back in 1994 using my computer to type messages to people over phone lines.

Who are you talking to? 😅 Even engs at my company are past that after they've seen cursor / claude / opencode.

I feel your pain. I never got beyond rudimentary regex. Now it takes maybe 30 seconds. I just need the LLMs to quit feeling like they need to explain what it does. Hell with that, just gimme the code.

People who aren't developers just don't see it as easily. I can do things with opencode and a good model that would look like sorcery five years ago. Not every field will see this upheaval, but it's revolutionary for software. (Not discounting the problems with GenAI: power, IP, megacorps, etc)

huh, surprising, wouldn't 8-bit have been sufficient? or was everything fp16? my memory of the specifics is hazy now

fair grievance to direct at microsoft and google, who are guilty of this

Pointing to bad individual product decisions does not equate to stating there is "no valid use case," which is what Ryan was referring to. These things are useful. They are also misused. Both of these things are true. Any critique that does not acknowledge the first point is unserious.

There are a lot of of people building software rn., and coding agents will extend that. Not many people need climate models on their laptops, but building an bitart game…

Just last night I had a frontier model nearly one-shot implementing associated types with fully passing tests in ts-go, mostly as a benchmark to see how far models have gotten. Obviously the next step is seeing if a model can design and implement the final boss that is higher kinded types 😈.

It does sometimes feel like punching down.