[infecto]

I don’t think AGI is imminent, but there’s already immense value in augmenting human workflows with LLMs. Yes, hallucinations and false positives exist—but I find that criticism often comes from people who don’t use these tools deeply. As a power user, the issue feels overstated and as an easy counter argument. We already are getting to a point where the tools are citing sources. The sources could be incorrect but that would be the same as a Human. As compute cost goes down or model efficient goes up, these problems would appear to be insignificant.

[the_snooze]

As a power user myself, LLMs don't feel like tools I can depend on. I try to use them for well-bounded low-stakes tasks like coming up with sports trivia and generating boilerplate "hello world" code for arbitrary targets (e.g., NES 6502), and they stink at it. Hallucinations aren't a problem you can just wave away because accuracy matters for most tasks. LLMs are less a hammer and chisel, and more of a slot machine that may or may not barf out something of value to me. If they fail at these simple tasks, I'd be a fool to rely on them for anything more substantial.

[infecto]

It’s interesting how varied experiences are. I don’t dismiss hallucinations, but my workflows avoid them by design—I’d never treat the model as a knowledge source, like generating trivia questions directly from it. So I wonder if it’s also about expectations and understanding of limitations. From my perspective I would never create queries like yours without supporting data sets.