Post

I wonder: could we as a society just run LLM-accellerated scientific programs for new battery/solar designs, CRISPR-like enzymes, cancer cures - the stuff where outcomes really matter - while also maintaining a humans-only sandbox for these people to experience the joys of the life of the mind in?

We already have this split for everything. The folks down at my local tennis court don’t think they’re about to be on the pro tour. It’s fine.

I think some people are not cognizant of what side of the split they are on. Explains some weird arguments about software and AI I have had with people who believe that software is a creative endeavor (true), and Literally Every Line Of Code Is A Sacred Expression Of Programmer Creativity (false).

"I'm not a CRUDslop generator, I'm a temporarily embarrassed Turing Award winner!"

I mean not even that, it's just. There is some amount of irreduceable mundane labor that must be done in the process of lowering our ideas into executable programs. AI can't think things out for you, but once you HAVE thought things out AI can help convert your thinking into code.

And the more you want to make the code the best possible expression of your thinking, the more effort it takes! Readable, moduler, maintainable, well-tested, easy-to-understand code that clearly lays out your thinking requires clearer thinking and MORE TYPING than a mess of spaghetti.

Funnily enough Shigeru Miyamoto recently said some very similar things regarding what he thinks about AI in game development? AI can mean the developer spends less time laboring on the code and more time thinking about what the right choice is to make the game the best it can be.

At least, I think he said that. The embers of the internet argument I had with someone who is insisting that can't POSSIBLY be what he means are still smoldering.

Eventually the AI will do everything and we all kill ourselves.The rest is cope.

yeah, hard agree, this is the story of so many abstractions. we let optimizing compilers turn our type-safe instructions for execution into machine code; we let query planners turn our SQL into loops and lookups; why wouldn't we let a jumble of weights and matmuls turn our specs into boilerplate

And the arguments that come back are, generally, "but if you have not written code you may not understand how to write a good spec for what you want" and "some code is very specific in how it is written so the spec is as much effort as the code."

To which my responses are 1. I never said you couldn't write code. I recently set up a Quake engine to build with buck2 on Linux and I did it all by hand because I wanted to really understand buck2. NOW that I have had that experience, I'll likely have AI write way mote of my Starlark.

"but how do you know the boilerplate's correct?" well, bsky.app/profile/arac...

Yeah the grand and beautiful irony of AI is that in order to utilize them most effectively you do actually need to have the empathy and humility to understand that all programmers are flawed and try to make your code testable and your interfaces as hard-to-misuse as possible.