The intent paradox of AI generated code(feruchemy.bearblog.dev) |
The intent paradox of AI generated code(feruchemy.bearblog.dev) |
Imagine a simple todo application. Instead of a conditional filter on completed items the code itself changes to always and only return completed items upon request, spoken or otherwise.
This will usher in the era of "hypermature optimization".
It's pretty crazy how many problems we have are people trying to solve already solved problems
I’ve need no evidence that AI writes particularly performant code (especially wrt to specific hardware.)
Nor have I seen any big player showing off models which were tuned for performant code generation.
Code performance work is almost always very specific to the codebase at hand and not at all general.
Not much changes between billing models, hr models, even custom financial models aren't THAT custom these days.
Consolidation of best practices across diverse technical fields has happened rapidly fast in the last 10 years.
So the AI catches things, business managers, project managers, business analysts, and programmers miss during initial project scoping and design.
Probably the same law that limits compute power wrt to space and energy.
Look at how much energy goes into training GPT 4.5 vs it’s improvements (on our own potentially bunk benchmarks, granted)
Also like… literally nothing in the universe advances exponentially forever.
Whether or not any of that is valid, I’d like an answer to my question.
Is the only reason you think that a blind belief that AI will advance exponentially in all domains?
I’m not sure any of that was relevant to what I was saying.
And you still didn’t answer my question.
If the improvement curve is linear, it would be prohibitively expensive as and if logarithmic, it’d be impossible.