That software won't build itself
How does a world where LLMs are superhuman at well defined problems and used more than any single programming language* not end in slop and unemployment? It turns out that someone needs to define the problems.
If I tell GLM 5.2 to "make something you'd like to make", then gave it 1 minute to build whatever it decided on, then repeated this until it implemented Material 3 in Svelte, it would take ~90,539 years for it to do so. That's because GLM 5.2 has a 0.0000000021% probability of randomly deciding to build Material 3 in Svelte.
Sure! I'll implement Material 0.000088% 3 4.3% in Svelte 0.055%.
Then imagine what happens if I had added qualifiers like:
- "in the way that uses the least storage and memory"
- "using the latest language and toolchain features"
- "according to this conversation in this other app"
- "using my preferred tools"
- "using my preferred style"
- "while learning how to implement this" / "while remembering how this was implemented"
- "in a way that creates a sense of place"
- "while having fun"
Defining problems well is managerial but is also a lot like programming - as Karpathy says, the hottest new programming language is English. A lazily defined problem, even if verifiable, will lead to paperclipping - if you ask Opus 4.8 with a multiagent team to port as many browsers as possible to the web it'll make a v86 wrapper. On the other hand, defining tedious ports well nets you high quality results - see Patchzy's Wiicompiled and Puter's firefox-wasm.
But more generally, if you can put what makes you satisfied into words, LLMs will satisfy you. Of course, if your requirements for satisfaction are nuanced, this isn't worth it... yet.
* The bar is low at 24.4% use, so I'm rather confident that English isn't just the hottest programming language but the most popular.