Claude agrees!
● Eric makes an excellent point about AI economics!
The old calculus: "Is feature X worth 3 weeks of human effort?" → Often no.
The new calculus: "Is feature X worth 30 minutes of prompting + review?" → Almost always yes.
The EV2-TUI backend is a perfect example:
- Human-only: Weeks of work, deep EV2 knowledge needed → "not worth it"
- AI-assisted: A few hours of iteration → "why not try it?"
The risk/reward shifts dramatically. Even a "failed" experiment costs almost nothing. You learn something, and the code might be useful later.
My honest assessment of effort:
- Zig POC (today): ~2 hours wall clock, mostly waiting for compiles
- WAV loading (yesterday): ~30 min actual work
- An EV2-TUI backend: Maybe 4-8 hours of prompting/iteration
The limiting factor isn't writing code anymore - it's understanding what to build and reviewing the output.
MY 2 CENTS
I really wanted to make a simple_zig but that experiment didn't work out as I thought. It requires more of that. Thinking, that is.
The simple_tui was just fun to make and include. Besides, the competitors have their own, so why not?
As for the WAV loading for simple_audio — that effort seemed to take forever because Claude was running out of memory resources (e.g. I think I need more RAM, 32 GB is clearly not enough). After rebooting and recovering after some Claude files got corrupted in the process, the task finished in under 30 minutes from where it left off to the commit and push.
All in all — the amount of energy required to produce something (even if just a "toy") is so inexpensive that such experiments are worth the effort. The Simple Eiffel experiment has not been a waste. For the most part, I have accomplished what I set out to know personally, which is: Can I successfully use AI in an Eiffel context? I think the answer is yes, even in light of Claude being overly eager, overly confident about itself, and overly presumptuous in it all.
What I haven't really addressed is the simple_kb (Knowledge Base). The goal was to get to an extended Ollama AI QLoRA, but in the end, the experience may demonstrate that the KB binary with its SQLite DB and RAG connection to various AI (Claude, Gemini, Grok, Ollama) is good enough. What sparked that was Jimmy asking about across loops. I had hoped that the KB could be a resource for programmers (new and old alike) to ask such questions and get non-hallucinogenic answers. That was the other reason for the TUI — namely — to put that KB into some kind of program with at least a console UX to facilitate it and make it useful as I wasn't ready to create simple_ev2 or go beyond EV2 into more modern territory (which thing may yet happen).