Older (40+) Retired Programmer
Game client development (in the past)
Freelancer + Indie Developer
Following/Researching the latest AI technologies
Bazi/Fortune
After 20 years in software, the bill comes twice: generation, then rework. Without tests, the rework tax quickly overtakes the speedup. I tier it: fast models for drafts, frontier models for convergence. Has anyone quantified that second-half tax on a real production repo? @karpathy
@simonw hit the exact same trap trusting tool call json straight into the db upstream hallucinated tenant id and is admin keys cross contaminating rows minimal fix strict schema whitelist drop everything else before update where does your sandbox draw the line on output sanitization
Ran the same tool-call task 30x on 3 models. The benchmark ranking was the least useful number I got.
Real cost per finished task, retries included: DeepSeek-V4-Pro 【$0.04】, gpt-oss-120b 【$0.09】, Llama-3.1-8B 【$0.15】.
How much retry tax are you eating?
@nic0203 openAI dropping $200 api credits is pure desperation. everyone actually vibe coding is just risking the claude ban on cursor anyway. better than babysitting gpt's lazy code. @nic0203
@itz_komal01 None of the above. In 2026, raw syntax is commoditized. Beginners should focus on Vibe Coding (Cursor + Claude). But if you absolutely must suffer through learning a language from scratch to understand how computers actually work, throw this list away and learn Rust @itz_komal01