The Eternal Sloptember

George Hotz (geohot) argues that AI coding agents are fundamentally incapable of real programming and that their widespread adoption will be one of software development's costliest mistakes. Drawing on six months of personal experimentation with agents in tinygrad and hardware reverse engineering, he concludes agents frontload progress but never achieve polish. He distinguishes between AI as a useful search/prototype tool versus a genuine software engineer, arguing it falls far short of the latter. He warns that large organizations with weak feedback loops will suffer most, as low performers produce high volumes of low-quality 'slop' without the self-correction ability of high performers. He aligns with LeCun and Marcus in believing current LLM architectures cannot truly program, and that real programming agents would require world models rather than RLVR-style hacks.