Carmack: AI writing software from scratch isn't optimal — well-regularized building blocks matter
ID_AA_Carmack · x · 2026-10-30
John Carmack argues that while AI may write massive software projects "from scratch," that's almost certainly not the fastest, most performant, most secure, or most token-efficient path — good building blocks matter.
Key points:
- Beyond RL on final task loss, we should train agents to regularize the building blocks too — what would a standard library look like that makes agents most effective?
- At the limit it may be some alien code organization, but given human bias in training data, he bets near-term it would be lovely code old programmers would appreciate
- He suspects the same applies to mathematics, though that's out of his lane
More from AGI Musings
- Blogger mocks 'pause AI' plan: humans can't even agree on basic treaties — VraserX · 2026-10-30
- Dean Ball: 21st-century institution building has much to learn from mycology — deanwball · 2026-10-30
- Why every lab is racing toward RSI: LLMs will be the ones discovering new architectures — haider1 · 2026-10-30
- The Obsoleting Project 'Is Building a Dystopia': The Bleak Reality of Job Hunting Now — GarrisonLovely · 2026-10-30
- AI-era entertainment has barely changed, argues signull — the paradigm shift is still ahead — signulll · 2026-10-30
- Shopify CEO Tobi Lütke: AI agents earn basis points — need tens of millions of users to work — rohanpaul_ai · 2026-10-30