Paul Graham vs Yann LeCun: What Would You Do at 17?

A question about what one would do at 17 drew opposing answers from two AI-industry heavyweights: Paul Graham recommended building LLMs from scratch, while Yann LeCun argued for exploring architectures beyond LLMs. The contrast reflects a broader split in AI between deepening the current paradigm and pursuing the next one.

Confirmed

• Paul Graham, co-founder of Y Combinator, said that at 17 he would learn to build large language models from scratch and use whatever hardware he could get to train the most powerful model possible; per @petrusenkomax, the advice was about mastering the technology first rather than rushing to start a company.
• Yann LeCun, replying to Graham, said at university or graduate age he would study why LLMs can write essays but cannot tidy a bedroom, and explore new methods and architectures beyond LLMs aimed at efficient physical-task execution like humans and animals.
• Elvis of dair.ai added that builders today, at any stage, should own their intelligence stack.
• @rohanpaulai recommended two "build an LLM from scratch" tutorial videos for 17-year-olds wanting to learn.

Why it matters

• The two answers form a sharp contrast: Graham emphasizes hands-on replication to deeply understand the dominant technology, while LeCun points to LLMs' fundamental limits on physical-world tasks and pushes research toward next-generation architectures — a useful reference point for young researchers choosing a direction.

2026-08-23 ~ 2026-08-25 · 5 related posts

• Paul Graham suggests 17-year-olds should build LLMs from scratch — samiramanabi · 2026-08-23
• Yann LeCun: If young, I'd study architectures enabling efficient physical tasks beyond LLMs — Fowe · 2026-08-24
• Two recommended videos for 17-year-olds to build LLMs from scratch — rohanpaul_ai · 2026-08-25

2 near-duplicate retellings: dair_ai · petrusenko_max