Learning to build LLMs from scratch
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Paul Graham advised 17-year-olds to learn building LLMs from scratch. Yann LeCun countered, urging focus on physical task AI.
Background
Paul Graham, co-founder of Y Combinator, posted on X that if he were 17, he'd learn to build LLMs from scratch and train them with available hardware, rather than starting a startup. His follow-up explained that deep technical knowledge would lead to better startup ideas later. Yann LeCun, Meta's former Chief AI scientist, replied that he'd instead investigate why LLMs can write essays but not clean a bedroom, and study methods beyond LLMs for physical tasks. The exchange sparked debate among users about whether to focus on building models from scratch or using existing open-source models.
How it unfolded
- Aug 23, 2026Paul Graham posted on X that if he were 17, he'd learn to build LLMs from scratch and train them with whatever hardware he could access.
- Aug 23, 2026Graham followed up, saying he would not start a startup but build a foundation of knowledge, as better startup ideas would grow from deep LLM understanding.
- Aug 24, 2026Yann LeCun replied, saying he'd try to figure out why LLMs can write essays but not clean his bedroom, and study methods beyond LLMs for physical tasks.
- Aug 24, 2026News outlets like Analytics India Magazine and Times of India covered the exchange, and social media users debated the practicality of Graham's advice.
Who’s saying what
- Paul Graham
- Learning to build LLMs from scratch is the best foundation for future startup ideas.
- Yann LeCun
- Focus on understanding why LLMs lack physical capabilities and study methods beyond LLMs.
- Public
- Some users argued that working with existing open-source models like Qwen and Kimi is more practical without expensive hardware.