Paul Graham: at 17, I'd learn to build LLMs from scratch

Key facts

• Paul Graham posted on X that if he were 17, he would not try to start a startup: instead he would learn to build large language models from scratch and train the most powerful ones he could on whatever hardware he could get access to.
• He argues that far better startup ideas would grow out of deeply understanding LLMs than out of trying to start companies based on what he knew at 17.
• He says he would also try to figure out why LLMs can write his essays but cannot clean his bedroom, then study topics in college and graduate school aimed at solving that problem.
• He adds that he would look for "methods and architectures beyond LLMs that can quickly learn to perform physical," but the post, as captured, cuts off mid-sentence at exactly that point, so it never names the task or field he means.
• The post names no specific hardware, model architecture, training technique, field of study, or date anywhere.

Why it matters

Graham stakes out a specific ordering of priorities for someone starting from technical skill rather than company-building experience: build deep, from-scratch understanding of how large language models work first, and let startup ideas grow out of that understanding later, rather than starting a company on the more limited knowledge he says he had at 17. He frames it explicitly as an alternative to founding a startup at that age, not as a general recommendation for people at other stages of their careers.

Who it affects

The post answers a question about what Graham himself would do at 17, so it speaks most directly to teenagers and students weighing whether to build technical depth in LLMs first or start a company right away. It also touches anyone interested in the specific gap he raises: why current LLMs can write essays but cannot clean a bedroom, a question he frames as worth studying in college and graduate school rather than one he has already answered.

How to use it

As described, the path has two parts. First, learn to build LLMs from scratch and train the most powerful ones feasible with whatever hardware is available, instead of moving straight to founding a company. Second, later in college or graduate school, study topics aimed at closing the gap between what LLMs can already do, such as writing essays, and what they still cannot, by looking for methods and architectures beyond LLMs themselves. The post names no specific course, hardware platform, or technique for either step.

How solid is it

This is a first-person opinion Graham posted directly to his own X account, not a data-backed claim or a research finding, so there is nothing external to verify beyond the wording of the post itself. The source identifies him only through the byline "Paul Graham on X," with no further detail about who he is or what, if anything, qualifies the advice.

Risks and caveats

The post, as captured, cuts off mid-sentence right after Graham writes "methods and architectures beyond LLMs that can quickly learn to perform physical," so it never says what task, activity, or field he means by that. It also gives no date and names no specific hardware, model architecture, training technique, or field of study for either the LLM-building step or the later college and graduate-school study. It is a personal, hypothetical answer to a question about his own life, not a study or a demonstrated track record of the approach working.

“Someone asked what I'd do if I were 17. I'd learn how to build LLMs from scratch, and then train ones as powerful as I could with whatever hardware I could get access to.”

— Paul Graham, on X