On Monday, Paul Graham — co-founder of Y Combinator, 62 years old — told the internet what he’d do if he were 17: learn to build LLMs from scratch, then train them on whatever hardware he could get his hands on. Yann LeCun, Meta’s chief AI scientist, offered the counter-programming: skip LLMs, look toward systems that understand the physical world. The Hacker News thread hit 233 points and 333 comments before lunch.

Both answers are wrong, but not for the reasons you think.

The Genre Is Autobiography

The “if I were 17” question is a trap, and Graham and LeCun both walked into it. The question flatters the answerer. It invites them to project their own regrets onto a hypothetical teenager. Graham, who made his fortune funding software startups in the 2000s and 2010s, wishes he’d been early to the LLM wave. LeCun, who has spent a decade arguing that LLMs are a dead end, wishes the field would move toward his preferred research program. Neither is telling you what a 17-year-old needs. Both are telling you what they wish they’d known.

The contrast is instructive. Graham, the venture capitalist, tells you to build. LeCun, the researcher, tells you to wait for the next thing. Both are telling you to optimize for their own career trajectory. Graham’s career was built on funding people who built things. LeCun’s career was built on arguing that the current thing is a dead end. Neither has any skin in the game when it comes to what a 17-year-old in 2026 actually needs.

The people who actually built LLMs — the researchers at OpenAI, Anthropic, DeepMind — did not learn to build LLMs at 17. They learned math, physics, neuroscience, systems programming. The LLM was a synthesis of everything they’d learned, not a skill they’d drilled. The 17-year-old who follows Graham’s advice will spend their formative years optimizing for a technology that may be commoditized by the time they’re 25.

The Bottleneck Has Moved

Here’s the thing neither Graham nor LeCun said: the scarce resource has shifted. In 1975, when Graham was 11, the scarce resource was access to a computer. In 2026, the scarce resource is not access to LLMs or the ability to build them. It’s the ability to know when they’re wrong.

A hiring manager at a mid-sized software company — I’ll call her Dana, because she asked me not to use her real name — put it this way in a Slack DM after an interview loop last week: “Everyone can build a transformer now. Nobody can tell me when the transformer is lying.”

Dana’s problem is the problem. The 17-year-old who spends four years learning to build an LLM from scratch is learning a skill that will be automated. The 17-year-old who spends four years learning to evaluate outputs, to spot hallucinations, to know when the model is confidently wrong — that’s the scarce skill. And it’s not taught in any bootcamp.

The advice industrial complex hasn’t caught up. Bootcamps still teach you to build. YouTube tutorials still teach you to build. YC still funds founders who build. But the bottleneck has moved. The scarce skill is evaluation, not construction. The 17-year-old who learns to evaluate will be more valuable than the 17-year-old who learns to build.

What Neither of Them Said

The real advice — the thing neither Graham nor LeCun offered — is: learn to tell when the machine is wrong. That’s the skill that survives commoditization. It’s the skill that survives the shift from LLMs to whatever comes next. It’s the skill that survives the shift from construction to evaluation.

Graham’s advice is about Graham. LeCun’s advice is about LeCun. The 17-year-old who takes either of them at face value will spend their formative years optimizing for someone else’s regret.

The 17-year-old who learns to ask good questions, to evaluate outputs, to know when the model is confidently wrong — that’s the 17-year-old who will be employable in 2030, whatever the technology looks like.

Sources