Paul Graham said in a recent post on X that if he were 17 years old again, the best use of his time would be to learn how to build a large language model from scratch and train the strongest model he could with whatever hardware was available to him.
The post quickly gained traction in the tech community and had drawn about 418,000 views at the time cited in the source material.
He said he would not start a company at 17
About six minutes after the original post, Graham added a clarification: 「Notice, what I would not do is try to start a startup.」 If he were only 17, he said, he would first build the knowledge base needed for a future company.
His argument was that startup ideas that emerge after someone deeply understands LLMs are likely to be better than ideas forced out of the limited knowledge a person already has at 17. That matches a view Graham has expressed for years: the best startup ideas often come from deep understanding of a specific field, not from deciding in advance to become a founder and then going looking for a problem.
In the current AI cycle, Graham's advice to young people was not to learn how to call APIs first, package ChatGPT, or quickly produce an AI app. He pointed them toward the model layer instead.
Why build from scratch when strong models already exist
A reply challenged the idea directly, asking why anyone should reinvent the wheel when mature models already exist instead of simply making the wheel better. Graham answered in one line: 「To learn how wheels are made.」
That answer captured the gap between training an LLM from scratch and using an existing model. For a 17-year-old learner, a homegrown model is unlikely to beat frontier systems from OpenAI, Anthropic, or Google on either capability or cost.
But Graham's point was not competition with GPT, Claude, or Gemini. The value sits in understanding how data, model architecture, training, fine-tuning, and compute constraints combine to shape model performance. Even if the final model has no commercial value, the knowledge gained through the process could later become the basis for identifying a real startup problem.
A 19-year-old asked whether to go all-in on LLMs
When another user said the more interesting question was what the model should actually be trained to do, Graham replied that if he were 17, he might train a model on something 「unimportant」 such as playing computer games, and he would still learn a lot from the exercise.
A 19-year-old user, @prathamkode, then said he had already trained a 100 million-parameter language model from scratch and was running new supervised fine-tuning, or SFT, experiments in hopes of getting better results at the same model size.
He asked Graham whether, at 19, the right move was to go all-in on LLMs or keep building different small projects at the same time. Graham answered: 「I’d probably try to build an LLM that could solve some specific problem.」
Yann LeCun pushed the discussion in another direction
The exchange also drew in Turing Award winner Yann LeCun. While Graham focused on deep understanding of LLMs, LeCun, who has long questioned whether scaling LLMs alone is enough to reach human-level intelligence, took the conversation elsewhere.
LeCun said he would try to figure out why an LLM can help him write articles but cannot help him clean his bedroom. From there, he said, he would study every field in college and graduate school that might help solve that problem, then look for methods and architectures beyond LLMs that would let machines learn tasks in the physical world quickly and efficiently, more like humans and animals do.
He added that this would not only be his answer at 17. 「In the end, this is also what I would do if I were 30, 40, 50, or 66.」