Paul Graham offered a concise answer to a question about what technical direction he would pursue if he were 17, identifying the construction and training of large language models as his preferred area of study.
In a post published through his account on X, Graham said he would learn to build LLMs “from scratch.” He added that he would then try to train models as powerful as possible using whatever computing hardware he could access. The statement presented both model development and the practical limits of available equipment as central parts of the proposed learning path. [Source 17764]
The post did not provide a curriculum, list particular programming tools or recommend a specific model architecture. It also did not define what level of hardware access would be necessary. Graham’s answer was instead framed as a direct response to the hypothetical question of how he would spend his time at that age.
His wording placed the emphasis on constructing models rather than only using existing language-model services. By pairing the goal of learning to build an LLM with an effort to train the most capable model that accessible hardware could support, the post described an approach based on direct experimentation within real resource constraints.
The reference to building systems from scratch was not accompanied by further explanation. The source therefore does not establish whether Graham intended that phrase to cover every component of a language model, or whether he was recommending a particular degree of independence from existing software and research. No such technical assumptions are necessary to understand the central recommendation: learn how the models are made and test that knowledge by training them.
The statement also stopped short of claiming that this path would guarantee a job, produce a commercially successful system or outperform other forms of technical education. Graham did not compare LLM development with university study, conventional software engineering or other possible subjects for a 17-year-old. His post identified what he personally would choose under the scenario presented.
Although short, the recommendation joins two distinct objectives. The first is educational: acquiring the ability to construct large language models. The second is practical: applying that ability at the largest scale permitted by the hardware available to the learner. That formulation leaves room for differing levels of access rather than naming a minimum computing threshold.
Beyond those points, the supplied post offers no further details about timing, cost, datasets, safety practices or intended applications. Graham’s message is consequently best understood as a focused statement of personal technical priorities, not as a comprehensive guide to entering the field.
