What's this about?
Paul Graham's advice to a hypothetical 17-year-old — learn to build language models from first principles — became a 673-comment argument about whether that is good career advice or nostalgia.
What exactly does "from scratch" mean here?
Not training a competitive model, which needs a data centre. It means implementing the mechanics yourself at small scale — attention, tokenisation, the training loop — on a laptop, to understand how the thing actually works rather than only how to call it.
What's the community saying?
HN 592 points, 673 comments
- The most-replied comment lays out the dilemma: in theory there is enormous demand for real machine-learning engineers, in practice very few companies train anything, so the positions barely exist.
- A strong defence of the advice on non-career grounds — learn it not to build LLMs for a living but to be equipped for the next class of problems.
- The blunt dissent: horrible advice today, because the real optimisation work is proprietary research-adjacent work you cannot break into by understanding toy models.
- A survivorship-bias caution about taking career advice from very successful people at all.
- Surprise at how negative HN was, given that hacking on the newest technology in high school is a generational tradition here.