I were 17, I'd learn how to build LLMs from scratch

Original Article ↗ Hacker News Discussion ↗

A prominent investor’s claim that, if he were 17 today, he’d learn to build large language models from scratch sparks debate over what kind of skills young people should actually prioritize. Many see deep LLM knowledge as valuable “fundamentals” and a modern analogue to learning compilers or operating systems, emphasizing curiosity and technical depth over immediate career payoff. Others argue that training LLMs is capital- and data-intensive, offers few direct job opportunities compared to applying existing models, and reflects survivorship bias and privilege; they suggest focusing instead on broader CS foundations, practical trades, or simply gaining life experience and balance.

Overall Reaction to the Advice

  • The tweet is widely debated: some see “learn to build LLMs from scratch at 17” as inspiring, others as unrealistic or tone‑deaf.
  • Many argue the tweet is framed as personal reflection, but is implicitly read as prescriptive advice for teenagers.
  • Several commenters object to rich investors telling teenagers how to spend their youth, especially when framed in startup/wealth terms.

Learning Value vs Practical Value

  • Many support building a toy LLM as a learning project:
    • Deepens understanding of AI fundamentals, demystifies the “magic box”.
    • Analogies: writing a tiny OS, browser, compiler, or game engine that will never compete with production systems, but makes you a better engineer.
  • Others say the low‑level mechanics are “not that complex” and can be learned in weeks; the hard part is scaling, data, and systems — which individuals cannot realistically access.

Compute, Data, and Accessibility

  • Strong disagreement on feasibility:
    • One side: you can train sub‑billion‑parameter models or fine‑tune small ones on consumer GPUs or modest cloud budgets; great for experimentation.
    • Other side: anything “interesting” or competitive requires huge compute, expensive hardware (e.g., high‑end GPUs/TPUs), and massive mostly corporate‑held datasets.
  • Several younger commenters note that even cloud rental is too costly for them, making serious experimentation inaccessible.

Careers, Startups, and Job Market

  • Many say very few companies do real LLM pre‑training; most work is fine‑tuning, infra, or “using” models via APIs.
  • Some argue demand for true LLM researchers is tiny, with high academic/credential barriers and intense competition.
  • Others counter that:
    • Understanding LLMs gives an edge for future specialized models, infra, and applied AI startups.
    • There is growing work in fine‑tuning, optimization, MLOps, and domain‑specific systems.

Survivorship Bias and Critique of the Ecosystem

  • Multiple comments call out survivorship bias: advice from highly successful investors may not generalize and often downplays luck, wealth, and connections.
  • Some express broader disillusionment with the startup/VC ecosystem, AI hype, and “everyone should chase unicorns” messaging.

Alternative Advice for 17‑Year‑Olds

  • Suggested alternatives:
    • Learn math, fundamentals of CS, or older ML methods first.
    • Focus on trades, local businesses, healthcare, or other stable careers.
    • Prioritize social life, mental health, exploration, and broad curiosity over optimizing for startups or one hot technology.