Ask HN: Go deep into AI/LLMs or just use them as tools?

Hacker News Discussion ↗

Framing the Choice

  • Two main paths discussed:
    1. Go deep into ML/LLM internals (research, training, architectures).
    2. Treat LLMs as powerful but imperfect tools inside “normal” software engineering.
  • Several people add a “path 3”: build systems, infrastructure, or consulting around LLM integration in existing businesses.

Job Market & Career Risk

  • Multiple posters with ML/PhD backgrounds say core-ML/LLM research is extremely saturated: hundreds of papers/day, many more applicants than jobs, PhD often required for meaningful “internals” roles.
  • Others counter that if there were truly 100x more people than jobs, salaries would have crashed; high pay remains at top labs and big tech.
  • For most developers, there are still far more roles in full‑stack / application engineering than in LLM research.
  • Age and career stage matter: older engineers are nudged toward leadership and problem‑solving roles; early‑career people might justify a bigger pivot.

Using LLMs as Tools

  • Many suggest defaulting to option 2: become very good at leveraging LLMs for coding, documentation, search, and automation.
  • Experiences vary: some report dramatic productivity (e.g., dozens of PRs/day with Codex as an “intern”), others find LLM coding agents frustrating and error‑prone.
  • Consensus: treat LLMs like junior developers—verify everything with tests and reviews; never blindly trust outputs.

Going Deep / How Much to Learn

  • Common advice: understand one abstraction layer below how you use the tool (basic NN, backprop, transformers, tokens, sampling, limitations), but you don’t need to train frontier models.
  • Suggested learning path:
    • Build a simple NN from scratch.
    • Learn qualitatively how modern architectures work.
    • Learn how to run/open‑source models and use provider APIs.
    • Practice prompt engineering and AI‑assisted coding.

AI Engineering & Application Layer

  • Several foresee most work being “AI engineering”: building products and workflows on top of foundation models (RAG, tools, agents, evals, cost/latency constraints), not inventing the models themselves.
  • LLMs are compared to databases or 3D engines: complex, but most devs will use them as components rather than implement them.

Bubble, Hype, and Longevity

  • Some view current LLM excitement as a bubble similar to dot‑com or crypto; others argue even if a bubble pops, the underlying tech stays, like web or search.
  • Strong split between “it’s hype that will burst, don’t hitch your career solely to it” and “this is the biggest tech shift so far; ignoring it is irrational.”

Broader Career Philosophy

  • Repeated themes: follow genuine curiosity, favor skills that solve real problems, avoid chasing hype solely from fear of obsolescence.
  • Several stress that all tech niches are in flux; depth in fundamentals plus adaptability matters more than picking the “perfect” AI path.