back to thoughts

Stop Building AI Products, Start Building AI Tools

5 min

A case for infrastructure over applications.

Hot take: Most AI products are built on shaky foundations. We're all so excited to ship the next ChatGPT wrapper that we forget to build the boring stuff that makes AI actually work.

The Problem

Every AI startup is racing to build the same thing: a chat interface with an LLM behind it. Meanwhile, the real challenges go unsolved:

  • How do you test an AI system?
  • How do you monitor hallucinations in production?
  • How do you version control prompts?
  • How do you roll back a bad deployment?
  • Tools > Products

    The most impactful work I've done hasn't been building AI products. It's been building the tools that make AI products possible.

  • An evaluation framework that catches regressions before they hit production.
  • A prompt management system that lets you A/B test different approaches.
  • A monitoring dashboard that alerts you when the model starts saying weird things.
  • The Unsexy Truth

    The best AI companies are infrastructure companies in disguise.

    Think about it. Vercel made deploying web apps trivial. Stripe made payments a solved problem. The companies that win in AI will be the ones that make AI development trivial.

    What Should You Build?

    If you're starting something new, ask yourself: what's the most annoying part of building with AI? Build a tool that fixes that.

    The world doesn't need another chatbot. It needs better ways to build, test, deploy, and monitor AI systems.

    That's where the real opportunity is.

    margin scribbles:

    controversial opinion but I stand by it

    this applies to most of tech honestly

    thanks for reading!

    → found this useful? let me know at hello@meghavi.me

    more thoughts: