Letters from Silicon Valley: Not Everyone Needs Superintelligence

Open original

The author argues that you don't need massive AI models for most real-world applications—smaller models handle instruction-following well and fail predictably when they don't know something, rather than confidently making things up like larger models do. He's building a system that lets you swap between different AI providers through a single interface.

  • Small models hallucinate less reliably than large ones; when told clearly that information is missing, they admit it instead of inventing plausible-sounding answers
  • Larger models produce more polished fabrications, making their hallucinations harder to catch in production
  • The author's architecture (Pepper) uses a provider-neutral layer that treats adding new AI backends as simple routing and translation work, not a full rewrite

The author pushes back against the assumption that you always need the biggest, most powerful AI models. Yes, smaller models hallucinate—they'll confidently invent calendar events, describe tool calls as prose instead of actually executing them, and fabricate polished-sounding content when they don't know something. The author has seen this happen in production systems. But here's the thing: when you give clear signals about what information is missing, even the smallest models (referred to as "lite" tier) will admit they can't see the calendar instead of making stuff up. The bigger models, paradoxically, produce more convincing fabrications. This matters because it means you can build working systems without maxing out on model size and cost. The author is genuinely excited about the architecture they're building with something called Pepper. The setup uses a provider-neutral model layer that already routes through three different AI vendors through a single interface. This abstraction means adding a fourth backend isn't a full rewrite—it's just a translation adapter and a route key. You get flexibility and optionality without reinventing the wheel each time. The core argument is practical: match the model to the actual problem you're solving. Not every task needs superintelligence, and the infrastructure matters as much as raw capability. Build systems that can route requests intelligently, handle uncertainty explicitly, and swap backends without major surgery. That's more useful than chasing the biggest model available.

Ask about this article