Why Smaller, Smarter Models Will Win
28 October 2025
The AI industry is obsessed with scale. Bigger models, more parameters, more compute. But I think we're approaching an inflection point where smaller, purpose-built models start to outcompete the giants for most real-world applications.
What People Actually Need
Consider what most people actually need from AI:
- Summarizing documents
- Answering domain-specific questions
- Generating structured outputs
- Classifying content
These tasks don't require a model that can write poetry, debug code, and discuss philosophy all at once.
The Power of Focus
A 7B parameter model fine-tuned specifically for medical triage will outperform a 70B general model on that task—while running on a phone instead of a data center. The same principle applies across domains.
On-device AI is the future for most consumer applications. It's:
- Faster
- More private
- Works offline
- Costs nothing per inference
The constraint of running locally forces focus: you can't build a model that does everything, so you build one that does your specific thing brilliantly.
Apple Gets It
Apple understands this. Their approach with Core ML and on-device processing isn't a limitation—it's a competitive advantage. When your AI runs locally, you can offer features that cloud-dependent competitors simply can't match.
The winners in the next phase of AI won't be whoever builds the biggest model. They'll be whoever builds the most useful models for specific, well-defined problems.