Why I Believe Small, Purpose-Built LLMs Beat One Giant Model
Everyone in AI right now seems obsessed with one question: How big can we make the next model? Bigger datasets, bigger parameter counts, bigger everything. But I think the more interesting path isn’t about size at all. It’s about specificity.
Most of the time, when people use AI, they aren’t asking it to solve the entire universe. They want one thing done well:
- Summarise this report.
- Help me debug a line of Swift.
- Draft an email.
- Turn a photo into a story.
Each of these is a single, clear goal. You don’t need a “universal brain” for that. You need a tool that’s tuned to the problem and doesn’t waste energy trying to be everything at once.
This is why Apple’s approach makes sense to me. Instead of chasing one giant model that does everything (and often poorly in certain areas), they are sprinkling smaller, specialised models across the system. One for language, one for images, one for context. Each is focused, efficient, and fits naturally into how you already use a device.
It reminds me of how apps evolved. On the web, we started with giant platforms trying to do everything. Over time, the best experiences came from apps that owned a category and solved it completely. Spotify isn’t trying to be your messaging app. Figma isn’t trying to be your ride-share. They are great because they are specific.
So why should AI be any different?
There’s also the trust factor. Smaller models can be more transparent, easier to run locally, and simpler to audit. That matters in sensitive areas like health, privacy, or education, where you don’t want a request going through a trillion-parameter black box just to get a straightforward answer.
Of course, there’s a tradeoff. A single giant model feels magical because it can switch tasks mid-conversation. But in practice, most use cases don’t need that kind of flexibility. They need reliability, speed, and clarity.
I think the future isn’t one mega-brain. It’s a network of smaller, specialised models, each doing one job extremely well, connected in a way that feels seamless. Like a great team.
And honestly, that feels much closer to how humans work too. None of us is the best at everything. We specialise, we collaborate, and together it works. Why wouldn’t AI evolve the same way?