Frontier models are the least interesting thing happening in AI right now. I know that reads as heresy on a site where half the inbound questions are some variation of “which giant model should I pay for,” but I’ve spent enough hours watching people route trivial tasks through enormous models to be confident about this: the scaling story has stopped being where the practical gains live. The interesting work has moved down-market, to models small enough to run somewhere you actually control.
Two signals landed recently that support this more than any benchmark chart. MIT Technology Review published a piece arguing that AI’s recursive self-improvement — the idea that models will start bootstrapping themselves into rapid gains — might not arrive as quickly as the industry’s loudest voices have promised. And the AFR reported that Australia is leading a rise in small language models. Those two stories are not usually filed together. They should be.
Why the self-improvement wobble matters
If you believed the aggressive version of recursive self-improvement, the rational strategy was to wait. Don’t over-invest in tooling, don’t build careful evaluation harnesses, don’t spend a quarter optimizing your prompt pipeline, because the next model would flatten all of it anyway. Patience was the play.
Take that assumption away and the calculus inverts. If capability gains arrive gradually rather than as a step function, then the difference between a working AI product and a demo comes from engineering: how you route requests, where you cache, what you validate, how you fail. That’s unglamorous work, and it’s the kind of work where a small model that you understand deeply beats a large model you’re renting through an API and hoping stays consistent.
I’ve reviewed enough agent frameworks to notice a pattern. The ones that hold up under real use tend to be the ones that treat the model as a component rather than a personality. They handle a narrow task, they get evaluated against something measurable, they get swapped out when a better option appears. Small models fit that shape naturally. Large models encourage you to hand over the whole problem and hope.
The Australia detail is not a footnote
The AFR framing — Australia leading the rise of the small language model — tells you something about who ends up building in this direction. It isn’t the organizations with the most GPUs. It’s the ones with constraints: bandwidth, budget, regulation, distance from the nearest hyperscaler data center. Constraints force decisions. Decisions produce architecture. And architecture is what’s been missing from most of the AI tooling I test.
Meanwhile, Rest of World has been covering Americans choosing Chinese AI, which is its own signal about how quickly assumptions around provider loyalty dissolve when something works well enough and costs less. Users are not sentimental about model providers. They’re sentimental about their own workflows. If a smaller, cheaper, closer option handles the job, that’s the option they’ll take, regardless of which flag is on the training run.
The hardware is quietly agreeing
AppleInsider’s 2027 iPhone 18 rumors mention 9GB or 12GB of RAM and a 2nm A20 chip. Rumors are rumors, and I’m not building a thesis on a leak two years out. But the direction of travel is consistent with what small-model advocates have been saying: the phone in your pocket is being designed with the assumption that meaningful inference happens on-device. Nobody specs that much memory into a handset for browser tabs.
The comparison I keep coming back to is Top Gear’s roundup of 44 new electric cars. Forty-four. Not one dominant model that made the rest obsolete, but a spread of options optimized for different budgets, ranges, and use cases. That’s what a maturing technology looks like. AI has been stuck in the phase where everyone argues about which single option is best, as if there were one correct answer for every task.
What this means if you build things
My practical read, and I’ll own it if I’m wrong:
- Stop routing everything to the biggest model available. Most tasks in a typical pipeline are classification, extraction, or formatting. Those do not need a frontier model.
- Build your evaluation setup now. If gains come gradually, measurement is your only reliable edge.
- Treat model choice as reversible. Anything that locks you to one provider’s quirks is a liability.
- Take on-device seriously as a target, not a novelty. The silicon is heading that way whether your stack is ready or not.
The frontier will keep moving, and the announcements will keep being loud. But the useful, boring, compounding work is happening in models small enough to fit in places the big ones can’t reach. That’s where I’d be spending attention.
🕒 Published: