MIT Technology Review recently made a point that deserves more attention than it got: the idea of AI recursively improving itself, bootstrapping its way to superintelligence in a hurry, may not happen nearly as fast as its loudest promoters suggest. I read that and felt something close to relief. Not because I want AI to stall, but because it shifts the conversation to the part that actually touches your life. If the giant models aren’t about to fold themselves into an infinite loop of self-upgrades, then the real story of the next few years is happening at the other end of the scale. Small models. The ones that fit on the thing in your pocket.
What a “small” model actually means
When people talk about small language models, they usually mean AI systems trained with far fewer parameters than the headline names. Think of parameters as the internal dials a model adjusts while learning. More dials can mean more general knowledge and more range. Fewer dials means the model is cheaper to run, faster to respond, and possible to run on hardware you already own instead of a data centre you rent by the minute.
The trade-off used to be brutal. Small meant dumb. That gap has narrowed enough that the question is no longer “which model knows the most” but “which model is good enough at the specific job I need done.” Those are very different questions, and the second one is the one businesses and app developers keep asking.
Why your next phone is part of this story
AppleInsider has been tracking rumours about the 2027 iPhone 18, including a 2nm A20 chip and either 9GB or 12GB of RAM. I know, a memory spec is not thrilling reading. But RAM is the ceiling on what an AI model can do locally, without pinging a server. Every extra gigabyte is room for a slightly more capable model to live on the device itself.
That matters for anyone who has felt uneasy about where their data goes. An AI agent running locally can read your calendar, your notes, your messages, and never send any of it anywhere. It also works on a plane. It doesn’t slow down when a million other people are using it. And it doesn’t cost the company behind it a fraction of a cent every time you ask it something, which changes what they’re willing to give you for free.
The builders aren’t all where you’d expect
The Australian Financial Review reports that Australia is leading the rise of the small language model. That is a genuinely interesting detail, because it says something about who gets to participate. Training a frontier model is a project for a handful of extremely well-funded organisations. Training a small, focused model is a project a university lab, a mid-sized company, or a well-organised startup can take on.
Meanwhile, Rest of World has been covering Americans choosing Chinese AI products. Put those two stories side by side and you get a picture of a field that is spreading out rather than consolidating. Users are picking tools based on what works for them, not based on which national champion they’re supposed to root for. Developers are building in places that were never going to win a spending race. Smaller models are a big part of why that’s possible.
What this changes for AI agents
This site is about agents, so let me connect the dots. An agent is AI that takes actions rather than just answering questions. Booking the thing, filing the thing, checking the thing, then telling you it’s done.
Agents are chatty by nature. A single task might involve dozens of back-and-forth steps behind the scenes. With a large model, every one of those steps costs money and time. With a small model handling the routine steps, an agent becomes cheap enough to run constantly in the background and fast enough that it feels immediate.
The likely shape of this is a team rather than a solo act. A small local model handles the ordinary work: sorting, summarising, deciding what to do next. When something genuinely hard comes up, it calls out to a bigger model. You won’t see the handoff. You’ll just notice that your assistant stopped feeling laggy.
A useful comparison
Top Gear just published a rundown of 44 new electric cars on the way. Most readers will skim past the flagship hypercar and stop at the one that fits their driveway and their budget. AI is arriving at the same moment. The record-breaking models will keep making headlines, and they should. But the ones that end up in daily use will be the practical ones, sized for the job.
So if you’ve been waiting for AI to feel less like a novelty and more like a tool, watch the small end of the market. That’s where the useful stuff is getting built.
🕒 Published: