The end of brute force: why the future of AI isn't bigger models

For about a decade, progress in AI has followed one rule that kept working: make it bigger. More parameters, more data, more compute. Each time the curve bent upward, and each time the field drew the same lesson from it.

I don't think that rule is wrong. I think it is running out of road, and that the next real gains will come from somewhere else.

What scale bought us

Scale gave us systems that write, translate, summarise and code at a level few people predicted. That is a real achievement and I am not interested in pretending otherwise.

Look closely at what scale buys, though, and a pattern shows up. Most of the gain is breadth: more facts, more styles, more of the patterns already present in the data. Much less of it is the ability to handle something genuinely new.

Give a large model a problem that resembles its training data and it is superb. Give it a small puzzle built on a rule it has never met, the kind a child solves after two or three examples, and the picture changes. Benchmarks like ARC were designed to measure exactly this, and for years they have been the place where size helped least.

The cost of scale

Every doubling now costs more than the last. Frontier training runs are planned like power stations, with their own energy contracts and their own supply chains, and serving the result to millions of people carries a running cost that never goes away.

That cost decides more than a budget. It decides who gets to build frontier AI, where it can run, and what it is allowed to be used for.

• A hospital that cannot send patient records off site.
• A robot working in a field with no signal.
• An instrument that needs an answer in milliseconds, not after a round trip to another continent.

None of these are well served by intelligence that only lives in a warehouse of GPUs.

Intelligence is not the same as size

The most capable general learner we know of runs on about twenty watts. A child does not need a billion examples of a door to learn how doors work. People infer a rule from very little, test it against what they see, drop it when it fails, and carry what works into the next problem.

That is much closer to what we mean by intelligence than recall at enormous scale. So the question I care about is not how large a model can get. It is how much capability you get per unit of energy, and how well a system copes with something it has never seen.

What we are betting on

At SHV Groups we build small on purpose. Eunoia Omega, our reasoning model, is designed to stay under half a billion parameters and to be judged on problems it has never encountered.

We are not going to describe its design here. It is early, and a claim like ours should be judged by results, not by diagrams.

What we will publish is the test. Before building anything, we wrote down what would count as success, what would count as failure, and the point at which we would abandon the approach. When the results arrive they get published, whichever way they go.

What this is not

This is not an argument that large models are useless or that scale has stopped mattering. Some problems will keep rewarding size for years.

It is an argument about where the next real step comes from. I think it comes from systems that learn more from less, need less to do more, and can run where people actually are. If we are wrong, the evaluation will show it, and we will say so here.