What Intelligence Should Cost
Most people buy intelligence the same way they buy a car. Bigger, and hope the next one is cheaper. We run a small lab that asks what intelligence actually costs, and whether you can buy the same capability for less.
Opinions are not my employer's.
If you build with AI, you pay for it. The bill comes as a subscription, or a datacenter, or a GPU running in the corner. For a long time the only way to bring the cost down was to build the model bigger … and hope the next hardware generation caught up. That bet has worked for a while now. That is not the only bet on the table.
What we are actually asking
The big labs are asking how to make it better. We spend our time on a different set of questions. How much attention does a model actually need … and at what size do a few attention layers stop earning their compute? Where does attention beat a simpler recurrent core, and where does it not? What is the floor of a corpus, and how far above that floor is the model really sitting? What is the cheapest part of a model that rescues one that would otherwise fail to train?
Those are not questions you answer in a keynote. They are questions you answer by building a lot of small things, measuring them honestly, and writing down what happened … including when nothing happened.
How the work is done
Most small lab research is a set of claims with a benchmark bolted on. We do not work that way. Every idea runs against the same fixed protocol. Same data. Same number of training steps. Same compute budget. Only the architecture is allowed to change. When you hold everything else still, you can tell a real difference from noise.
Every run goes into a journal. Not a highlight reel. A journal. If an idea does not work, it is in the record, and it changes what we try next. We write the hypothesis before we run. We do not tune a model after we have seen its score. Those rules are boring. They are also the reason the results are worth trusting.
We also work small, on purpose. The big labs get to test their ideas at billions of parameters. We test ours at a fraction of that. That is a real constraint, and we would be lying if we said it is not. But it is also the point. If an idea cannot be shown to matter at a scale you can afford, you cannot afford to scale it up and find out it did not. The cheap version is where the bet gets made or broken.
Why it is worth doing
This lab exists for the same reason you would run a model on your own hardware. A model someone else hosts is a meter they own … a capability bought through someone else’s architecture is a price they can move. Every time we show a capability costs less to build than the default, we hand someone holding their own tools a little more room to negotiate … or to walk away from the meter entirely.
We are not trying to replace the frontier models. We are trying to find the parts of the stack that are overpriced, and to prove it with numbers instead of opinions. The frontier will keep moving. The question is what it should cost, and who gets to decide.
What we are after
We are a small lab in Reno doing the unglamorous version of AI research. Build it small. Measure it honestly. Keep the receipts. Let the evidence say what is cheap and what is not. If you have ever looked at a model bill and thought there should be a cheaper way to get this … that is the whole lab in one sentence.
The findings are coming. Some will be boring. A few should be interesting. We will write them down when we are ready, and we will write the failures too.