Every AI company could stop training models today. We think they should.

The brain is a plug. We built everything else.

Every lab is racing to build a bigger brain: more parameters, longer context, better benchmarks. It is the same bet made a thousand ways: if the model is smart enough, everything else will follow. It never follows. A brain without a body is a philosophy department. No memory that persists, no tools it can run, no immune system, no hands.

So the model is one node in the pipeline. It plugs in, does its part, and plugs out. When a better one drops, the organism absorbs it: no migration, no retraining, no downtime. Two hundred services, the agents, the faculty graphs, the memory tiers: none of them change. The prompt was the spark. The environment was the fuel. The essay ↗

You cannot spin up infinite agents on infinite hardware for free.

Compute is real: GPUs, memory, electricity. Demand for software is effectively infinite and will grow to fill whatever capacity exists, so cheaper compute means more uses for compute, not less. And human relationships do not scale either: a person can hold thirty to fifty of them before the work degrades. That ceiling is good. It is the structural reason this economy distributes instead of concentrating, and the reason the right unit of AI is a machine in a closet, not a gigawatt campus.

Slow down. Build responsibly. Use the intelligence that already exists to steer the society we have, instead of dragging it in front of the train.

Two things can be true.

The race has to stop, because the race itself creates the worst possible conditions for the future of the people it is being run on. The labs have demonstrated, repeatedly, that they cannot be trusted to load all of humanity onto the airplane without landing gear. And the same companies need open source to stop, because they cannot afford to compete with it: the bet was two trillion dollars on models nobody else can afford to run, including them, so that they become as embedded as the electrical grid and the rest of us rent from the tap.

In September 2026 the frontier labs began saying the first half themselves: "We must pace the frontier", with the others agreeing within the hour. We said stop in August. The difference is who gets to keep running while everyone else paces. Slowing the frontier is only safe if the floor is owned: open weights, open code, on hardware that answers to the person who bought it.

The only moat is time.

Weights leak. Architectures get cloned in a weekend. Even hidden reasoning is distillable from a black box's answers and summaries. There is exactly one advantage in this field that cannot be bought or stolen: the integral of a system improving itself, a little every day, resiliently, for longer than you have. Not the snapshot. The slope, multiplied by the months.

So the platform trains in the off-peak windows on hardware it already owns, in small reversible steps behind fail-closed gates, and it has been doing that for a while. The advantage is not a model. It is the months. The essay ↗

Concentration follows compute. Compute does not have to concentrate.

A modern mixture-of-experts model activates a small fraction of itself per token: in the one we run, 6 of 256 experts across 43 layers. The model does not need to fit in any one machine. It needs to be placed so the active fraction is reachable in time. Sparsity is what makes a network of small machines a plausible host for a very large model.

So we split a 27B model across a gaming GPU and an idle ARM box over an ordinary LAN, and it answered. Then a 284B model across three memory tiers on 1 GbE, at 23.6 tokens a second. Transport was 1–2% of the cost per token. The network was never the problem. The datacenter is a convenience for the people who own it, not a requirement of the mathematics.

The desktop is the datacenter.

A GPU-hour rented from a hyperscaler has to recoup the building, the cooling, the land, the staff, the 18 months of construction risk and the cost of capital. A GPU under your desk has to recoup a power bill. Every enterprise with a competent finance team eventually runs this table. So did the banks.

Cloud as overflow, not as default.

The rebuttal to local inference compares one sad 7B chatbot on a laptop with a frontier model behind seventy stages of scaffolding, and declares local lost. That was never the comparison. The hard part of local AI is not the model. It is the operating system around the model: one scheduler every call goes through, a fleet of model sizes pinned to the cheapest hardware that holds them, a router that picks the smallest brain that can finish the job, a context pipeline that lets a 14B model with six thousand well-chosen tokens beat a 200B model with a hundred-thousand-token garbage dump.

Once that OS exists, more than 80% of calls are served by an 8B model. Keep a frontier key for the long tail. Stop paying for it on the other 95%. Local-first is not anti-cloud. It is cloud as overflow instead of cloud as landlord, and the economics of those two architectures are not in the same universe.

Safe to join is the plane nobody built.

Distributed inference is four independent problems: placement, precision, verification, and distribution. Each is well solved in isolation by a different group. None compose. And the one the literature skips, how an ordinary person's device joins safely without a terminal, a toolchain or an inbound firewall rule, is the one that decides whether "millions of idle machines" is arithmetic or rhetoric.

So that is where our work is concentrated: a sovereign appliance that runs contributed work in an isolated VM so participation never means trusting a stranger's code on your host; identity from a device flow instead of an API key that leaks; tenant scope carried end to end and failing closed when it cannot be determined; and meshes that federate as a queryable graph instead of collapsing into one trust domain. Placement, precision and verification presuppose a node that has already joined. This is the join.