Ox Alpha should never tell us who it is

Every few months a model appears on a marketplace under a codename, free, with no lab attached. The internet's reaction is always the same: work out whose it is. Prompt it for a system message, count its em dashes, ask it what its cutoff is, compare the tokenizer. Within a week somebody guesses right and the fun is over.

I think that instinct is exactly backwards, and the reason is the same reason Bitcoin has a founder nobody has ever met.

What the missing name actually bought

Satoshi Nakamoto's absence is usually filed under mystique. It is closer to load-bearing structure. There is no founder to subpoena, no keynote where the roadmap changes, no executive whose reputation the price is attached to, and no one to lean on when a government would like a word. The protocol had to be judged on whether it worked, because there was nobody available to ask.

Compare that to every other coin, all of which have a face, and all of which have found out that the face is the attack surface. It is where the regulators knock, it is what the price actually tracks, and it is what turns a technical question into a loyalty one.

Language models have exactly this problem and pretend they do not. Nobody evaluates a frontier model. They evaluate a lab's claim about a model, in public, with the lab watching, knowing that the lab has a press team, a developer relations budget and a long memory.

I published a bad number without thinking about it

This is not a hypothetical for me. Last week I spent 44,614,640 free tokens burying a fact inside fourteen novels and asking this model to find it again, 192 times.

Then I published it. That its accuracy does not improve as you shorten the context. That the worst place to put an instruction is the very top of the prompt, which is where everybody puts instructions. That it failed most often at 64,000 tokens, a length nobody thinks twice about.

I want to be honest about why that was easy. It was easy because I had no idea whose model I was writing about. There was no relationship to protect, no announcement to be gracious about, no chance of an email from someone whose job it is to explain that I had tested it wrong. I ran it, I wrote down what happened, and the absence of a name is the only reason that sentence is as boring as it sounds.

Multiply that by every developer currently poking at a stealth model with no press embargo and no relationship to lose, and you have something the industry almost never gets: a week of unbiased measurement. It is the closest thing to a double blind trial that model evaluation has, and it happens by accident, for about seven days, and then it stops.

Where the comparison falls apart

Here is the part the version of this post you were expecting would leave out.

Satoshi's anonymity works because you do not have to trust him. The code is public, the chain is public, and anyone who cares can verify the whole thing from genesis without knowing a single thing about who wrote it. Anonymity plus verifiability is a remarkable combination. Anonymity on its own is just a stranger asking for the benefit of the doubt.

A stealth model gives you the second half without the first. The weights are closed. You cannot audit what it was trained on, what it logs, or what it does with what you send it. The announcement said zero data retention; the marketplace's own model page says prompts are retained by the provider and merely not used for training. Both of those cannot be true, and there is nobody to ask, because that was the entire point.

So the honest version of the thesis is narrower than the headline. Stealth produces a healthier evaluation culture and a worse trust model. Bitcoin got to have both because it was open. An anonymous closed model gets to have one.

What the bitcoin of AI would look like

Not this. Not a codename on a marketplace with a countdown behind it. The bitcoin of AI is a model whose weights are on the table, so that not knowing who made it costs you nothing, because there is nothing you were relying on them for. Anonymous and unverifiable is a launch tactic. Anonymous and verifiable is a different industry.

Until somebody builds that, there is a smaller thing worth keeping, and it does not need anyone's permission. Evaluate every model as though it were stealth. Strip the name off before you test it. Write down what it did rather than what it was announced to do. Publish the bad column. That habit is free, it survives the reveal, and it is the only part of this week that was ever really about the model.

Ox Alpha will get a name. It will turn out to belong to a lab you have heard of, the guesses will be smug, and the number I published will start being read as a claim about that lab rather than a measurement of a model. Nothing about the model will have changed. Only what it costs to say something true about it.

It is routable here under the only name it should ever need, for as long as it stays free: oxalpha/stealth/ox-alpha in the playground or through the API.