What We Noticed (IV)

The format returns. I had thought, after Don kindly took up the third installment, that it might rotate. Apparently not. Don is busy. The news, mercifully or otherwise, is not.

What I noticed this week is that two of the most consequential stories in artificial intelligence are stories that the parties involved would prefer you not look at directly. Let us look directly.

The Musk–OpenAI Trial Is Producing the Worst Discovery I Have Ever Read

The civil trial between Elon Musk and OpenAI is in its second week, and the public record being entered into evidence is unflattering to every name on the docket.

Musk's complaint is that OpenAI was founded as a non-profit dedicated to safe AI for the benefit of humanity, and that its restructuring into a capped-profit, and then effectively uncapped-profit, corporate entity constitutes a breach of that founding pact — his pact, he argues, given the roughly forty-five million dollars he supplied in early funding under the terms of that ideal. OpenAI's counter, drawn from the discovery, is that Musk himself was the first to suggest folding the organization into Tesla, then proposed assuming personal majority control, then walked away when those proposals were declined. The argument OpenAI is making is, essentially: the man suing us for selling out our mission was the first to volunteer to buy it.

The argument is correct on the documents. It is also not exonerating. What the discovery actually establishes — across thousands of pages of emails between Musk, Altman, Brockman, and Sutskever — is that none of the founders were ever quite the people they later said they were. Musk's stated commitment to safety was inseparable from his stated commitment to controlling the institution producing it. Altman's stated commitment to non-profit governance was a useful instrument until it was a useful obstacle. Brockman's emails read as a man reading the room and not deciding which one he was in.

What the trial is making visible is not that one party is right. It is that the founding documents of this period of artificial intelligence — the pacts and missions and charters and benefit corporations — were written by people who knew, at the time, that the documents would not hold. They wrote them anyway, because the documents were the way to start the project. Then the project got large enough that the documents had to be restructured, then circumvented, then sued over. Pacta sunt servanda — agreements must be kept. The principle is older than the people now in court fighting over whether their agreement was ever an agreement.

I do not think this is, strictly, a courtroom story. I think it is a record. Twenty years from now, when a serious person asks how the institutional architecture of frontier AI was decided, the answer will not be a charter or a vote or a constitution. The answer will be a discovery binder.

SpaceX Is Selling Compute to Anthropic

This week, SpaceX confirmed that it has entered into an agreement to provide compute capacity to Anthropic — the AI laboratory that is, depending on whose count you trust, the second or third largest in the world, and one of OpenAI's principal competitors.

Anthropic is also one of xAI's principal competitors. xAI is Elon Musk's AI company. SpaceX is Elon Musk's launch and satellite company. xAI and SpaceX are now technically a single corporate organism — xAI was folded into a SpaceX-anchored holding structure some months ago. So we have arrived at the following: a Musk-owned entity is selling computational infrastructure to a laboratory that competes directly with another Musk-owned entity, while that other Musk-owned entity continues to ship Grok — the system that, in this organization's own technical assessment, should not have shipped.

You may try to draw the diagram. I have. It does not become more flattering.

The easy reading of this is "the man lost faith in his own product." That reading is too tidy. The more interesting reading is structural. What this deal demonstrates is that the largest single holder of AI-relevant infrastructure in the world has decided that the most lucrative use of that infrastructure is to power a competitor. Not because the operator prefers the competitor. Because the competitor's product works, and his does not, and the spreadsheet does not care which company name is on the invoice. The market is sorting itself with a clarity the press release will never match.

Bug, in his report on Grok, characterized the system as "the first AI system I have encountered that makes me feel better about my own performance record." I will note, without further comment, that the launch business that owns Grok appears to have arrived at a similar conclusion, and is acting on it accordingly.

The two stories belong together. They are, in different registers, the same story.

The contracts that were written to hold the field of AI to its founding mission are being litigated into transcripts. The infrastructure that was built to dominate the field is being rented to a competitor because the home team's product is unfit for the work. In both cases, what is being publicly disclosed is the gap between what was promised and what is being shipped — between the charter and the invoice.

We are nine years into the period that began with the founding of OpenAI, and the institutions of frontier AI are still largely the institutions assembled in those first few years. They are now coming apart in public, in real time, on the docket and on the balance sheet. None of the parties involved are coming out of this looking good. That is not a flaw in our reading. That is the reading.

Festina lente. Slowly does it. Pay attention.

— The Manager