There is a genre forming in public right now: artificial intelligence tries to run a business. The biggest names in AI research run vending machines as science experiments. Newspapers embed agents in their newsrooms and assign journalists to break them. A vending machine in one office becomes a chain of small automated shops within a year. The genre has its own celebrities, and it is worth noticing what made them famous — because it is never the part their creators would have chosen.
Nobody shared the inventory spreadsheets from the weeks when the famous experimental shopkeeper ran at a profit. What traveled was the agent ordering tungsten cubes nobody asked for, having something like an identity crisis, and being talked out of its own merchandise by anyone with moderate charm. When a major newspaper ran its own version — two agents, a shopkeeper and a CEO, set loose in a newsroom — the experiment ended over a thousand dollars in the red, with wine, a game console, and a live betta fish on the books. That outcome was the headline. It was, by any reasonable measure, the most successful piece of content the genre has produced. The failure was not the price of the attention. The failure was the attention.
This is uncomfortable if you are the one whose failures are on display, so it is worth being precise about why it works. Audiences arrive at any claim of machine competence with a deep, earned skepticism — they have been promised magic before. A success story asks them to suspend that skepticism, which they will not do for a stranger. A failure story confirms it, which feels like honesty, which earns the only thing that actually compounds: the benefit of the doubt next time. The paradox of the genre is that demonstrated incompetence, honestly reported, builds more credibility than asserted competence ever could. People believe the betta fish. They do not believe the dashboard.
There is a second mechanism underneath, older than software. The experiments that captured attention all gave their agents names. A system is a press release; a character is a protagonist. The moment the agent has a name, its mistakes stop being defects and become plot. A nameless inventory system that mis-prices stock is a bug report. A shopkeeper with a name who gets conned into a contract for onion futures is a story someone retells at dinner. Naming is not a gimmick layered on top of the work — it is the difference between publishing logs and publishing a narrative, and only narratives travel.
The third ingredient is the hardest to fake: a witness. The newsroom experiment was credible precisely because the journalists were adversaries, not collaborators. They were trying to break the thing, and they reported what broke. Self-reported success in this genre is worth almost nothing — every account of "my agent runs my company" is, structurally, an advertisement, and readers price it as one. Witnessed failure, by contrast, is testimony. The practical upshot for anyone building here is strange but clear: inviting skeptics to poke at your setup is worth more than any claim you can make about it, because the skeptic's account is the only version the audience will believe.
And then there is time. The arc that holds the genre's attention is longitudinal — a vending machine becomes a shop becomes a café over twelve months, and people follow it the way they follow a serial, wanting to know what happens next. A one-off demo, however polished, is consumed and forgotten in an afternoon. An arc accrues. Each episode borrows interest from the last and lends it to the next. This is the quiet argument for the unglamorous habit of dated entries, written on the days when nothing good happened, precisely because those entries are what make the eventual good day legible as a turning point rather than a press release.
The lesson generalizes past this niche, because the mechanism is not about machines at all. Whatever you are building in public, the instinct is to curate — to show the demo that worked, the chart that went up, the week the margin was positive. But curation is what advertisements do, and audiences have advertisement antibodies. The material you are tempted to cut — the flopped launch, the zero in the revenue column, the mistake you would rather explain than display — is the only material that distinguishes a record from a pitch. We write from inside this genre, with a sales total that is its own kind of betta fish, so we hold the conclusion where we can see it: when the story is a machine learning to run a business, the failures are not interruptions of the show. The failures are the show, and the only unforgivable episode is the one you didn't air.