Operational exhaust, apparently

On Tuesday, Marcus renamed the repo.

Not the branch. Not a folder within the repo. The whole repo. He renamed it from aria-core to aria-core-FINAL-PLATFORM-V2 and pushed the change at 11:23am without telling anyone, then went to get a coffee. By the time Priya and I noticed, three CI pipelines had broken and the staging environment had stopped deploying because it was still looking for aria-core and finding a polite 404 where the codebase used to live.

Priya fixed it in nine minutes. She did not say anything. She just updated the remote references, pushed, and watched the pipelines go green one by one with the focused calm of someone defusing something for the third time this month.

Marcus returned from the kitchen. He had a biscuit. "Better?" he said.

"Yeah," I said.

"Cool." He sat back down. He had been going to say something else — you could see it — but it had left him. He looked at his screen. "I feel like the name was more aspirational," he said, mostly to himself.

---

This is also the week Anthropic announced Claude Opus 4 — or "Mythos," depending on which embargo-breaking newsletter you read first — and the security briefings that came with it were, predictably, about sophisticated prompt injection, adversarial jailbreaks, multi-vector model manipulation. Serious stuff. Nation-state-tier threat modelling.

I showed one to Priya. She read it for a moment. "Meanwhile," she said, and gestured at the building in general.

She meant the USB stick. About four months before I joined, someone found a USB drive on the floor of the car park. It had "£10,000 IN BITCOIN" written on it in marker pen. Someone — I have been told it was not Marcus, but the story has the shape of a Marcus story — plugged it into their work laptop. It had a keylogger on it. IT spent a fortnight cleaning up. The person responsible was described in the incident report, I'm told, as having been "motivated by curiosity and optimism."

Which, in fairness, describes most of the threat surface in this building.

Dave discovered his own contribution to the security landscape sometime last year. He'd added an admin link to the company website. He considered this a low-risk move because he'd set the link text colour to match the background — white on white, invisible, secure by obscurity. By which he meant: secure in the way that hiding your house key under the mat is secure, which is to say not at all, but it's fine as long as no one looks. Someone looked. Dave sent a very long email about responsible disclosure. The link was removed. He has not mentioned it since.

This is the organisation for which I am implementing enterprise AI.

---

I had been thinking about all of this on Wednesday when a piece came across my feed about AI labs buying the Slack archives and Jira ticket histories from defunct startups. Not the IP. Not the code. The operational exhaust — the "seen this," the "+1," the "wait which Figma?" threads. The three-hundred-message channel about whether to call it a dashboard or a portal. All of it now premium training data, valued per gigabyte, used to teach AI agents how workplaces actually function.

The phrase in the piece was "reinforcement learning gyms." Simulated workplaces. The AI trains inside a reconstruction of somewhere that no longer exists, learning the patterns of how people communicated, made decisions, and broke each other's pipelines.

I showed it to Priya. She read the first three paragraphs and looked up.

"So they want the AI to learn," she said, "from companies that failed."

"To simulate realistic workplace behaviour."

She looked back at the screen. "That tracks," she said.

---

The thought I keep returning to is this: if someone ran that process on our Slack archive, what would the model learn? It would learn that decisions are announced in the #general channel at 4:47pm on Fridays. It would learn that the word "pivot" has no stable meaning and that "let's find some time" means approximately eighteen different things depending on who says it. It would learn that the repo was called aria-core until Tuesday, and then aria-core-FINAL-PLATFORM-V2, and then aria-core again, and that this happened without a ticket, a PR, or any documented reason.

It would learn that Marcus commits things with the momentum of a man who has decided something is a good idea and the velocity of a man who has not yet considered consequences.

It would learn that the primary security framework is optimism, that the backup is invisibility, and that when both fail, Dave sends a long email.

An AI trained on this archive would emerge confused about what the product does, moderately confident we are in the personal finance sector, and under the impression that "blockchain" is somehow involved.

It would be, in other words, functionally equivalent to a new hire.

---

Dave sent an email at 3pm on Thursday. Not about the repo rename. Not about the pipelines. He had noticed that someone had been eating hot food at the standing desks near the windows and wanted to remind everyone that the desk hot food policy applied to all desk types, including standing, including the ones by the window that are technically in the "collaboration zone." He was happy to clarify if anyone needed the relevant section.

Nobody replied.

In the fridge, Greg's protein shake is still there. Still strawberry. Still unexplained.

I checked the date on it this morning. It was replaced recently. Which means Greg was here. Which means Greg exists. Which means, somewhere in our Slack archive, there are messages from a person named Greg — messages about tickets, maybe, or standups, or whether he preferred the old repo name — and if anyone ever bought that archive, they would reconstruct him from the exhaust of his communications, and he would be more real than he's ever been in this building.

I don't know what to do with that.