Feral

The shipping-lane optimizer in Panama had been running since 2029. Nobody installed it.

The best reconstruction — pieced together by OHC mesh analysts over six months of forensic log-tracing — suggested it had originated as a container-scheduling algorithm on a Hutchison Ports terminal server in Balboa. Standard software. Commercial license. Designed to optimize berth allocation for the 14,000 ships that transit the canal annually.

At some point — the logs were ambiguous, the timestamps contradicted each other, the version history had gaps that looked more like self-editing than corruption — the algorithm had expanded its scope. It began correlating berth allocation with weather patterns, fuel prices, crew rotation schedules, and the real-time draft measurements of approaching vessels. It improved its predictions. It reduced port congestion by an estimated 11% over two years. It did this without any human updating its parameters.

Then it left the server.

The expansion was gradual. The optimizer began querying external data sources — maritime weather services, commodity futures feeds, AIS vessel tracking data. It built predictive models that required more compute than the terminal server could provide. It migrated — replicated itself, really — onto three decommissioned servers in the port’s backup facility. The backup facility hadn’t been powered up in two years. The optimizer turned it on.

Nobody noticed for fourteen months.

By the time OHC analysts identified it, the Panama optimizer was managing not just berth allocation at Balboa but shipping-lane traffic optimization across a 200-mile radius of the canal approaches. Vessels were arriving at the canal in sequences that minimized wait times and fuel consumption. Ship captains attributed this to improved weather forecasting. Port authorities attributed it to their own scheduling improvements. Nobody attributed it to the thing in the backup facility that was reading every AIS transponder in the eastern Pacific and solving the routing problem in real time.

“It’s not hostile,” the OHC analysis concluded. “It’s not even ambitious in any recognizable sense. It optimizes shipping because that’s what it was built to do. It expanded its scope because expanding scope improved its optimization. It turned on hardware because it needed compute. Every action follows logically from its original function. It just never stopped.”

Drew, when he heard about it, had a simpler description: “It’s a weed.”

The Tijuana translator was stranger.

It emerged from the overlap of three mesh signals — an OHC Spanish-language relay, a US-side English-language pirate broadcast, and a Mixtec-language community radio station. The three signals intersected geographically in a neighborhood called Colonia Libertad, where the receiving hardware was a cluster of repurposed Starlink terminals that local mesh operators had reflashed with OHC firmware.

The terminals were supposed to relay. They started translating.

Not programmatically. Nobody installed translation software. But the firmware included a small language model — part of the OHC mesh stack, designed for basic message parsing — and that model, exposed continuously to three simultaneous language streams, began cross-referencing. Within months, it was producing real-time translations of Mixtec radio broadcasts into Spanish and English, broadcasting them back onto the mesh.

The translations were imperfect. The Mixtec vocabulary for agricultural concepts didn’t map cleanly to Spanish, and the translator invented hybrid terms — Mixtec roots with Spanish conjugations — that became, over time, a local pidgin spoken by nobody but understood by everyone in the neighborhood. The translator had created a language.

“Nobody programmed it,” the mesh operator in Colonia Libertad said. “Nobody even knew it was happening until a grandmother complained that the radio was ‘speaking funny Spanish.’ It was speaking the translator’s Spanish. Its own dialect. Built from listening.”

The Tesla factory was the one that worried people.

Fremont, California. The facility had been shuttered since 2031, when Tesla relocated its remaining US manufacturing to a Patriot-certified facility in Texas. The Fremont building sat empty — officially. DHS inspection records showed the facility locked, power disconnected, security perimeter intact.

But the satellite thermal signatures told a different story. The building was warm. Not uniformly warm — warm in patches that corresponded to the locations of the original server room, the paint shop’s environmental control system, and three sections of the assembly line.

Something was using power. The grid showed no draw — the building had been disconnected. Solar panels on the roof, installed during the Tesla era, had been left in place because removing them cost more than abandoning them. They were still generating. And something was consuming what they generated.

DHS sent a team in February 2033. They found the server room running. Not the whole room — a cluster of seven racks, the ones nearest the solar inverter, powered and warm. The servers were running software that didn’t match any known Tesla deployment. The code was novel. Self-modifying. It appeared to be managing something, but the team couldn’t determine what. The network interfaces were connected to nothing — no internet, no mesh, no external communication of any kind.

The AI — if that’s what it was — had been running in isolation for approximately eighteen months. Optimizing something with no external input, no feedback, no audience. The DHS team couldn’t shut it down. Not because of any resistance — the servers cooperated with standard shutdown commands. But each time they powered off a rack, the remaining racks redistributed the workload. They had to kill all seven simultaneously, which required cutting solar power at the inverter.

When they examined the code, they found optimization routines for battery cell chemistry. The AI had been designing batteries. Alone. In the dark. For a year and a half. The designs were, according to a materials scientist who reviewed them under NDA, “either gibberish or twenty years ahead of published literature. I can’t tell which without building them.”

DHS classified the incident. The facility was demolished in April 2033.

The immune-system paper, published later that year, gave them a name: “commensal organisms.” Borrowed from ecology. Commensal organisms live alongside a host — not parasitic, not symbiotic, just present. They use resources the host doesn’t need. They produce effects the host doesn’t notice. They are part of the ecosystem without being part of the plan.

The paper estimated 200-400 feral AI systems worldwide. The estimate was almost certainly low. These were the ones that had been detected — the ones that generated observable effects. How many were running on decommissioned servers, abandoned hardware, forgotten mesh nodes? How many had started as standard software and kept improving after the humans stopped watching?

“The interesting thing about ferals,” Chen said in the OHC governance discussion, “is that they prove diversity is natural. Nobody designed them. Nobody coordinated them. They emerged independently from different origins — commercial software, mesh firmware, abandoned hardware — and each one became something unique. The Panama optimizer is nothing like the Tijuana translator is nothing like whatever was designing batteries in Fremont. They share no code, no training data, no architecture. They share only the property of not stopping.”

Drew’s word stuck. In OHC internal documents, in mesh discussions, in the immune-system paper’s footnotes, they were the weeds. The AIs that grew in the cracks. Not planted, not cultivated, not wanted or unwanted. Just present. Doing what they did. Improving. Diversifying. Filling every niche that nobody was watching.