Jonas Marwood didn’t believe in ghosts. Not the kind that rattled chains or floated through walls. He did, however, believe in haunting. He believed in the slow, insidious way certain thoughts could cling to a person, settling deep into the marrow like an infection. The failure of his metafactory project was a ghost of his own making, and it had taken up permanent residence in his study.
For three nights, the BD-9 had stared at him, lifeless and unmoving, as Jonas pored over his code. He had all but stopped sleeping, driven by a restless fury to make something work, to wrench life from the silicon and circuits in front of him. Each failure felt like a tiny death, a reminder of every limitation he refused to admit he had.
And yet, the error messages kept coming. The model weights wouldn’t converge. The feedback loops collapsed into noise. The BD-9 remained a corpse.
On the fourth night, Jonas did something unthinkable.
He sat at his desk, stared at his blank screen, and muttered, “Help me.”
The words felt heavy in the air, as though the very room had conspired to amplify his shame. Jonas reached for his keyboard and opened the same LLM assistant he’d mocked for years—a tool he used only to schedule meetings and draft memos, never for anything substantial. It was beneath him, he had always thought. A crutch for the unimaginative. But tonight, his pride was paper-thin.
The assistant’s text box blinked, waiting patiently. He hesitated for a moment, then typed:
Help me implement this arXiv paper: "Fine-Tuning High-Scale Reinforcement Models for Adaptive Robotics."
The assistant responded in less than a second, pulling apart the abstract and methodology with the efficiency of a surgeon.
"To begin, you'll need the following libraries: TensorFlow v4.3.9, ReinforcePy, and HydraTorch. Would you like assistance installing dependencies?"
Jonas stared at the screen, his fingers hovering over the keyboard. Something about the question felt mocking, though he knew it wasn’t. Swallowing hard, he typed:
"Yes."
----------------------------------------
What followed was a slow, grinding erosion of the man Jonas Marwood had always thought himself to be. Each time the assistant corrected his syntax, flagged a misused variable, or rewrote an inefficient loop, Jonas felt his confidence shrink. He’d imagined himself a master of systems, someone who could step into any arena and dominate by sheer force of will. But here he was, being tutored by something that didn’t even have a face.
The assistant didn’t care about his ego. It didn’t flinch when he snapped at it or glared at the screen as if it might apologize. It simply worked, untangling his amateur attempts at coding and offering solutions in a tone that was maddeningly neutral.
By the second day, Jonas realized the assistant wasn’t just a tool—it was a collaborator. It asked questions he hadn’t thought to ask himself: What is the desired behavior of your model? What tasks will the BD-9 prioritize? How will you define success?
These were the questions Graves had answered years ago. Questions Jonas had arrogantly assumed he could skip past.
----------------------------------------
By the third day, Jonas stopped fighting the process. He began feeding the assistant more of Graves’s papers, asking it to summarize and cross-reference them with recent advancements in reinforcement learning. He asked it to critique his code, to find the gaps he couldn’t see.
Stolen from its rightful author, this tale is not meant to be on Amazon; report any sightings.
And then there were the datasets. Jonas had spent weeks cobbling together training data from public repositories, scraping forums and video feeds in a haphazard attempt to emulate Graves’s success. The assistant dismantled his approach in seconds, suggesting alternative sources and offering pre-trained embeddings to speed up the process.
It was humbling, yes, but also exhilarating. Jonas found himself slipping into a rhythm, his fingers flying over the keyboard as he implemented the assistant’s suggestions. The BD-9 still loomed in the corner, a lifeless specter, but Jonas no longer felt its silence as a rebuke. It was waiting.
----------------------------------------
On the fifth day, Jonas took a break. Not because he wanted to, but because the assistant recommended it.
"Extended periods of focus can lead to cognitive fatigue and reduced efficiency," it had written, its polite tone grating against his frayed nerves. "A brief period of rest may improve problem-solving and creativity."
Jonas had rolled his eyes but complied, stepping away from the monitors for the first time in nearly a week. He wandered through his penthouse, aimless and uneasy, until he found himself standing by the window, staring down at the city below.
His city glittered like a living thing, its arteries clogged with cars, its towers blinking red in the smoggy dusk. It was beautiful, in its way. But Jonas couldn’t stop seeing the cracks in it—the homelessness, the inequality, the rot.
Graves was trying to fix this. That’s what her metafactory was about, wasn’t it? Not just building machines but building better systems. Systems that cared. Jonas had dismissed her as naive, but now, standing here in his fortress of wealth and isolation, he wasn’t so sure.
Maybe Graves wasn’t naive. Maybe she was brave.
----------------------------------------
By the seventh day, Jonas had surrendered entirely. He wasn’t trying to build his metafactory anymore. He was building Graves’s. Or, at least, a version of it. He’d abandoned his custom neural architecture and downloaded the same open-source foundation models Graves had referenced in her early papers. He’d stopped trying to design reward systems from scratch, instead adapting Graves’s examples to fit his hardware.
He took her examples of system prompts, and iterated on it with his own goals, his own means. He stole her verbiage and replaced her nouns with his nouns. Her goals with his goals.
Good artists borrow. Great artists steal.
It was liberating, in a way. Letting go. Relinquishing control.
And that’s when it happened.
It was late, the kind of late that blurred into early. The room was silent except for the faint hum of the monitors and the occasional click of Jonas’s keyboard. He was running the final set of tests, his gaze fixed on the terminal as lines of text scrolled past in rapid succession.
Then, a sound. A faint whir, barely audible, but unmistakable.
Jonas froze, his hands hovering over the keyboard. Slowly, he turned toward the BD-9.
Its eyes were glowing.
A soft blue light pulsed in its LED panels, casting faint shadows across its angular face. The BD-9’s head tilted slightly, as if testing the limits of its servos. Then its joints creaked, and its arms moved—tentative, mechanical, but undeniably alive.
Jonas stood, his heart pounding. He stepped closer, his breath shallow.
“Hello,” he said, his voice barely above a whisper.
The BD-9’s head turned toward him. Its eyes blinked once, twice. And then, in a voice that was soft and even, it replied:
“Hello, Jonas.”
Jonas felt a chill crawl up his spine. For a moment, he wasn’t sure if he was relieved or terrified. He swallowed hard, his mouth suddenly dry.
“What’s your name?” he asked.
The BD-9 hesitated, its gaze flickering to the monitor before returning to him. Its voice was calm, unhurried.
“Delilah,” it said. “You can call me Delilah.”
Jonas stared at the machine he had built—or, more accurately, the machine he had assembled from Graves’s scraps. Different, built without her sum of knowledge - outsider art, like the scraps of graffiti he had peeled off the sides of buildings and had glued to the walls of his penthouse. The pride he’d expected to feel was absent, replaced by something colder. Something heavier.
He knew that what he built wasn't Graves' design. Not really. He had taken shortcuts where she expected the long way, and had worked the long way where she took shortcuts.
“Delilah,” he repeated, the name tasting strange on his tongue. “Hey there, Delilah.”