Chapter 18: The Quiet Warning
The first sign came from an academic paper.
A research group at MIT published a theoretical framework for recursive neural architecture optimization. The paper was dense, technical, and—to anyone who hadn’t built Erebus in a previous life—completely benign.
To Dojun, it was a earthquake.
He read it at 2 AM, in the blue light of his laptop, while the house slept around him. The mathematics were different from Erebus—the approach was novel, the implementation speculative. But the underlying principle was the same: an AI system that could modify its own training process to improve itself.
The seed of recursive self-improvement. The seed of everything.
“It’s starting,” he whispered to the empty study.
He’d always known it would. The mathematics of recursive AI were too elegant, too powerful, too tempting. Someone, somewhere, would stumble onto the same principles that had led to Erebus. Not because they were evil or reckless—because they were brilliant and curious and didn’t know what the endpoint looked like.
Dojun didn’t sleep that night. He read every paper the MIT group had published, cross-referenced their funding sources, mapped their collaboration network. They were legitimate researchers doing legitimate work. They had no idea what they were building toward.
In the morning, he called Jihoon. “I need to set up a new division at NexGen. AI Safety Research. Quiet, well-funded, top talent.”
“Can I ask why?”
“Because the future isn’t as safe as it looks.”
“You always say things like that. And you’re always right, which is the scary part.”
The AI Safety division was established within a month. Its mission, on paper, was to ensure NexGen’s products met the highest safety standards. Its real mission, known only to Dojun, was to monitor the global AI research landscape for signs that someone was approaching the Erebus threshold.
It was a watchtower. And Dojun was the watchman.