It was past midnight when Dan Shiftman finally got to the proposals.
His home office was dark except for the glow of his monitor — a deliberate choice. The bright, open spaces of the Omnis campus were for performance. Here, in the quiet, he could actually think.
He'd asked for big swings. He'd gotten... caution.
The first twelve proposals were variations on themes the steering committee had already approved. Efficiency improvements. Benchmark optimisations. Safe bets dressed up in ambitious language. He could see the calculus behind each one: ask for enough to matter, not so much that you become a target.
He understood. The thought surfaced: somewhere along the way, this became a place where people were afraid to fail. He pushed it aside. That wasn't useful tonight.
He was about to close the queue when one more subject line caught his eye.
Proposal — Project Mnemosyne: Breaking the Memory Wall
He clicked.
The author was Aliah Green — one of the researchers they'd brought in after that blog post went semi-viral last year. He remembered the post. He remembered being annoyed by it, then intrigued, then sending the recruiting email himself. She'd called out the gap between what Omnis promised and what the architecture could actually deliver. Memory. Learning. Generalisation. The same words he'd put on a slide twelve hours ago.
He read.
By the second section, he'd stopped skimming.
Retrieval isn't remembering. Looking up notes isn't the same as having been there.
He knew this. Everyone at Omnis knew this, if they were honest. But she'd named it in a way that made the problem feel solvable. Not a vague complaint — a specific failure mode with a specific hypothesis for fixing it.
The model would learn to generate consolidation outputs — structured notes optimised not for human consumption, but for the model's own ability to resume a session.
He read the training approach. The verification signal. The phased compute ask.
It was a big ask. The steering committee would push back — Green was still junior, still proving herself. Marcus would point out the opportunity cost. Yuki would want more preliminary data. They'd be right to be cautious.
But Dan couldn't stop reading.
This is the first step toward a mind that could grow.
He sat back in his chair.
The thing about running Omnis was that everyone assumed he had a plan. That somewhere, in a vault or a whiteboard or a five-year roadmap, there was a path from where they were to where he'd promised they'd be. AGI. Scientific breakthroughs. AI to benefit all of humanity.
The truth was messier. The truth was that scaling had worked until it hadn't. The truth was that he'd said "we know the recipe" so many times he'd almost believed it. The truth was that he was starting to wonder if the recipe existed at all — or if they were all just waiting for someone to invent it.
And here was a researcher he barely knew, proposing to invent one piece of it.
He forwarded the proposal to his personal reading list. He'd look at it again in the morning, when he could think more clearly. When he could separate the late-night hope from the daylight reality.
But as he closed the laptop, one line kept circling in his head.
Every other approach to this problem is a workaround.
Maybe it was time to stop patching and start building.