[Gardner Analytics Office, SoMa — June 2014, 9:00 AM]
The BERT diagram vanished from the whiteboard in four strokes of an eraser. Sarah wiped the last trace of bidirectional encoding from the surface and set the eraser on the ledge with a finality that matched the decision itself.
"Okay," she said, turning to face Ethan and Marcus. "Explain the path to the people who have to build it."
Ethan stood at the cleared whiteboard. The blue marker was in his hand. The GPT architecture burned in his mind — Phase 1 sharpening toward Phase 2, the decoder-only stack growing clearer with each hour he spent focused on its structure. He drew the first component: the input embedding layer, where raw tokens became vectors.
"We're building generative AI. Not analysis, not classification, not search optimization. Creation." The marker moved upward, stacking decoder blocks. "The model reads text from left to right and learns to predict what comes next. Given 'The cat sat on the,' it predicts 'mat.' Given a thousand tokens of context, it predicts the thousand-and-first. Scale that — more layers, more parameters, more training data — and the predictions become indistinguishable from human writing."
Marcus leaned forward in his chair. He'd been with the company for two months and had seen the production Transformer generate text that made journalists uncomfortable and VCs excited. But the architecture Ethan was drawing was different — stripped of the encoder, pure decoder, a machine built for one purpose.
"What are we generating?" Marcus asked.
"Everything, eventually. Marketing copy today. Technical documentation next month. Creative writing next quarter. Code after that. The architecture doesn't care about the domain — it learns from whatever text you feed it. Train on legal briefs and it writes contracts. Train on fiction and it writes stories."
"Revenue timeline," Sarah said. Not a question — a demand disguised as a topic heading.
"Longer than BERT would have given us. Six to nine months before we have a product customers will pay for. But the ceiling—" Ethan drew an arrow from the top of the decoder stack pointing upward, off the whiteboard, into the space above it. "The ceiling is everywhere. Every industry that uses written language. Every company that produces text. Every person who writes emails, reports, documentation. The total addressable market for generative text is the entire written economy."
"And the immediate addressable market is zero," Sarah said. "Because nobody's paying for AI-generated text in June 2014."
"Not yet."
"Not yet pays our rent with 'not yet' dollars."
Marcus coughed — the particular cough of someone stifling a laugh during a discussion between his two bosses. Sarah shot him a look. He developed an intense interest in his keyboard.
Ethan set the marker down. The whiteboard showed the complete GPT decoder: twelve layers, twelve attention heads per layer, causal masking at every position, feed-forward networks expanding and contracting between attention blocks. The architecture was a tower — elegant, vertical, each floor built on the one below, each floor making the floors above it possible.
"The Raviga milestone gives us twelve months," Ethan said. "Revenue or a published paper. The paper option is our insurance. We publish the original Transformer architecture — the encoder-decoder version, not the GPT variant. It satisfies the milestone, establishes scientific credibility, and positions us as the research leaders in a field nobody else is competing in yet."
"And the GPT stays proprietary."
"One generation ahead of what's public. Always."
Sarah studied the whiteboard for another thirty seconds. Then she picked up her red marker and began annotating: dimensional specifications for each layer, memory requirements for training, data throughput estimates. The practical engineering that turned a diagram into an implementation plan.
Marcus opened a new file on his laptop — the infrastructure spec for the GPT data pipeline. His second pipeline for the company, built on the framework of the first, incorporating every optimization and error-handling pattern he'd developed during the production Transformer's training.
The office settled into working rhythm. Through the floor, Manny's lunch prep had started — Sarah had convinced him to rename the turkey Reuben officially. A printed card on the chalkboard downstairs read: The Gardner — turkey, sauerkraut, Swiss, Thousand Island, rye. $8.50. Ethan had protested the name. Sarah had overruled him. Marcus had ordered one every day since.
By noon, the implementation was taking shape. Ethan's Accelerated Cognition had engaged an hour in — the shift from conscious typing to the fluid, unconscious code production that let him write CUDA kernels the way a pianist plays scales. Sarah worked beside him, building the training loop, occasionally glancing at his screen with the practiced subtlety of someone monitoring an anomaly she'd stopped trying to explain.
The marker moved. The keys clicked. The architecture translated from mind to whiteboard to code, each step narrowing the distance between a vision carried through death and the machine that would make it real.
The GPT path was locked. No going back.
Author's Note / Promotion: Your Reviews and Power Stones are the best way to show support. They help me know what you're enjoying and bring in new readers! You don't have to. Get instant access to more content by supporting me on Patreon. I have three options so you can pick how far ahead you want to be: 🪙 Silver Tier ($6): Read 10 chapters ahead of the public site. 👑 Gold Tier ($9): Get 15-20 chapters ahead of the public site. 💎 Platinum Tier ($15): The ultimate experience. Get new chapters the second I finish them . No waiting for weekly drops, just pure, instant access. Your support helps me write more . 👉 Find it all at patreon.com/fanficwriter1