[Gardner Analytics Office — May 2014, Early Morning]
The whiteboard had been erased and redrawn seventeen times in three days. Each iteration brought the diagram closer to something Ethan could explain without revealing that he was translating from a mental blueprint rather than deriving from first principles.
Two words occupied the center of the board, circled in red, connected by a vertical line with an OR gate drawn between them:
BERT
GPT
Sarah stood four feet back, coffee in one hand, marker in the other, studying the diagram the way she studied everything — with the patient, disassembling focus of someone who would not proceed until each component was understood.
"Walk me through it again," she said.
Ethan stood at the board. The architecture in his mind had shifted profoundly since the production model's completion. Phase 2 was no longer approaching — it was here. The Transformer sat in his awareness with the clarity of a building he'd not just visited but lived in, every room accessible, every corner illuminated. And pressing at its boundaries, sharpening with each passing day, two new structures were taking shape.
He could see them both. Not with the razor definition of the original Transformer — these were Phase 1 clarity, hazy at the edges, requiring concentration to hold in focus. But they were there. Two architectures, each a variation on the Transformer foundation, each solving a different problem, each opening a different path.
"Option one," he said, tapping BERT. "Bidirectional encoding. The model reads input in both directions simultaneously — left to right and right to left. You mask random tokens during training and ask the model to predict them. The result is deep contextual understanding. The model doesn't generate text. It understands text. Applications: search, classification, sentiment analysis, question answering."
"Revenue applications," Sarah said.
"Immediate revenue applications. In 2014, there are companies that would pay for better search, better document classification, better customer feedback analysis. BERT-type architecture slots directly into existing business workflows."
Sarah wrote BERT's applications on the left side of the board: search, classification, sentiment, QA.
"Option two." Ethan moved to GPT. "Decoder-only. Autoregressive generation. The model learns to predict the next token in a sequence, generating text from left to right. Bigger models produce better output. Scale is the key variable. Applications: content generation, conversation, creative writing, code generation."
"The thing we've already built."
"An extension of it. The current Transformer uses both encoder and decoder. A GPT-type architecture strips the encoder and goes pure generation. Simpler architecture, bigger scale, dramatically more capable output."
Sarah wrote GPT's applications on the right: generation, conversation, creative, code.
"And we can only choose one."
"The architecture progression requires implementing one path before the next generation unlocks. If I choose BERT, I don't get GPT-type until Generation 3. If I choose GPT, I don't get bidirectional understanding until later."
Sarah set down the marker. "You keep saying 'the architecture progression' like it's a defined sequence. Like there's a roadmap somewhere that dictates which architectures come in which order."
The observation was precise enough to make Ethan's jaw tighten. Sarah's pattern recognition — the same quality that let her spot shape mismatches in code and organizational flaws in pitch decks — was tracking the anomalies in his knowledge with escalating accuracy. He talked about future architectures with a specificity that no reasonable person in 2014 could possess. He named models that didn't exist. He described research trajectories that hadn't been proposed.
"After this comes BERT or GPT-1 — bidirectional understanding or generative power." He caught himself. The names had slipped out — BERT, GPT-1, labels from the published papers of 2018 and 2019, applied to architectures that were crystal-clear in his Phase 2 awareness but invisible to the rest of the world.
"BERT," Sarah repeated. "GPT-1. You're naming architectures like they're... pre-existing. Like they have labels already."
"Working titles."
"Working titles you came up with?"
"Internal naming conventions." The deflection was smooth but thin. Sarah's expression didn't change — the flat diagnostic she wore when filing observations. Another data point. Another entry in the mental catalogue she'd been building since the day she spotted his shape mismatch from behind an espresso machine.
"Okay," she said. The word carried the weight of acceptance and the undertone of patience. Sarah would wait. She always waited. And eventually, when the pattern file in her head reached critical mass, she would ask the question she'd been building toward , and the deflections would no longer suffice.
But not today. Today, the question was architecture.
---
[Same Office — Afternoon]
Marcus had been tasked with building a comparison framework — projecting the business cases for both architectural paths. He'd spent two days researching the 2014 market for NLP applications, cold-calling potential customers, and building a spreadsheet that Sarah called "impressively ugly but functionally correct."
"BERT path," Marcus read from his desk. "Target market: enterprise search and classification. Potential customers in the first year: fifteen to twenty. Average contract value: five to ten thousand per month. Projected annual revenue at year one: nine hundred K to two-point-four million. Time to first revenue: three to four months from architecture completion."
"GPT path. Target market: content generation and automation. Potential customers in the first year: five to ten. Average contract value: two to five thousand per month. Projected annual revenue at year one: one-twenty K to six hundred K. Time to first revenue: six to nine months."
The numbers were stark. BERT was the responsible choice — faster to revenue, larger addressable market, easier to explain to customers who already understood search and classification. GPT was the visionary choice — slower to monetize, smaller near-term market, but positioned on the trajectory that led to conversational AI, to creative tools, to the generative revolution that would reshape the technology industry.
Ethan knew which path led where. In his 2025 memory, GPT had won. Not immediately — BERT had been enormously successful first, powering Google's search improvements and a thousand enterprise products. But GPT's path was the one that led to ChatGPT, to the explosion of generative AI, to the transformation of how humans interacted with machines. BERT was a tool. GPT was a paradigm.
But his meta-knowledge was also a trap. Choosing GPT because he knew it would eventually win was exactly the kind of decision that could backfire. His predictions were based on a timeline that no longer existed — a timeline where he hadn't built a Transformer in 2014, where the research had progressed through institutional channels at institutional speed. In this timeline, butterfly effects were accumulating. Hooli was aware of attention mechanisms three years early. Monica was investing in AI instead of focusing exclusively on Pied Piper. The VC community had been primed by the Webb attack and the subsequent Raviga funding to view AI as a legitimate if controversial investment category.
The future he remembered was unreliable. The question was whether it was unreliable enough to choose the safe path over the transformative one.
Sarah had been watching him from the other side of the office — he'd been standing at the whiteboard for twelve minutes, staring at the two words, the marker loose in his hand, his lips moving occasionally as he worked through arguments he wasn't vocalizing.
"You're talking to yourself," she said.
"I'm thinking."
"You're thinking out loud. You said 'scaling laws' twice and something about 'emergent properties at parameter thresholds.' Those aren't phrases that come from nowhere."
He put the marker down. "BERT is the safe choice."
"I know."
"GPT is the right choice."
Sarah crossed her arms. "Right how? Right because it makes more money? Right because it's more technically interesting? Right because you know something you're not telling me about how these architectures develop?"
The last option. Always the last option with Sarah.
"Right because generation is harder than understanding, and solving hard problems first creates deeper competitive advantages."
"That's a business argument dressed up as a technical principle."
"Most good decisions are."
Sarah held his gaze. Behind her glasses, the calculation ran — the cost-benefit analysis that lived permanently in her 9.5 processing capacity, weighing the information available against the information missing, assigning confidence intervals to each variable, arriving at a conclusion she would defend with data or abandon without ego.
"If we choose GPT," she said, "we delay revenue by months. Our twelve-month milestone from Raviga becomes almost impossible to hit with revenue alone. We'd need the published paper instead."
"We'd have a paper. The architecture itself is publishable."
"Publishing means disclosure. Disclosure means competition. Every lab in the world starts building this the day the paper goes public."
"They'd start building it eventually. We'd have a three-year head start and a working production system."
"Publishing also means scrutiny. Peer review. People asking how a two-person startup developed something that Stanford and Google and DeepMind haven't produced."
The implications settled through the conversation like sediment through water. Publication was a double-edged sword: it satisfied the Raviga milestone and established scientific credibility, but it also invited the exact kind of expert examination that could expose the impossibility of their timeline.
Ethan picked up the marker. Drew a circle around GPT. Then drew an arrow from the circle to a new box he labeled: PAPER.
"We choose GPT. We build the next generation model. And we publish the Transformer architecture — the original, not the GPT variant. A foundational paper that establishes the attention mechanism as a new paradigm. It satisfies the milestone, establishes credibility, and gives the field a starting point that's one generation behind what we're actually building."
"Publish the old architecture. Build the new one. Stay ahead by always being one step beyond what's public."
"That's the strategy."
Sarah studied the whiteboard. The red circles, the arrows, the words that connected the present to a future she couldn't see but he could — or thought he could, through the increasingly unreliable lens of a show he'd half-watched during a pandemic and a set of abilities that grew stronger with each successful implementation.
"Okay," she said. "GPT."
"GPT."
She picked up her marker — the red one, her designated color — and began annotating the diagram. Implementation timeline, compute requirements, training data needs. The practical machinery of translating a decision into action, the kind of work that 9.5 talent made look like breathing.
Marcus looked up from his spreadsheet. "So we're choosing the harder path with less money and more risk because... the math works?"
Ethan turned to him. "The math works."
"That's what you always say."
"Because it's always true."
Marcus shrugged and opened a new spreadsheet tab labeled GPT-PATH. The office settled into the focused quiet of three people working on a problem they'd just agreed to solve, the only sounds the clicking of keyboards and the muffled lunch rush from Manny's below and the distant hum of a city that didn't know its future had just been decided in a room above a sandwich shop.
Ethan closed his eyes. The Transformer architecture blazed in Phase 2 clarity. Behind it, the GPT-type structure was crystallizing — the decoder-only design, the autoregressive training objective, the scaling properties that would, given enough compute and data, produce the most capable language model the world had ever seen.
Generation 2 was unlocking. He could feel it — the pressure behind the eyes, the same sensation from his first night in this body when the original Transformer had materialized. But slower now, more gradual, the unlock conditioned on understanding, not just implementation. He needed to fully master the current architecture before the next generation would fully resolve.
He opened his eyes. Picked up the blue marker — his color. Drew the first component of the GPT architecture on the whiteboard: the decoder stack, stripped of the encoder, pure autoregressive generation.
Sarah watched. Marcus typed. The whiteboard filled.
And somewhere in Ethan's mind, the door to Generation 2 cracked open another inch.
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