[Gardner Analytics Apartment — January 2014, 9:00 PM]
Sarah typed the way she made espresso — with an economy of motion that bordered on mechanical, each keystroke landing exactly where it needed to without hesitation or backtrack. She'd been coding for three hours straight, and in that time she'd refactored the entire positional encoding module, restructured the data pipeline to handle variable-length sequences, and identified four bugs that Ethan's accelerated coding sessions had introduced through sheer speed.
"Your batched_dot implementation assumes contiguous memory allocation," she said, not looking up from the screen. "Theano doesn't guarantee that. You need to explicitly call T.as_tensor_variable on the input before the operation, or you'll get silent shape errors at training time."
Ethan leaned over from the whiteboard where he'd been sketching the decoder architecture. "That should throw an exception, not a silent error."
"In Theano? Theano swallows errors like they're vitamins. I once spent three days tracking a gradient that computed to NaN because of an implicit broadcasting mismatch that the framework decided was intentional." She fixed the line, saved, ran the unit test. Green. "Your code is fast. Impressively fast. But it's got the fingerprints of someone who writes faster than they debug."
A fair criticism. The Accelerated ML Cognition let him produce code at a rate that outpaced his ability to verify it. Every session left behind a trail of subtle issues — not structural bugs, but efficiency gaps, edge cases, the kind of problems that only surfaced under load or with unusual input. Sarah was finding them at a rate that suggested her 9.5 rating underestimated her reviewing capability.
"The decoder," Ethan said, tapping the whiteboard. "I want your take before I start implementing."
Sarah pushed back from the desk and crossed to the whiteboard. The apartment had developed an ecosystem in the three hours since she'd arrived — her jacket on the back of the desk chair, her notebook open on the kitchen counter, a grocery bag she'd brought containing two sandwiches from a deli and a bag of dark chocolate almonds that she ate one at a time like medication.
She studied the decoder diagram. Six layers, mirroring the encoder, but with an additional cross-attention mechanism that let each decoder layer attend to the encoder's output. Masked self-attention in the decoder — preventing the model from looking at future tokens during training. A linear projection to vocabulary size, followed by softmax for next-token prediction.
"The cross-attention is where the power lives," Sarah said. "The encoder compresses the input. The decoder generates from that compression. But the masking has to be exact. If a single future token leaks into the attention weights during training, the model learns to cheat instead of predict."
"I know."
"Do you know because you've tested it, or because you just... know?"
The question hung between them. Sarah asked it without accusation — the same flat, diagnostic tone she used for code review. But the implication was there. She'd been watching him. Noticing things.
"I've worked with this architecture before," Ethan said. Which was true, in the way that a cook who'd eaten at a restaurant had "worked with" its kitchen. He'd used Transformer models daily in 2025. He'd fine-tuned them, deployed them, complained about their inference costs. He'd never built one from scratch — that knowledge came from the blueprint in his head, not from experience.
Sarah held his gaze for two seconds. Then she turned back to the whiteboard and picked up the marker.
"The masking implementation in Theano is going to be ugly," she said. "We'll need a custom Op for the triangular mask. Theano's built-in masking doesn't support the kind of dynamic sequencing you need for variable-length generation."
"Can you do it?"
"I already know how. Give me two hours."
---
[Same Apartment — 12:30 AM]
Ethan's hands were moving again. The Accelerated Cognition had kicked in forty minutes ago, triggered by the transition from whiteboard planning to actual implementation. The decoder's self-attention was taking shape on screen — lines of Theano appearing in bursts of five, ten, fifteen at a time, each one slotting into place with the precision of a jigsaw piece.
Sarah was at the kitchen counter, her own laptop open — she'd gone home to retrieve it during a break — working on the masking Op. She'd been typing steadily for ninety minutes. Occasionally she muttered to herself. Once she said "no, that's stupid" and deleted an entire function. Twice she looked up at Ethan's screen from across the room, watching his hands, her expression unreadable.
The third time she looked, she spoke.
"You're not using autocomplete."
A statement. Not a question.
"No."
"You're writing CUDA kernels at — I've been watching — roughly a hundred and twenty words per minute. Sustained. For forty minutes. With zero typos."
"I type fast."
"Nobody types this fast. Not writing original code. Copying, maybe. Transcribing, sure. But you're implementing novel Theano operations at a speed that implies you already know what you're writing before you write it."
Ethan's hands stopped. The cursor blinked at the end of a half-finished function.
He turned to look at her. Sarah's face was the same flat diagnostic she always wore — no anger, no fear, no accusation. Just observation. Data collection. The face of a scientist cataloging anomalous results.
"I get in the zone," he said. The deflection from the character bible — the prepared response for exactly this question. "Caffeine helps."
Sarah looked at the counter. They'd been drinking the coffee Ethan had bought — better than Folgers, worse than her café's beans. He'd had two cups. She'd had three.
"Caffeine doesn't explain this," she said. Then she turned back to her screen and kept coding.
She didn't push. She noted it, filed it, and moved on. Ethan recognized the behavior — it was the same pattern she'd used with the architecture questions, the cloud provider questions, the "how do you know things" questions. Sarah was building a case. Not with confrontation. With patience and observation. Collecting data points, storing them in the filing system behind those wire-frame glasses, waiting until the pattern resolved into something she could confront with evidence rather than suspicion.
A 9.5 didn't just mean technical brilliance. It meant a mind that tracked everything.
---
[Same Apartment — 3:15 AM]
The headache arrived like a tide. Not the sharp spike of the blueprint test — that had been overuse, too many hours of architectural examination. This was cumulative. Five hours of accelerated coding layered on top of two hours of whiteboard architecture review. The cognitive cost stacking, each hour's toll adding to the last.
Ethan pressed his palms against his temples. The code on screen — the decoder's cross-attention layer, three-quarters complete — blurred at the edges. Not his vision. The architecture in his mind. When he tried to reference the blueprint for the next implementation detail, the mental image wavered. Dimensions that had been crisp an hour ago were now soft. The attention heads looked the same but felt approximate, like remembering a phone number by rhythm rather than digits.
He reached for the aspirin bottle in the desk drawer. Shook out two. The bottle was lighter than it should have been — he'd been hitting it regularly since the first headache, and the supply was running low. He swallowed the pills dry, grimacing at the chalk taste.
Sarah looked up. She'd been watching again — that quiet, cataloging gaze.
"Headache?"
"Yeah."
"You've been coding for five hours."
"Closer to six."
"Take a break. The cross-attention can wait thirty minutes."
"The money can't." The words came out sharper than he'd intended. The headache was compressing his social filter. "Every day we're not training a full model is a day closer to zero."
Sarah set down her laptop. "How much do you have?"
"Eight thousand and change."
"And the cloud provider costs how much per training run?"
"The test model was thirty-four hundred. A full model will be ten times that. Minimum."
Sarah did the math silently. Her expression didn't change, but something shifted in the way she held her shoulders — a micro-tension, the physical tell of someone recalculating the risk of a decision they'd already committed to.
"So we need thirty to fifty thousand dollars before we can run a real training job."
"Closer to fifty. Allowing for failed runs."
"And the VCs think we're building Siri."
"Exactly like Siri."
She almost smiled. Not quite — Sarah's version of a smile was a slight softening around the eyes, a millimeter of lip-curve that disappeared if you looked directly at it. But it was there.
"Okay," she said. "Then we don't pitch VCs. Not yet. We need a demo that makes the Siri comparison impossible. That means we need a full model. That means we need money. That means—"
"We need a different kind of money."
"What kind?"
Ethan leaned back. The chair creaked. Outside, San Francisco was dark and quiet, the particular 3 AM stillness of a city that hadn't yet decided whether it was asleep or merely resting. Through the window, a single streetlight cast its cone of orange on the empty sidewalk.
"Grants," he said. "Research grants. DARPA. NSF. Google's been funding university AI labs. Facebook just hired Yann LeCun. There's money flowing into foundational AI research — not VC money, but institutional money. Money from organizations that won't ask us to compare what we're building to a phone assistant."
Sarah's eyebrows lifted. "You want to apply for government research grants? With a defunct startup and no academic credentials?"
"Gardner Analytics is a registered corporation. We can apply through the small business innovation research program — SBIR. The DoD funds early-stage technology through SBIR all the time. The proposals are technical, not pitchy. We'd be writing to an audience that knows what a neural network is."
"SBIR proposals take months to process."
"We don't need the money immediately. We need a credible path to money. Something we can show VCs as validation. 'We've been selected for a DARPA SBIR review' is the kind of sentence that makes skeptics listen."
Sarah picked up a dark chocolate almond from the bag on the counter. Chewed it slowly. "That's not terrible."
"High praise from you."
"Don't get used to it." She ate another almond. "I'll draft the technical section. You handle the business case and budget. We can have something submittable by end of week."
The headache was still there, pressing behind both eyes, but the conversation had shifted its weight from burden to background noise. A plan. Not a good plan — a desperate plan, full of assumptions and contingencies and the faint hope that a government bureaucracy would move fast enough to save a startup running on fumes. But a plan was better than whiskey at two in the afternoon and bartenders asking if he was building Siri.
Ethan stood. Walked to the kitchen. The instant coffee was terrible — Sarah had said so when she'd made it four hours ago, pronouncing the brand "an insult to the entire concept of roasting" — but it was hot, and at 3 AM, hot was the only metric that mattered. He filled two mugs. Handed one to Sarah. She took it without looking up, already typing.
They drank bad coffee in the blue light of two laptop screens while the apartment hummed with the particular energy of two people building something that didn't exist yet. Sarah's typing was steady. Ethan's was slower now, the accelerated cognition fading as fatigue accumulated, each line requiring conscious effort rather than automatic flow.
By 4:30, the decoder's self-attention was complete. The cross-attention layer was half-done. The masking Op that Sarah had built was elegant — cleaner than anything Ethan would have written, with edge-case handling he wouldn't have thought to include. Together, they'd produced roughly five hundred lines of reviewed, tested, functional code. The encoder was done. The decoder was taking shape. The training loop needed building, and the tokenizer was still a placeholder, but the core architecture — the beating heart of the Transformer, the attention mechanism that processed language the way no system in 2014 could — was real.
Ethan saved the latest version. Checked the line count. One thousand three hundred and forty-seven lines of Theano. Approximately sixty percent of a complete Transformer implementation.
The first light of dawn was turning the window from black to gray. Sarah had stopped typing. Her eyes were closed, her head tilted back against the desk chair, her breathing even. The laptop in her lap still displayed the masking Op, cursor blinking at the end of the last line she'd written.
Ethan pulled the North Face jacket from the closet — the one he wore as a blanket when the heating shut off — and draped it over her shoulders. She didn't wake. Her notebook had slipped from the counter to the floor, falling open to a page of systems diagrams. He picked it up. Closed it without reading. Set it on the counter beside her coffee mug.
The architecture in his mind was dim. The headache had subsided to a low hum. His hands were cramped from six hours of typing, the joints stiff and reluctant.
He sat back down at the desk and opened a new document.
SBIR Phase I Proposal: Novel Attention-Based Architecture for Natural Language Generation.
The title stared back at him. Through the window, San Francisco was beginning its morning routine — the first bus rumbling down the street, a delivery truck double-parking outside the diner across the road. The model's loss curve on ChronoCloud had settled at 2.1, the training complete, the test model waiting.
Somewhere on Sand Hill Road, Alan Rao was preparing for a day of meetings with founders whose products made sense in 2014. Somewhere in Palo Alto, Richard Hendricks was debugging Pied Piper's compression algorithm, riding the wave of his Disrupt victory. Somewhere in a Hooli conference room, Gavin Belson was planning a product called Nucleus that would try to crush what Richard had built.
And here, in a dead man's apartment in San Francisco, two people were building the future one line of code at a time, funded by an evaporating savings account and the stubborn conviction that the math worked.
Ethan started typing the proposal. Slower than his abilities allowed. Carefully. Each word chosen for an audience of government researchers who would read it in three months, if the money lasted that long.
The jacket shifted on Sarah's shoulders. She murmured something in her sleep — a fragment, inaudible, probably a variable name.
Ethan kept writing.
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