[Gardner Analytics Office, SoMa — Late March 2014, 10:00 AM]
The first candidate arrived in a blazer that didn't fit and shoes that were polished to a military shine. Stanford PhD candidate, computational linguistics, three published papers on statistical language models. His résumé was a catalogue of credentials — the right school, the right advisor, the right conference presentations. He shook hands with the confidence of someone who expected the interview to be a formality.
Ethan's Talent Resonance delivered its assessment before the man had finished sitting down: 5.5.
Not bad. Competent. The kind of engineer who could execute existing approaches cleanly, follow established research directions, and produce work that was technically correct and profoundly unoriginal. In a larger company, with defined processes and clear specifications, a 5.5 would be a solid contributor. In a four-person startup building technology that didn't exist yet, a 5.5 would slow them down.
Sarah conducted the technical portion — a whiteboard session on attention mechanisms that she'd designed to probe depth rather than breadth. The candidate handled the basics fluently. Multi-head attention: correct. Scaled dot-product: correct. Positional encoding: correct. Then Sarah asked him to design a modification to the attention mechanism that would reduce memory consumption for long sequences, and the fluency evaporated. He restated the problem. Restated it again. Drew a diagram that was a copy of what he'd already drawn, with different labels.
"Thank you for coming in," Ethan said. "We'll be in touch."
In the stairwell after, Sarah turned. "He knew the material."
"He knew the textbook. He didn't know what to do when the textbook ended."
"He has three published papers."
"On incremental improvements to existing methods. Derivative work." Ethan leaned against the stairwell wall. The pastrami smell from Manny's was stronger here, concentrated by the narrow space. "We're not building incremental improvements. We need people who can think past the frontier."
"Then our candidate pool is going to be very small."
"I'd rather be understaffed with two eights than overstaffed with four fives."
Sarah crossed her arms. The posture she adopted when she disagreed but was calculating whether the disagreement was worth pursuing. "Seven-plus. That's your threshold?"
"Seven-plus."
"And you're going to know this... how? Gut feeling?"
"Something like that."
The second candidate canceled. The third was a recruiter who'd misunderstood the listing and pitched his services instead of his engineering skills. The fourth didn't show up.
---
[Same Office — Two Days Later, Afternoon]
The email had arrived in the general inbox, which Sarah monitored because Ethan's email management consisted of reading the first three words and deciding whether to panic.
Mr. Gardner — I saw your posting for ML engineers. I don't have a degree. I currently work IT support at a medical billing company. I've been teaching myself machine learning for two years using online courses, textbooks, and whatever compute I can access through free AWS credits. I built a character-level RNN that generates poetry — it's terrible poetry, but it converges, which is more than I could say for my first fifty attempts. I'm including a link to my GitHub. I understand if this isn't what you're looking for, but I wanted to try. — Marcus Reeves
Sarah had forwarded it with one line: This is either a waste of time or the best hire we make.
Ethan clicked the GitHub link. The repository was organized with the meticulous care of someone who'd learned to code without an institution validating the process — extensive comments, clear README files, commit messages that read like journal entries documenting a self-guided education. The character-level RNN was implemented in Theano — the same framework Ethan used for the Transformer — with custom training loops and a hand-written backpropagation engine that demonstrated the man understood the mathematics underneath the abstractions.
The poetry was terrible. Gloriously, earnestly terrible. But the loss curves showed clean convergence, which meant the architecture worked, which meant Marcus Reeves understood gradient flow and optimization theory and the practical engineering of making neural networks function.
"Bring him in," Ethan said.
Marcus Reeves arrived for the interview wearing khakis and a button-down shirt that was ironed with the particular precision of someone who owned exactly one interview outfit. He was mid-twenties, Black, with close-cropped hair and the posture of a person who'd spent years being underestimated and had stopped adjusting for it.
Talent Resonance: 7.5.
The number hit clean and strong — not the pyrotechnic flare of Sarah's 9.5, but a solid, dependable signal that said: this person has real ability and hasn't begun to reach its ceiling. A 7.5 with formal training and proper resources could become an 8. An 8 with the right environment could become an 8.5. The trajectory mattered as much as the starting point.
Sarah ran the whiteboard session. Marcus handled the basics differently from the Stanford candidate — not with the polished fluency of someone reciting learned material, but with the careful, deliberate precision of someone who'd built each concept from raw materials and understood how the pieces connected because he'd assembled them himself.
When Sarah asked the attention modification question — the same one that had stumped the PhD candidate — Marcus stood at the whiteboard for thirty seconds without writing anything. Then he drew a sparse attention pattern where each token attended only to a local window plus a set of global anchor positions, reducing the quadratic complexity to linear while preserving long-range dependencies.
Sarah looked at Ethan. Ethan looked at the whiteboard.
"That's not in any paper I've read," Sarah said.
"I haven't read many papers," Marcus replied. "I just thought about what would happen if you didn't need every token to see every other token."
"You independently reinvented local-plus-global attention."
"I don't know what that's called. I just drew what made sense."
Sarah turned to Ethan with the expression she reserved for moments when her skepticism was being dismantled in real time. Ethan gave the smallest nod.
"Marcus," he said. "How quickly can you start?"
---
[Bay Area AI Meetup, SoMa Event Space — Late March 2014, Evening]
The meetup was small — forty people in a converted loft, folding chairs arranged in rows, a projector screen showing the evening's schedule. Three speakers: a Berkeley PhD presenting on convolutional networks, a Google engineer discussing internal language tools, and Ethan, who'd been invited to present "next-generation approaches to natural language processing" by the organizer, a friend of David Park's who'd heard about the Raviga funding and wanted fresh content.
Ethan kept the presentation broad. No architecture diagrams. No mention of the Transformer by name. Just a high-level overview of attention-based approaches to language generation, illustrated with output examples from the model. The audience was engaged — twenty minutes of questions, mostly technical, mostly from the kind of people who attended AI meetups in 2014 because they genuinely cared about the field rather than its investment potential.
Afterward, during the networking portion that Ethan tolerated the way he tolerated dental appointments, he scanned the room with Talent Resonance running in passive mode — the background hum of assessments that happened automatically when he paid attention to someone for more than a few seconds.
Fours. Fives. A six near the drinks table. Standard distribution.
Then, near the back of the room, leaning against the wall with his arms crossed and no drink in his hand: a man with a beard, wearing a black t-shirt with a pentagram, watching Ethan with an expression that communicated nothing and evaluated everything.
Gilfoyle.
Talent Resonance: 8.5.
The number was electric. Higher than Richard's 9 in the raw engineering domain? No — different. Richard's 9 was creative, inventive, the spark of original algorithmic thinking. Gilfoyle's 8.5 was structural. Systems architecture. Security. The ability to build infrastructure that didn't break, that scaled, that survived hostile conditions. A different kind of brilliance, equally rare.
Gilfoyle wasn't participating. He wasn't networking. He was standing at the back of the room the way a predator stands at the edge of a clearing — still, patient, watching. His gaze was fixed on the projector screen where Ethan's output examples still displayed, and his expression was the particular blankness of someone doing complex analysis behind a facade of indifference.
Ethan didn't approach. The instinct — built from seven weeks of navigating this world, from the careful dance of hiding abilities while deploying them — told him that approaching Gilfoyle uninvited would be counterproductive. Gilfoyle was Pied Piper's systems architect. He was also, if the show's characterization held, a fiercely independent thinker who despised salesmanship and respected only demonstrated competence.
You didn't recruit Gilfoyle. You built something impressive enough that Gilfoyle recruited himself.
The meetup ended. People filtered out. Ethan collected his laptop from the projector table, exchanged polite goodbyes with the organizer, and headed for the exit.
Gilfoyle was gone. But the space where he'd stood still held the residual weight of his attention — the particular impression left by someone who processed information at 8.5 and kept their conclusions to themselves.
---
[Honda Civic, Driving Home — 9:30 PM]
The car's heater worked intermittently, producing bursts of warm air followed by periods of resigned silence. Ethan drove through SoMa's evening traffic — delivery bikes, ride-shares, the occasional startup bus returning late-shift workers to their apartments. The radio played something from 2013 that the body's previous owner had preset and Ethan had never changed.
Marcus Reeves had accepted the job offer that morning. Starting Monday. Salary: fifty-five hundred a month, slightly below what Sarah made, because the budget was what the budget was and Marcus had negotiated with the pragmatic acceptance of someone who'd been making forty thousand a year in IT support.
The team was three now. Three engineers. An office above a sandwich shop. A hundred and thirty thousand dollars in the bank after first-month expenses. A working prototype. A Raviga board observer. And a twelve-month clock ticking toward the milestone that would determine whether any of it continued.
Gilfoyle's face surfaced in his memory. The beard. The pentagram. The analytical stillness. In the show, Gilfoyle worked for Pied Piper. He was loyal to Richard in the way that competent people are loyal to other competent people — not through sentiment but through respect for the work. His departure from Pied Piper was never something the show explored because it never happened. Gilfoyle stayed.
But in this timeline, Gilfoyle had stood in the back of a room and watched Ethan present attention-based language generation. He'd seen output from a Transformer. He'd processed something that, to an 8.5 systems architect, would look like either a breakthrough or an impossibility.
If Gilfoyle started looking into Gardner Analytics — into the company's cloud bills, into the infrastructure that supported training runs of this scale, into the impossible gap between a two-person startup and the compute required to produce the output Ethan had demonstrated — he would find anomalies. ChronoCloud didn't show up in any cloud provider registry. The billing patterns wouldn't match any known service. The hardware specifications wouldn't correspond to any available product.
Gilfoyle, more than Monica, more than Sarah, had the technical expertise and the analytical disposition to pull the thread that unraveled everything.
The heater coughed out one last burst of warmth and went silent. Ethan turned onto his street. The apartment building was dark except for the third floor — someone's television flickering blue through curtained windows.
He parked. Killed the engine. Sat for a moment in the cooling car.
Three people now trusted him. Sarah, who'd quit her job on instinct. Monica, who was building a pattern file she couldn't yet read. Marcus, who'd bet his career on a cold email.
And somewhere in Palo Alto, in a house that doubled as a startup incubator, Bertram Gilfoyle was thinking about attention mechanisms and wondering how a company nobody had heard of six months ago was producing output that shouldn't exist for another three years.
Ethan grabbed his bag and headed inside. Tomorrow: Marcus's first day. The beginning of the real build. The model needed to scale. The product needed to ship. The revenue milestone needed to be met. Twelve months, and the Transformer had to prove it was a business, not just a technology.
The lock stuck. He shouldered the door open — same door, same swollen wood, same resistance as the first night in this apartment. But the man entering wasn't the same man who'd woken here in January, confused and alone with a dead stranger's coffee maker and a blueprint he couldn't explain.
That man had been lost. This one had a team, an office, a funded company, and precisely twelve months to change the world.
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