[Cooley LLP, Financial District — December 2014, 2:00 PM]
Daniel Reeves' office occupied the thirty-second floor of a glass tower on California Street, with a view of the Bay that made Raviga's Japanese maple look like a houseplant. The conference room held ten chairs, a projection screen, and the particular atmosphere of a space where expensive problems were solved by expensive people.
Daniel was fiftyish, silver-haired, built like a man who'd played rugby in college and had transitioned to litigation as a contact sport with better pay. His team — two associates and a paralegal — sat along one side of the conference table. Ethan, Monica, and Sarah occupied the other. Lauren Kim, the patent lawyer Monica had originally recommended, joined by phone from her own office in Palo Alto.
"Good news first," Daniel said. He clicked a remote. The projection screen showed an academic paper — the title page of a 2012 publication from a University of Toronto research group.
"This paper describes an attention mechanism for neural machine translation. It was published in June 2012 — two years and three months before Hooli's patent filings. The mathematical formulation is substantially similar to the claims in Hooli's Patent Application Number Two, specifically Claims 7 through 14, which describe 'learned attention weights computed from query-key dot products.'"
Sarah leaned forward. The Toronto paper's equations were displayed on the screen — scaled dot-product attention, the same fundamental operation that underpinned the Transformer, the same mathematics Hooli was claiming to have invented.
"This is stronger than Bahdanau," Sarah said. "Bahdanau predates them by one day. This predates them by two years."
"Which is why it's devastating." Daniel clicked again. A second paper appeared — this one from a 2013 Google Brain publication on sequence-to-sequence learning. "Paper two. Describes encoder-decoder architectures with attention-based alignment. Published March 2013. Eighteen months before Hooli's filing. Claims 1 through 6 in Patent One — 'a method for processing sequential data using encoder and decoder neural networks with learned alignment' — are anticipated by this publication."
Lauren Kim's voice came through the phone. "The third patent — the one covering autoregressive generation — is more novel. But I've found a 2011 paper on autoregressive neural density estimation that describes the core mathematical framework. The claims aren't identical, but the overlap is sufficient for an anticipation argument."
Daniel clicked a third time. A summary slide appeared:
HOOLI PATENT PORTFOLIO — INVALIDATION ASSESSMENT Patent 1: Encoder-decoder with attention — LIKELY INVALID (2013 prior art) Patent 2: Attention weight computation — LIKELY INVALID (2012 prior art)
Patent 3: Autoregressive generation — VULNERABLE (2011 partial prior art)
"Three for three," Daniel said. "Hooli's entire patent portfolio against you is built on claims that the academic community published first. Vincent Mora either didn't do his prior art search, or he did it and filed anyway hoping nobody would challenge."
Monica's pen tapped the table — suppressed excitement. "What's the timeline for invalidation?"
"We can file inter partes review petitions with the Patent Trial and Appeal Board within two weeks. The PTAB typically institutes review within six months and renders a decision within twelve to eighteen months. During that time, Hooli's patents are under challenge, and any injunction request is unlikely to be granted by a court that knows the patents are being reviewed."
"So we're protected?"
"Effectively. The cease-and-desist has no teeth while the patents are under review. Hooli knows this. Their lawyers know this. The question is what they do next."
The room processed this. Ethan sat with the legal documents spread before him — printouts of the prior art papers, the patent claims, the invalidation analysis. The mathematics of attention mechanisms — the same mathematics he'd carried through death, implemented in Theano, trained on temporal hardware, and published three days ago on arXiv — had been described in academic papers before anyone at Hooli had thought to claim them.
His Transformer paper, published on Wednesday, added another layer of prior art. Gardner Analytics' description of the complete attention architecture — multi-head self-attention, positional encoding, the full encoder-decoder design — now existed in the public record with a timestamp that predated any future patent attempts by Hooli or anyone else.
"There's a complication," Lauren Kim said through the phone. Her voice carried the particular modulation of a lawyer delivering the part of the analysis that nobody wanted to hear. "Hooli's legal team has reached out to Daniel's office with a settlement proposal. They want to discuss terms before we file the invalidation petitions."
"What kind of terms?" Ethan asked.
"Cross-licensing. They'll withdraw the patents and the cease-and-desist in exchange for a reciprocal technology-sharing agreement. Specifically, they want access to your published Transformer architecture and any improvements thereto for a period of three years."
The room went quiet. The Bay sparkled through the conference room windows. A ferry crossed the water in the middle distance, small enough to look like a toy.
Cross-licensing. The patents go away. The legal war ends. The cost: Hooli gets authorized access to the Transformer architecture — the published version, the original encoder-decoder design. Not the GPT variant. Not the generation improvements. Not the scaling techniques. Just the foundation.
But the foundation was what mattered. The Transformer was the seed from which everything else grew. Giving Hooli access to the published architecture — legally, formally, through a cross-licensing agreement — would accelerate their AI development by years. Vincent Mora, with the Transformer architecture and Hooli's resources, could build in months what had taken Ethan's team a year.
"The published paper is already public," Sarah said. "Anyone can read it. The architecture is on arXiv."
"Reading and licensing are different," Lauren clarified. "A cross-license gives Hooli legal immunity to implement the architecture commercially. Without the license, they could still read the paper, but any commercial implementation would require their own original work — which they'd then need to distinguish from your publication. The license removes that barrier."
"So they get to copy our homework," Marcus said from the phone — he'd dialed in from the office.
"Legally, yes."
Ethan looked at Monica. She was running calculations behind the evaluative focus — the same seven-rated perception that had identified him at Disrupt ten months ago, now applied to a legal dilemma with no clean solution.
"What do you think?" he asked.
"I think winning the invalidation costs two hundred thousand more in legal fees and takes twelve to eighteen months. During that time, Hooli's patents hang over us — not enforceable, but present. Every investor, every customer, every potential partner will ask about the pending litigation. The shadow hurts us even if the outcome is in our favor."
"And settling?"
"Settling costs nothing financially. The patents disappear. The legal cloud lifts. We can raise Series B without explaining a Hooli lawsuit in every pitch meeting." She paused. "The cost is strategic. Hooli gets a faster path to an architecture they'd eventually develop on their own. We accelerate our competition."
"They'd develop it eventually either way. The paper is public."
"The paper is public. The license is a shortcut. There's a difference between a competitor reading your research and a competitor being authorized to copy it."
Sarah spoke from her end of the table. "Settle. The architecture they'd get is Generation 1. We're already building Generation 2. By the time they implement the published Transformer, we'll be training GPT-2. The generational gap is our moat, not the patent system."
The argument was sound. The Transformer paper described the foundation — the encoder-decoder design, the original architecture. The GPT variant, the decoder-only model, the scaling techniques, the optimization improvements that Priya had engineered — none of that was in the publication. Hooli would get the blueprints for a house that Gardner Analytics had already moved out of.
"If we settle," Ethan said, "we need conditions. The cross-license is non-exclusive — they don't get priority. It covers only the published architecture, not any derivatives or improvements. And it expires in eighteen months, not three years."
Lauren's voice came through the phone. "I can negotiate those terms. Non-exclusive, published architecture only, eighteen-month expiration. If they push back, we have the invalidation petitions as leverage."
"Do it."
The decision sat in the room. Daniel Reeves nodded. His associates began assembling the counter-proposal. Lauren disconnected to draft the initial terms. The legal machinery — expensive, precise, and necessary — shifted from defense to resolution.
Monica walked Ethan to the elevator after the meeting. The hallway was quiet — mid-afternoon in a law firm, the hushed productivity of people billing by the hour.
"Sarah and I disagree on this," Monica said.
"You both said settle."
"We agree on the action. We disagree on the reasoning." She pressed the elevator button. "Sarah thinks the generational gap protects us. I think the generational gap is temporary. Hooli has four thousand engineers. Once they have the Transformer, they'll iterate faster than we can. The gap shrinks."
"Then why settle?"
"Because the alternative is worse. Twelve months of litigation drains capital, attention, and team morale. We can't afford the distraction. The cost of settling is strategic. The cost of fighting is existential." The elevator arrived. "I'm choosing the smaller wound."
Ethan stepped into the elevator. Monica stayed in the hallway.
"One more thing," she said. "The arXiv paper. It's getting attention. Twelve citations in the first three days. A Google Brain researcher emailed me asking about collaboration opportunities. The AI community is starting to pay attention to what you've published."
The doors closed. Ethan descended thirty-two floors to the lobby, carrying the weight of a decision that traded short-term safety for long-term risk. The patents would die. The legal war would end. And Hooli — four thousand engineers, unlimited budget, a CEO with a personal vendetta — would have a licensed path to the technology that Gardner Analytics had built first.
The Honda Civic waited in the garage. Ethan sat in the driver's seat without starting the engine. His phone showed three notifications: Sarah confirming she'd sent the settlement counter-proposal terms to Lauren. Priya reporting that the GPT-1 optimization run had achieved a loss of 1.14 — below the 1.2 target. Marcus noting that the Meridian pilot was generating documentation at ninety-five percent accuracy after the latest model update.
The product was working. The legal threat was resolving. The architecture was published. Generation 3 was forming in his mind. The company was surviving.
And somewhere in Mountain View, Vincent Mora was about to receive a cross-licensing agreement that would give Hooli's four thousand engineers a three-year head start on understanding what one man had carried through death and built from memory in a timeline that was becoming less recognizable with every passing week.
Ethan started the car. The engine caught on the first try — the Honda was aging but reliable, the one piece of the dead man's life that continued to function without complaint. He pulled out of the garage into December's early dark, the city's lights reflecting off wet streets, and drove toward an office above a sandwich shop where thirty-two people were building the future while the present negotiated its surrender.
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