Chapter 89: The Photograph Stops Developing
[Sequoia Capital, Sand Hill Road, Menlo Park October 14, 2015, 2:00 PM]
Monica's deck was forty two slides, cut down from fifty eight over the weekend, and every cut had cost her something she did not show. The valuation slide was clean. The ARR trajectory was clean. The unit economics were clean enough that Andrew Lau had nodded once, early, and the nod meant the financial story would not be the obstacle.
Ethan's section was slides forty three through fifty one. Nine slides. Architecture, training methodology, the compute roadmap, and the timeline to convergence written three weeks ago, still correct.
He stood at the screen and walked the room through the transformer architecture: encoder stacks, positional embeddings, the attention mechanism that made the whole thing go. The four partners across the table tracked him closely, the intent focus investors give a founder off script. Not reading his slides describing the model from memory, precisely, completely. It was a familiar performance, run a dozen times before. The image was as clear in his head as the room around him.
Rebecca Zhou was the only one whose pen moved.
Baidu was on her résumé production NLU shipped at scale. She did not ask what the model did only how it failed, and he liked her for that. The first two questions were fair: gradient instability at depth, the vanishing signal problem at thirty plus layers, why layer norm over batch norm. The shape was there, and he answered from it.
The third question was different.
"Head allocation at scale," Rebecca said. "You're at twelve heads now in your published work. At one seventy five billion, what's the allocation strategy? Fixed, learned, or something else?"
The blueprint was there. It had been there all morning, and he reached for it the same way he always did. The model hung in his head, a photograph he had memorized. Every layer, every connection, visible and stable and his.
The photograph was out of focus.
Not gone. Not dark. He could still see the broad strokes: the encoder stack, the residual connections, the general sense of where the heads should sit. But the fine detail, the part where the allocation strategy lived, had gone blurry, indistinct, unreadable. Five minutes ago, he had been able to read it clearly. For two years, always.
Four seconds passed. He felt them. The room did not.
"At that scale," Ethan said, "fixed allocation breaks the gradient economy. You learn it. The heads diverge into role specialized groups. Some attend to syntax locality, some to semantic distance, some to position. The supervision signal is sparse early and you don't constrain it. Let the loss landscape find the partition."
The photograph had gone soft. No longer looking at it, he built the answer from the mechanism instead: how attention worked, what he knew about gradient flow and head specialization in shallow models. Extrapolated upward. Reasoned through rather than retrieved. He had never done this in a room of investors, had never needed to.
The answer was correct, known by checking the steps, the way you verify a proof, not by simply recognizing the theorem on sight.
Rebecca wrote something down. Not the answer. Something else. Her pen moved, and he could not read what it said from where he stood.
Andrew Lau leaned forward. "The convergence timeline. You're saying January, early February at the outside. What's the gating factor?"
"Data quality," Ethan said. The fog was still there. He was not reaching for the photograph anymore he was talking. "We're at nine point four billion tokens now. The pipeline scales to twenty by November. The model doesn't converge on bad data and we don't ship bad data."
Monica took the valuation conversation at slide forty six, smoothly, mid sentence. Two words in, he realized his answer to the pricing question was going to be the honest one: the round was worth what they were asking. He was about to say so in a register that read as confidence, not negotiation. She stepped on it before he could, said the number, said what it was based on, and sat down. The room did not notice the handoff, except that Ethan stopped talking and Monica started. In any pitch room, that was the point.
The meeting ended at three forty. Andrew shook his hand. The other two partners, Keith Park, who had said almost nothing, and a senior associate whose name he had not caught, shook his hand. Rebecca shook his hand and said the architecture was impressive and she would follow up on a few items.
Her pen had written two things during the meeting. One was the answer. One was not.
[Sand Hill Road parking structure 3:52 PM]
The Honda was on the third level. The AC rattled even though it was sixty one degrees and he did not need it, turned on out of habit and left running.
Monica was in the passenger seat, laptop open, rebuilding the comp model from the meeting's feedback in real time. She had the valuation at $185M to $195M, pre money, with a likely term sheet range of $50M to $65M. She did not say this out loud, only showed him the screen.
"That works," he said.
"Andrew's going to push for a board observer seat," she said. "Pre empt it. Offer it before he asks."
"Okay."
She typed. He sat with his hands on the wheel and the engine off and the AC rattling.
Sarah was in the back seat, had been for all four pitches. Her role was to listen and take notes on a legal pad, and to speak only if the technical section was challenged. It had been challenged, once, by Rebecca. Sarah had not spoken.
"You went still," she said now.
His eyes went to the concrete wall. "When?"
"After the head question. You went still for about four seconds and then you answered. The answer was good. The stillness was new."
"I was thinking about scaling problems."
Sarah's red marker was in her hand, uncapped, not going to be used. Her hand knew where it went, a habit older than this meeting.
"Okay," she said.
She recapped the marker. She did not ask what scaling problems meant, had never asked how he knew the things he knew, not once. Not going to start now, not in a parking structure on Sand Hill Road.
Monica closed her laptop. "Series B terms within three weeks. I'd expect Sequoia first, then Felicis, then maybe Baseline. Trinity will pass."
"Trinity will pass," Ethan agreed.
"Rebecca's going to follow up," Monica said. "She does that. She follows up on the one thing she didn't buy completely."
Ethan started the car. The AC rattled. Sarah's legal pad had one line on it he could read in the mirror.
Chen went still @ 2:14 PM, ~4 sec.
Into the October light, he pulled out of the structure. The fog remained in his head. Not fog. The photograph, the blueprint of the model he was building, clear and stable for two years, now seen through water, remained there. It remained there. It was not as sharp.
He did not know what that meant, and he drove.
[Noe Valley apartment 9:15 PM]
Monica was asleep, gone to bed at eight thirty, which meant the meeting had cost her more than she had shown in the car. On the couch, he sat with his phone.
Rebecca's email arrived at 9:07 PM. Two paragraphs. The first was warm: impressive architecture, strong methodology, clear roadmap, looking forward to continuing the discussion. The second paragraph was four sentences.
For the diligence file, could you put the head allocation answer in writing? The mechanism you described, learned role specialized partition via gradient economy, is something I want to reference specifically. We typically document technical claims from founders as part of our process. Happy to discuss over a call if easier.
Twice, he read it. The email was asking him to put in writing the one answer he had not retrieved from the blueprint. The one he had built from reasoning the one she wanted in a file.
The phone went face down on the cushion. The AC in the apartment rattled too, or maybe that was the refrigerator. The Honda was four blocks away. It had started on the first try this morning, which was unusual, and which he had not noticed then and did not think about now.
The photograph in his head remained out of focus, and he sat with it, not knowing what to do with it, not knowing what it meant. An ability he had always treated as permanent, seeing the model clearly, completely, on demand, was now behaving differently. Rustier. Slower to answer.
He picked up the phone and typed a reply to Rebecca. Two sentences: the written answer would come by Friday, and thanks for the question.
The second paragraph of her email would become a line in a term sheet, though he did not know that. What he knew was that the meeting had gone well, and the fog had not cleared. Ahead lay the harder part: describing a model he could no longer see perfectly, to people paying one hundred and sixty five million dollars to watch him build it.
He sent the reply. The phone went dark. The refrigerator rattled. The photograph stayed soft.