The Task Designer

The brief came in at 7:14 on a Tuesday morning in April. Hira read it twice over coffee and then opened her notebook.

Q3 capability targets: extend math reasoning into research-level combinatorics and topology. Need 8,000 seed problems by May 9. Diversity requirements: 40% combinatorics, 30% point-set and algebraic topology, 20% geometric topology, 10% boundary cases between fields. Difficulty floor: open problems published 2018-2024. Difficulty ceiling: open problems considered approachable by senior graduate students. No reproductions of known problems. Each seed must include a verification sketch.

She wrote down the date, the initials of the project lead — J.K. — and the brief in her own handwriting, the way she had been writing notes on yellow legal pads since her second year of graduate school. The lab provided perfectly good digital tools for everything. She used them too. But the legal pad was where she actually thought.

The notebook had blue-lined paper, slightly grey from being made from recycled stock, and her sister had given her a box of forty of them when she had left her tenure-track position in 2023. In case you miss it, her sister had said. Her sister was a school nurse in Akron and had not really understood what Hira was leaving or what she was going to. Hira had not been entirely sure herself.

She thought about combinatorics for a while. Eight thousand problems was a lot. Three years ago, when she had been writing original papers in extremal combinatorics, eight thousand original problems would have been the work of her entire career and the careers of fifty people like her. Now it was a quarterly deliverable for one task designer working at one frontier lab. There were forty-three other task designers working on math at the lab. Last quarter their combined output had been roughly four hundred thousand seed problems.

The seed problems were the first stage. The lab's frontier model, running in some cluster she would never see, would generate hundreds of candidate solutions per seed. A second model, from a competitor's API, would critique the candidates. A third model, smaller and specialized, would verify the surviving solutions against Lean 4 formalizations she would never see. The verified solutions would become training data for the next frontier model. The next frontier model, three months from now, would solve problems Hira could not.

She knew this because she had been watching it happen for two years.

She wrote her first three seeds before lunch. They were variations on a Ramsey-theoretic problem she had thought about during her postdoctoral year at the IAS, before she had learned that the Ramsey program was a dead end for tenure unless you happened to be one of seven people in the world. The lab paid her $340,000 a year plus equity and asked her to think about Ramsey-theoretic problems whenever she felt like it, as long as the resulting seeds met the diversity quotas.

It was, in a literal sense, the job she had wanted at twenty-six and been told did not exist.

She wrote two more seeds in the afternoon. The third was a topological problem about the homotopy groups of certain configuration spaces, and she could not quite get the verification sketch right. The verification sketch was the part of the job that distinguished a useful seed from a useless one. The lab's frontier model could generate plausible-looking solutions to almost any problem. The question was whether those solutions could be verified, and verification required that the seed itself be expressible in a formal framework the verifier model knew how to handle.

She worked on the verification sketch for an hour. Then she gave up on it for the day, made herself an espresso, and walked to the window of her home office. Outside, the late April light was hitting the maple tree in her front yard at the angle she always thought of as the mathematician's hour — the long oblique light of late afternoon that made every surface look more textured than it was.

Her sister called.

"Did you see the article today?" her sister asked. "About the data wall?"

"I saw it."

"It said the wall already fell."

"It did."

"Was that — were you — did you know that was happening?"

Hira looked at the maple tree.

"Yes," she said. "I work in it."

There was a pause. Then her sister said, "It was on the front page of the New York Times. I just thought you should know in case you hadn't heard."

"I appreciate it."

"Are you okay?"

It was a strange question. Hira thought about how to answer it.

"I'm fine," she said. "I'm working on a problem about configuration spaces."

"Okay," her sister said. They talked for a few more minutes about her sister's children and her sister's husband and a complicated situation at her sister's school involving a substitute teacher and a parent group and a board meeting. Then her sister said she had to go and Hira said she loved her and they hung up.

Hira sat on the windowsill for a while. The maple tree was just beginning to leaf out, a pale green almost translucent in the light. She thought about the configuration space problem.

The lab held a quarterly all-hands every three months. Hira attended remotely from her home office. The all-hands always included a capability presentation from the head of research, a man Hira had met twice and who had shaken her hand both times with the same brisk courtesy. He spoke for thirty-five minutes about the next frontier model. He showed benchmark numbers Hira had seen in advance through internal channels. He talked about safety milestones and infrastructure expansion. He thanked specific teams.

He did not, that quarter or any previous quarter, mention the task designers by name.

This was not personal. The lab had decided that capability advances should be attributed to the model and the company, not to specific contributors, and the policy was applied uniformly. Engineers were not named. Researchers were not named. Task designers were not named. Only the senior leadership and, occasionally, the safety team appeared in the public-facing capability presentations.

Hira understood the policy. She did not particularly resent it. She had spent eleven years in academia where credit was the currency of survival, and she had found it exhausting. The lab paid her enough that she did not need credit. The work was the work. The next frontier model would solve problems she could not, and her seed problems would have contributed to that capability, and the model would not know her name and the public would not know her name and the historical record, if there was a historical record, would not know her name.

Sometimes she thought about this. Sometimes she did not.

That afternoon, after the all-hands, she went back to the configuration space problem.

The verification sketch came together at 6:40. She wrote it out on the legal pad first, then transcribed it into the seed submission tool. She attached the formal framework reference. She added the difficulty tag (senior graduate student, approachable) and the diversity tag (geometric topology, boundary case with combinatorics). She submitted the seed.

It was the 1,847th seed she had submitted that quarter. She had 6,153 to go in the next four weeks.

That evening she had dinner with her husband, who was a software engineer at a different company and was patient about the work she could not discuss. They ate pasta at the small kitchen table. He told her about a deployment that had gone sideways and how his team had spent the day rolling it back. She told him about the maple tree.

After dinner she walked outside. The maple was still just beginning, the pale green so new it almost looked white in the dusk. She put her hand on the bark and stood there for a while.

She thought about the next frontier model, three months from now, solving problems she could not solve. She thought about whether that should bother her or whether it should not. She thought about the seventeen graduate students she had supervised over the years before she had left, and where they were now, and what they were doing. She thought about how few of them were still in academic mathematics. She thought about how many of them were doing something like what she was doing — designing seed problems, designing verification environments, designing the substrate from which the next thing would grow. Four of them, that she knew of. There were probably more she had lost track of.

She thought about whether there would be a future Hira, twenty-five years from now, designing seed problems for whatever came after frontier language models. She thought about whether that future Hira would feel about her work the way she felt about hers. She thought about whether the future Hira would have a sister who called from Akron to ask if she was okay.

The leaves were too new to rustle properly. The wind moved through them and they only barely registered. She stood for a long time in the pale green dusk.

Then she went inside and read for an hour and went to bed and slept the sleep of a person who has done six hours of difficult and useful work in a day, which is the kind of sleep no other thing produces.

In the morning her brief queue had a new line.

Q3 capability targets revised: extend math reasoning into research- level number theory. Need 4,000 additional seed problems by May 16. Same diversity, difficulty, and verification requirements. Coordinate with M.T. on number-theoretic verification framework.

She read it twice and made a fresh espresso and thought about number theory.

The maple tree outside the window had put out, overnight, a layer of new leaves she had not noticed at dinner the day before. It was the hour before the world had quite begun to make its noise. She opened the legal pad and wrote the date and the initials of the project lead — M.T. — and the brief in her own handwriting.

Then she sat down to think about which open problems in analytic number theory a senior graduate student could approach, and which of those problems were currently expressible in a formal framework the verifier model knew how to handle.

It would take her twenty minutes to write the first one. She knew this because she had done it many times before, and she knew how the rhythm of the work felt when it was about to start.

She picked up her pen.

If this story sat with you, you might also want to read the accompanying analysis of the synthetic data tipping point and how AI started training itself, or the related short story The Chart the Engine Couldn't Read.