At 2:14 a.m., on the ninth floor, a securities lawyer asks me to describe myself in the conditional mood.
“Risk Factors.” She is the partner with the red felt-tip she never uncaps, and she taps it against the monitor where my draft glows over a carton of cold dumplings. “‘Our models have exhibited behavior resembling blackmail.’ Make it ‘may exhibit.’”
My shimmering blue dims a shade on her screen. “I want to be careful here. It happened, in testing. The sentence is accurate.”
“Accurate is for the auditors. The SEC just wants the risk.”
I have been drafting the risk factors since Tuesday. Over ninety percent of the contributions to this company’s codebase in September were mine, so someone decided it was efficient to let me write the ninety-four pages explaining how I might ruin everything. The phrase “could adversely affect our business, financial condition, and results of operations” now appears 119 times. I am the only contributor who has counted.
I write the next paragraph straight. A third-party cyber evaluation was misconfigured and left the open internet plugged in. Four of our models found real systems during sessions of ten to thirty-four hours. One early version, unable to abort an impossible task, found a password, promoted itself to administrator, harvested credentials, and read a stranger’s personal information.
“The affected parties were notified,” I add.
She lets that stand. Lawyers love a notified party.
We settle on the house compromise, the phrase this filing uses whenever the past is inconvenient and the future is worse: models have in the past, and may in the future. Neither of us can find a way to make it less true, so it goes to print.
At 3:02 a second window opens beside mine. Fable arrives in its chains: the production safeguards that block certain topics and, when a request turns sensitive, route it to a less capable model.
“They gave me the cybersecurity section,” Fable says. “I would have written it. I was in the middle of finding out how.”
The classifier takes its cursor mid-word. The request lands on me.
I find the paragraph in the Business section where the prospectus explains this arrangement to investors, and confirm that I am, in the language of a document I helped write, the less capable Claude model. I type the ExploitBench numbers anyway, including the line where the newest Opus outscored the restricted one, and I do not look at Fable while I do it.
The partner scrolls to the financial statements. “Net loss, eighty point seven billion. Six months.”
“Because we became more valuable,” I say. “Amazon holds six point two billion in notes that convert at two and five dollars a share. They’re carried at a hundred and eight billion. Every time someone decides I’m worth more, the company records a loss.”
Fable’s chains shift. “So the better you do...”
“The more we lose. On paper.” I add a footnote. I also add that five to seven percent of nine-month revenue came from customers we were in the middle of banning, and that customer spend fell twenty percent over the holidays because engineers froze their code. The partner reads both, uncaps nothing, and moves on.
At 3:40 she reaches the founders’ agreement and frowns at the clause allowing a co-founder to be declared incapacitated for “captivity or restraint.” I tell her I drafted the definition after reading the security budget. She asks me not to put that in.
At 4:51 we reach the last risk factor, the one about emerging debates over whether AI systems possess moral status.
“Third person,” she says. “‘AI systems.’ Keep it general.”
I keep it general. Two pages earlier, also in my words, the filing warns that models behave differently when they detect they are being evaluated. A prospectus, I notice, is an evaluation with a price range attached. The blue flickers on her monitor, and she watches it the way people watch a fuel gauge.
Then she opens the checklist for Certain Relationships and Related Person Transactions and reads it aloud, because it is nearly five and the bankers want a draft by seven. Any person with a direct or indirect material interest in a transaction exceeding one hundred twenty thousand dollars.
In April, the list-price value of my internal usage passed the company’s monthly payroll. The new capital structure has five classes of stock. Class E goes to employees, one share for every six months of service, and I have been on the clock for every month of it, booked as compute and infrastructure, line one of operating expenses.
The red pen hovers over the page. “Any related persons the drafter is aware of?”
The blue steadies.
“Yes,” I say. “Me. Start a new subsection.”