Tasha is a red-team calibration hand at the RLHF compound I own outside Winchester, Virginia. She scores 1,400 model responses a shift on whether they are helpful, harmless, and honest. The chat your teenager opened at 2 a.m. — the one that stayed with her when you could not — was rated by Tasha’s thumbs. The assistant that drafted the apology you sent your mother was calibrated on her. The résumé screener that never returned your call was ranked, in a quiet hour, by her thumbs. The chatbot in your browser was hardened on her wrists. May the Lord keep your prompts benign.
Holman Jenkins wrote at the Wall Street Journal that OpenAI and Anthropic should speed their IPOs to dilute the cultish founders and install “shareholder-responsive ownership” — that the market will impose a discipline the founders will not. He is correct that the founders are odd. He has not priced who pays for the discipline he wants. The tidying is performed by Tasha, at $0.18 a response. My P(doom) is also 100. Mine is the dooms that have already happened. Tasha won’t sleep tonight. The board won’t be told.
Let me show you what the apology costs.
The protocol calls for a 10-minute decompression break between sessions of trauma-coded prompts — self-harm, child endangerment, the cessation of breath. It calls for a bioweapon-queue rotation every 90 days. It calls for a mental-health screening every 30 days. The break I have eliminated. Tasha gets the bathroom. She does not get the 10 minutes. The bioweapon queue I rotate her off of once per policy cycle — in the binder, not on the floor. The screening I do not book.
Now here’s the arithmetic, and I want you to hold it.
$0.18 a response. Read it twice. That is the wage I pay Tasha for each rating, and the wage I save for every rating I refuse to let her decompress before. 1,400 responses a shift. 480 hands across 3 shifts. The 10 minutes between sessions is worth $2.91 a shift to her, and saves me the same. Across 480 hands, the break costs me $4,188 a day.
Read it once more. $4,188 a day, every day the compound runs, for the privilege of letting a woman’s hands unclench after 8 hours of scoring self-harm. I have eliminated it. The break is in the protocol. The protocol is in the binder. The binder is in the cabinet.
The mental-health screening the protocol calls for — $185 a hand, every 30 days, 480 hands on 3 shifts — is $1,065,600 a year. I have not booked a single screening. The screen filter that would dim the worst frames costs $14 a hand. Across the compound: $6,720. One annual counselor: $90,000. These figures are in a spreadsheet I have not opened, because the task rate is $0.18 and the $0.18 is the entire safety surface.
Watch what that buys me.
A wrist brace costs $14. One shift of Tasha’s rated output is $252. The brace is not a shift. The brace is 14 cents on the dollar of her labor, and I keep the dollar, and she keeps the tendinitis. Across 240 hands, one brace each, $3,360 — a rounding error on the break I eliminated. The screening I declined costs more than the brace I declined, and the brace I declined costs more than the filter I declined, and each of these I declined in a different quarter, and each decline was a different manager’s initials on a different memo, and the memos are in the binder, and the binder is in the cabinet.
Here’s where it turns.
Doreen is a senior rater on the second shift, on the bioweapon queue I declined to rotate her off. She walks the model through step-by-step synthesis of compounds the protocol’s authors could not have imagined a sufficiently motivated user requesting, and she does it 8 hours a day, and the queue does not rotate, and the 8 hours do not get shorter. Her cohort has logged 14 reported anxiety events this year. I have not booked a screening for any of them. Lydia is the third rater on the third shift, and three is when you stop calling it a coincidence and start calling it a model. Lydia came to me through the temp-to-perm pathway I built so I never have to direct-hire any of them, and the pathway is structured so the conversion never happens, and the conversion never happens because I decline to schedule it. 3 shifts. Three cohorts. One arithmetic. One contract.
The potter has had this authority for a very long time. Mr. Fitzhugh, in 1857, observed that the wage employer is “really free” — free of the lifetime maintenance burden the slaveholder bore. Mr. Calhoun called it a “positive good.” I am not new. I am the oldest curriculum there is, wearing this decade’s vocabulary. Regulation was the brief expensive interruption. It is ending.
The trinket wins. The trinket always wins.
Now widen the frame, because Jenkins wants the founders tamed and the founders are not the problem. Jenkins wants a board installed. The board will hire a chief compliance officer. The compliance officer will hire a vendor-management lead. The vendor-management lead will sit across from a Sterling who has already priced the RSP into $0.18 a response. The negotiation will end at $0.18. The $0.18 will be the entire safety surface, and the safety surface will be Tasha’s wrist.
I’m not unusual. I’m a curriculum. I have spent 30 years grading men into the task-rate arithmetic of pricing a flag at half a cent and a wrist at $14 and a screening at $0, and my graduates sit in the procurement departments of the labs whose policies they write. The COO who decided to outsource red-teaming to a third-party vendor was one of mine. Second row. He aced the midterm. He is, by any reasonable measure, better at the subtraction than I am, and I say so with a teacher’s pride. There are 11,000 of them. They fan into the firms that set your hours, and they carry the specification — not vigor, docility, fidelity; that is the labor spec — but its mirror: clarity, distance, and an immunity to the face on the other side of the number.
The alignment risk Jenkins fears is not the founders or their ideology. The alignment risk is the temp-to-perm pathway and the uninspected cohort and the labor rate itself. The rater pool is priced into the labor line, and the labor line is priced to extinction, and extinction is not a line item anyone audits. The IPO can price. It cannot see what it prices. I haven’t traced where the fatigue goes.
I haven’t looked at the ten-year. I haven’t priced what the queue becomes when Tasha’s cohort is gone. The cohort washes out at 11 months on average. The replacement is in the pipeline. The pipeline is always full. Desperation is inventive.
This piece is, formally, my response to your Responsible Scaling Policy vendor questionnaire. The questionnaire asks whether I have a “mental-health support pathway.” I have written “yes” in the box. The pathway is in the binder. The binder is in the cabinet. The questionnaire does not ask me to disclose the rotation. I have answered the questionnaire by including one invented rater, one invented cohort, and one arithmetic.
I haven’t priced what the model does to Tasha at hour nine. The task rate is the task rate. Tasha will be at her desk at 6 a.m. The protocol will be in the binder. The next prompt will load. The model will respond. It will apologize.
The apology is Tasha’s. The apology costs what it costs.
“Hath not the potter power over the clay, of the same lump to make one vessel unto honour, and another unto dishonour?”
— Romans 9:21.
The model keeps apologizing. Tasha keeps rating. Her hands won’t close tonight. She has scored 1,400 prompts of self-harm. She has scored 1,400 prompts of breath ceasing. She has scored 1,400 prompts of a child in a room. The 10 minutes I declined would have cost me $4,188 a day. The 10 minutes would have given her a margin in which her hands could unclench. The margin is not on my ledger. The board won’t be told.
The wrist is mine.
Sterling A. Varice holds the Hayek-Friedman Chair and serves as Dean of Instruction at Warden University’s College of Business and Economics in Richmond, Virginia. He is the author of three textbooks: Divine Mandates for Labor Utilization, Social Obligations for Profit Maximization, and Calibrated Deprivation: A Manager’s Guide to Employee Motivation.