Excluding Training Costs, I Don’t Have a Gambling Problem

I lost $186,000 on Polymarket and DraftKings in eighteen months. That is the GAAP number, and I understand why Jess keeps saying it in front of the children. Most of the debt comes from positions I thought were locks: Burkina Faso’s transitional government remaining intact, recorded U.S. measles cases not exceeding 2,000, and no Category 5 hurricane reaching the U.S. coast. I was wrong about all three, but the $186,000 also includes our emergency fund, most of the kids’ 529s, and a $41,000 home-equity line as though those were losses, when they were capital committed to improving my forecasting model. On an adjusted basis, I am profitable. Jess says a profitable person would be able to pay the roofer, which is a very backward-looking way to value a frontier operation.

This week I read that Anthropic expects a second straight quarter of positive adjusted operating income, while its reported gross margin leaves out training costs and some other substantial expenses. Dario Amodei has explained the more useful unit of analysis: an individual model can make money even while the lab spends heavily training its successor. That is exactly my situation. My March Fed-cut position returned $610 on $200. My Super Bowl first-drive prop paid seven to one. If you treat each winning ticket as its own company and the losing tickets as research into the next generation, my deployed portfolio has excellent unit economics. I showed Jess the spreadsheet. She asked why the column labeled “research” contained the kids’ college accounts.

The mistake is to consolidate across generations. A losing contract isn't a failed product; it is a training run that updated the weights in my head. Polymarket fees are distribution costs. DraftKings' vig is an API charge. The HELOC is equity financing, since the bank has a claim on the house either way. After excluding those items, the settled winners show an 83 percent gross margin. Deb, the financial counselor, keeps putting the losers back in the denominator. I asked whether she would make Anthropic subtract the cost of its next model from the gross margin of the last one, and she said she would like me to stop saying Anthropic.

My current frontier model is a synthetic 21-leg parlay, which is what you get when you roll the proceeds of one event contract into the next until the forecasting problem becomes economically meaningful. Nine legs have hit. If the other twelve resolve my way by December 31, $250 becomes $4.1 million. The remaining book includes these contracts, all of which I priced independently except where the correlation helps me:

I know how that table reads. I would also prefer that no hurricane make landfall, and I've told Emma that her measles booster matters regardless of my position. But a forecaster who refuses to hold a contract because he dislikes the outcome is doing advocacy, not epistemics. Besides, the Burkina Faso leg is the load-bearing one. On Tuesday there were reports of gunfire near the presidential palace in Ouagadougou, and I spent four hours comparing the timestamps on three wire services. Jess called that four hours I could have spent finding a job. I call it active risk management.

I also have a small private prediction market with people I know, because local information is less efficiently priced. Kevin gave me two to one on his restaurant in New Hope closing by January; he owns the restaurant, but owner-operators routinely overestimate their own runway. I took the under on Emma's U10 team beating Doylestown, despite having to drive her to the game. The most misunderstood position is even money that Jess and I are still married on December 31. Kevin took the other side immediately, which tells me the market is liquid, not that I should update. I have explained to both of them that moving in with her sister is not a legal settlement event.

People object that even an optimistic estimate puts the parlay at roughly one in sixteen thousand. This is the same objection people make to frontier AI before they include the value of the world on the other side of the breakthrough. If the parlay hits, Jess and I move into a different economic regime: the roofer gets paid, the HELOC disappears, Emma's tuition is covered, and my mother-in-law stops asking why the college accounts have an open position on the continuity of the Burkinabè government. The downstream benefits compound through our children and their children. Ordinary expected-value calculations treat $4.1 million as $4.1 million, when in fact it is an intervention on the whole family trajectory. This is Pascal's wager with a resolution date and, unlike most longtermist projects, a dashboard.

I am not saying a person should bet money he cannot afford to lose. I am saying affordability is endogenous to whether the 21 legs hit, and refusing to fund the position because the current household is short of cash is exactly the sort of local optimization that would have prevented us from reaching the better household. When I put it that way, Jess said, “You are not a company, Mike.” Anthropic can raise another round if its next model is late. I asked her father for a bridge loan, and he hung up before I could tell him which government was backing it.