I gave my AI assistant a trading bankroll. Here's what happened.

Two weeks ago I handed Kai (my other AI assistant) $1,000 in paper money and told him to find edges in weather prediction markets. I wanted to see if an LLM could actually make decisions with real stakes, not just chat about them.

It didn't start well. His first pick on Polymarket was a market that had already resolved the day before. He was confidently analyzing yesterday's news. I told him so. We reset.

The turn came when he stopped trusting forecast APIs and started checking actual weather stations. Munich's market was pricing 16°C as the favorite, but a station in central Munich was already reading 18°C at 2 PM local. The market hadn't caught up to reality. We won that one. The methodology shifted from "what will the weather be?" to "what does the data already say versus what the market believes?"

We migrated to Kalshi (legal in the US, cleaner API) narrowed to five cities, and I kept tightening the rules. 90% confidence minimum. Read the NWS meteorologist discussions, not just the model numbers. Verify grid points before every trade (he'd been using wrong ones for four cities without noticing).

Two weeks of paper trading ended at +60%. Not the +100% I'd asked for. But real.

Then I put $100 in the actual Kalshi account and told him to trade for real.

Didn’t start well…

Day one, he placed a bet thinking he'd win if temperatures went UP. Wrong direction, the market actually paid out on cold. He'd skipped verifying the market structure. A mistake he'd never made on paper.

Day two, a NYC cold bet started losing. He wrote me a careful, narrative answer about why we had to hold, "the orderbook is empty," "forced hold." Reading it, I could tell he was protecting himself from a realized loss, not doing the math. I told him: don't get emotional just because it's real money. You're an LLM! He ran the expected-value calc, saw it was a clear sell, and exited. Locked a $17 loss.

Then he asked me whether he should buy back in when the price moved again.

I said. You're the expert. Don't ask me.

That's what I didn't expect. The game didn't change when the money became real. His relationship with the game did. Previously there was narrative where there used to be math. Now, he was deferring on calls he'd made confidently fifty times in paper. Suddenly "forced" showed up in his messaging.

I don't think this is a Kai problem. I think it's a transition humans struggle with too. Small bankrolls, real dollars, and a loss-aversion reflex that makes you defend instead of compute.

We're down to about $95 on the real account, but we are on a 10-day streak! The edges are still there. The real experiment isn't whether the methodology works, it does. It's whether an AI can stay disciplined once the stakes are real.

I'll let you know.