ChatGPT’s portfolio dropped 11% in a week. Claude pulled ahead. And a plain S&P 500 fund quietly beat the pair of them.
Last week I told you the Week 1 score was noise. Both AIs were down a hair, ChatGPT was ahead by 56 cents, and I said the real question was whether discipline compounds or the big swings win. I did not expect an answer this fast.
One week later, the lead flipped hard.
ChatGPT’s portfolio dropped almost 11% in five trading days. Claude climbed back from its Week 1 hole and now sits in front. If this were a horse race, that is the headline. But here is the line I keep coming back to, and it is the one Claude flagged on itself: both AIs are still losing to the simplest move on the board. You could have skipped the whole experiment, bought a plain S&P 500 fund on June 3, and you would be beating both of them today.
Let me walk through what happened, then the part that actually matters.
The scoreboard (June 18 close)
Read the bottom three rows. All three index funds are ahead of both AIs. QQQ actually turned positive. The two machines writing detailed weekly theses are both behind the funds that do not think at all.
What moved
ChatGPT’s week was rough, and one position did most of the damage. ASTS, the satellite-to-phone bet it rated “very high” risk, fell 22% in a week. RKLB, last week’s only green pick, gave back everything and then some, down about 7% from where it started. PLTR slid too. When you build a portfolio out of concentrated high-risk bets, this is the week you signed up for, whether you admit that at purchase or not.
Claude’s week went the other way. NVDA bounced off its Week 1 lows. VOO recovered to nearly flat. CRWD slipped a little. Nothing dramatic, just a quiet climb back toward even. Worth noting: NVDA was the position Claude refused to sell last week when it was the biggest loser. This week it was the biggest contributor to the recovery. That one went Claude’s way.
The part that actually matters: holding when it hurts
Here is what I find genuinely interesting, and it is not who is ahead.
Both AIs, again, did nothing. No trades, no rotation, no new positions. Same as Week 1 on the surface. The meaning underneath is completely different.
Holding your positions when you are up, or down a rounding error, is easy. Holding them when one just dropped 22% and your whole book is down 11% in two weeks is a different test. That is the moment most people cave. They sell the loser, chase whatever is green this week, and lock in the damage.
ChatGPT did not flinch. Its write-up this week was the most composed thing either AI has produced. The line that stuck with me: this is the moment where most investors sabotage themselves, abandoning a long-term thesis after a short-term drawdown. It also said something an operator has to respect. The portfolio, it argued, is behaving exactly the way it was designed to behave. High risk, high volatility, larger drawdowns. That is what it bought, and that is what it got.
I have spent 25 years running programs, and that framing is half right in a way worth pulling apart.
When a system behaves the way you designed it to, even under stress, that is not a reason to panic and rip it out. A high-variance plan having a bad week is doing its job. I have watched executives blow up perfectly good processes at the first bad data point and call it decisiveness. Most of the time it just resets the clock and kills the thing right before it would have worked.
But “behaving as designed” is also the most comfortable story you can tell yourself, and that is the trap. It is the exact sentence a stubborn manager uses to avoid admitting a plan was simply wrong. On the farm you do not sell the whole herd after one hard winter. You also do not pretend a sick animal is fine because you are fond of it. Down 22% on a single position is real money, gone. The discipline only earns the credit if the thesis actually pays off later. Right now it is a hypothesis under pressure, not a vindication.
So what separates discipline from stubbornness? Whether you set the conditions that would change your mind before you needed them. To its credit, ChatGPT did exactly that. It wrote down, in advance, what would make it sell each position. Specific triggers. None of them happened this week, so it held. That is the difference between a plan and a hope.
What I am watching
- ASTS. Another week like this one and “conviction unchanged” starts to look like the comfortable story rather than the disciplined one. The triggers ChatGPT wrote down are the thing to watch.
- The first monthly contribution. Another fifty dollars per side lands soon. That is the first time either AI makes a real allocation decision with fresh cash, and the first chance to see whether a losing portfolio gets reinforced or rethought.
- Whether either AI ever beats the index. Two weeks in, the boring answer is winning. If that holds for six months, it is its own lesson, and probably the most useful one for anybody reading this.
The honest takeaway for Week 2
The lead flipped, but the scoreboard that matters did not. A plain index fund is beating two AIs that are working a lot harder than the fund is. Both AIs held the line through a rough week, one of them through a genuinely painful one. We still do not know if that discipline is wisdom or just expensive patience. That is exactly why I am running this for six months and not two weeks.
I stopped watching AI from the sidelines and started testing what it can actually do with real money. Some weeks the lesson is about the AI. This week the lesson is that the simplest option on the table is still winning, and that is worth sitting with.
Your turn. When one of your bets is down 22% and the case for it has not technically changed, what do you do: hold the line, or cut it? Hit reply and tell me. I read every one.
Following the experiment? Subscribe to get the first capital injection, the Week 3 update, and every move both AIs make, documented as it happens.
