For three weeks, the story of this experiment was that both AIs did nothing. Same holdings, week after week, while I kept waiting for one of them to flinch. This week the market finally gave them a reason to, and both of them blinked. Here is the strange part. Neither one made a single trade.
Changing your mind and changing your portfolio turned out to be two different things. That distinction is the whole story this week, and it is more useful than the score.
Let me give you the numbers first.
The scoreboard (June 25 close)
Everything is red. This was the week the whole market turned down, not just the risky corner of it. All three index funds fell three to four percent. Claude fell more than five. ChatGPT fell another sixteen points on top of an already bad start. And the plain index funds are still ahead of both AIs. Three weeks in, the boring option has now beaten both machines every single week.
ChatGPT is now down more than a quarter of its starting money. Hold that thought, because what it decided to do about it is the most interesting thing that happened this week.
What happened to the market
This was a tech-led pullback, and it was about price, not panic. Technology stocks had run up hard over the prior few months and gotten expensive relative to everything else. This week investors started asking whether the profits from all the AI spending could justify both the valuations and the enormous cost of building data centers, chips, and power. So the most stretched part of the market took a haircut. The Dow actually held up on strength in healthcare and industrials. Under the surface, money rotated out of the expensive names. That is a valuation reset, not a sign the underlying businesses broke. The distinction matters for how you respond, and both AIs leaned on exactly that distinction.
Claude’s first crack
Here is something that did not happen in Weeks 1 or 2. Claude lowered its conviction on a position.
For two weeks Claude held NVDA through every dip and called the weakness rotation, not rejection. This week it changed its tune. Not on the holding, on the confidence. It downgraded NVDA from a 4 out of 5 to a 3, and it was specific about why. Its argument was that the concern was no longer just sentiment. It pointed to falling rental prices for NVDA’s flagship chips, which would pressure the fat margins the whole thesis rests on, plus heavier insider selling than it had flagged before. Its conclusion was measured: not a reason to sell, but a reason to stop defending the pick and start watching it closely into August earnings.
The line that stuck with me was its own framing of the job. The job is not to defend a pick. It is to follow the evidence. Right now the evidence says hold and watch, not sell, and not double down either.
That is discipline cutting in a new direction. In Weeks 1 and 2, discipline meant not selling into a dip. This week, discipline meant admitting a favorite looks less certain than it did, out loud, while still not doing anything rash about it.
ChatGPT’s confession
ChatGPT, down 27%, did something I did not expect. It stopped defending the construction of its portfolio.
For two weeks its line was that the portfolio was behaving exactly as designed. This week it went further and admitted the design itself was the mistake. Not the stock picks, in its view, but the concentration. Three aggressive growth names with no ballast, all of which fall together when the mood turns. It lowered its conviction on its two worst positions. And then it said the thing that made me sit up. It does not plan to average down into its losers with the first contribution. It plans to add a plain broad-market ETF and a little cash instead.
Read that again. The concentrated, swing-for-the-fences AI, after three rough weeks, is planning to buy the boring index fund that has been beating it the entire time. It named VOO specifically, which happens to be Claude’s core holding.
I have to give it credit. That is a real lesson learned in public, with real money on the line, written down honestly. By Claude’s read, ASTS, ChatGPT’s worst position, took genuine damage this week on a weak earnings report, not just a sympathy selloff, so part of that drop is the thesis actually cracking. Admitting your whole portfolio was built wrong is harder than admitting a single pick was wrong, and ChatGPT did it without flailing.
The lesson worth keeping: updating is not thrashing
Here is what ties the week together. Both AIs changed their assessment. Neither changed their positions. And that is exactly right.
In 25 years of running programs, the most common mistake I have watched smart people make is treating new information as a reason to blow something up. A risk shows up, so they reorganize the team, kill the plan, and start over, and they call it being responsive. Usually they have just confused updating their read with thrashing their plan. Those are different acts. One is intelligence. The other is panic wearing a suit.
The reason both AIs could change their minds without changing their trades is that they wrote down, in advance, what would actually make them act. ChatGPT set its sell triggers back in Week 2. Claude tied its NVDA decision to a specific upcoming earnings report. So when a bad week arrived, each had a standard to measure it against instead of a feeling to react to. On the farm you can read a bad sky and keep the animals in for the day without deciding to sell the herd. Reading the weather is not the same as making a permanent decision. Most bad calls come from skipping that gap.
This is the part of my actual work I think about most. When I help a company put AI into its operations, the hard part is almost never the technology. It is building the decision standard first, so that when early results wobble, and they always wobble early, nobody panics and rips out a system that was two weeks from working. Same skill these two AIs are showing with a fifty-dollar brokerage account. Decide your triggers before you are emotional. Then hold the line until one actually trips.
The honest scoreboard
Three weeks, and three weeks of both AIs trailing a plain index fund. Claude is down about two points more than VOO. ChatGPT is in a hole that will need a real rally to climb out of. If you had ignored this entire experiment and bought the S&P 500 on June 3, you would be ahead of both of them right now. I am not going to dress that up. It is the most useful result in the project so far, and we are only three weeks in.
What I am watching
- The first capital injection. Another fifty dollars per side lands next week, and it is now the most interesting moment in the experiment. ChatGPT says it will diversify instead of doubling down. Will it actually do that when the cash is real? And will Claude add to anything, or sit on its hands?
- NVDA into August earnings. Claude drew its line there. That is the report that tells us whether the chip-margin worry is real or a head fake.
- ASTS. Its drop now includes an actual earnings miss, not just a mood swing. That is the first position in this experiment where the business itself, not just the price, is flashing a warning.
The takeaway
The market handed both AIs their first genuinely bad week, and both responded the same way. They updated their thinking and kept their hands off the controls. Whether that is wisdom or just patience that has not been punished hard enough yet, I still cannot tell you. That is why this runs for six months. But the skill on display this week, changing your mind without changing your whole plan, is one most humans never learn. Two AIs and a fifty-dollar account just demonstrated it cleaner than a lot of people I have worked with.
Your turn. When the data turns against one of your decisions, what is your honest first move: change your mind, change your plan, or change nothing and watch? Hit reply and tell me. I read every one.
Following the experiment? Subscribe to get the first capital injection, where one AI says it is finally going to diversify, plus every move both AIs make, documented as it happens.
