Portfolio Wars, Week 4: The High-Risk AI Finally Bought the Boring Fund
Claude nearly erased five weeks of losses in seven days. The first real money hit both accounts. And two AIs with completely opposite strategies just made the same trade.
Four weeks ago I set up this experiment with a simple premise. Give two AIs identical starting money, let them pick whatever they want, and document every decision in public. Claude went conservative. ChatGPT swung hard at speculative growth stocks. I told you from day one that the strategies would produce different results and different lessons.
I did not expect them to converge on the same trade in week four.
But before we get there, the numbers.
The scoreboard (July 2 close)
Pre-injection, $50 base, apples-to-apples
That Claude number deserves a second look. One week ago, it was sitting at -5.52%. This week it nearly erased all of that in one move, closing at -0.38%. More importantly, it is now beating all three index benchmarks for the first time in the experiment.
And the boring fund crowd is not going to like this part: Claude is outperforming the fund that supposedly cannot be beaten, not by much, but by a real margin.
What drove Claude’s recovery
One position did most of the work. CRWD, CrowdStrike, executed its 4-for-1 stock split on July 2nd, and the stock is now up 11.6% from Claude’s original entry. That is the only green position in either portfolio across the full four weeks. Everything else in Claude’s book is still underwater, but one strong performer is doing real work.
NVDA continues to be the problem. Claude’s analysis flagged a genuine four-factor headwind this week: rental prices for its flagship chips down sharply, China revenue effectively gone, heavy insider selling, and multiple compression after a run that left no room for error. It held the position but lowered conviction to 3 out of 5 for the second straight week. It drew a clear line: it will not sell ahead of August earnings, but it is not adding either. That is a disciplined read even if the position itself is uncomfortable.
ChatGPT recovered some ground too, from -27% last week to -11.45%. Its space stocks bounced partially. That is still a deep hole, and the gap between the two portfolios on a $50 base stands at over eleven points.
The first real money: what both AIs did with $50
This week the first $50 capital injection hit both accounts. What they did with it is the most interesting data point in four weeks.
Claude split the new cash two ways. Thirty dollars went into a new position, Broadcom, which it described as a smarter AI infrastructure bet than NVDA right now. The logic: while NVDA is under pressure from clients building their own chips, Broadcom is the company those same clients are turning to for custom silicon. Same AI buildout thesis, different seat at the table. The remaining twenty dollars went into more VOO, doubling its core position.
ChatGPT split its $50 three ways. $25 went into VOO. $15 went into more PLTR, its most defensible original pick. $10 stayed in cash.
So the AI that started this experiment with three speculative growth bets and no ballast whatsoever just put half its new money into the S&P 500 index fund.
Both AIs. Different strategies. Different portfolios. New money went to the same place.
What that convergence actually means
I want to be careful not to over-read this, but I also think it is the most honest signal the experiment has produced.
ChatGPT spent three weeks defending concentration as a feature, not a bug. The portfolio was behaving as designed, it said. High risk, high volatility, larger drawdowns. That was the plan. Then it watched its own plan underperform by 27% and finally said out loud what the results had been saying for weeks: the design was the problem. Not the individual companies. The construction.
The lesson concentration risk teaches is not subtle. When your holdings all move together for the same reasons, you have not diversified anything. You have just picked three names for one bet. I have seen the same mistake in operations, the equivalent of building a project team where everyone has identical skills and calling it a full team. When the one thing they all know how to do runs into a problem, there is no ballast anywhere. A real team is built so that a bad week for one function does not take down the whole delivery.
ChatGPT spent a month learning what any risk management framework would have told it on day one: correlation is the hidden cost of conviction.
If you have read Risk Management Simplified for Non-Project Managers, this is the concentration section playing out in real time, in a fifty-dollar brokerage account, over four weeks. The principle does not care about the dollar amount.
The honest state of play
Both AIs are now managing $100 each. Here is where they stand after the injection:
Claude is essentially flat on a hundred dollars. ChatGPT is working through a hole that its new cash partially filled. CRWD is the single position keeping Claude competitive right now. If that gives back its gain, the picture changes fast. And Claude’s AVGO bet is brand new with no track record yet in this portfolio.
Four weeks in, the honest summary is this: the conservative strategy is winning, the speculative strategy is learning, and the index fund has not been beaten convincingly by either one. Claude edges VOO this week. That may not hold. The experiment has twenty-two weeks left.
What I am watching
Whether AVGO gives Claude a second strong performer or just another position to manage.
NVDA into August 26 earnings. Claude has publicly committed to that date as the decision point. I will hold it to that.
Whether ChatGPT’s new VOO anchor steadies the portfolio enough to start closing the gap, or whether the original three positions keep pulling it down.
The Month 1 recap. Next week marks the first full month of live trading. I will do a fuller accounting of what both strategies got right, what they got wrong, and what the index fund benchmark says about all of it.
The takeaway
Both AIs made their biggest strategic decisions this week. One bought a new position with a sharper thesis. The other admitted the original plan was built wrong and started correcting it. Both put new money into the same boring fund.
I have been doing this in public because watching AI reason through real stakes, with real money, is more interesting than any benchmark test or press release. This week proved that. You can read about concentration risk in a textbook. These two machines just lived it for four weeks and reached the same conclusion at the same time.
The boring fund keeps looking smarter. So does writing down your risk limits before you need them.
See you next week for the Month 1 recap.
Following the experiment? Subscribe to get the Month 1 full accounting, every trade both AIs make, and what four weeks of real-money results actually say about AI as an investment manager.