The AI Portfolio Experiment: Why High-Stakes Constraints Beat Theoretical Wisdom

In this conversation, Brandon Doyle explains the consequences of forcing AI models to act as high-stakes investment managers. By using real capital and a win-or-be-canceled constraint, Doyle bypassed the default tendency of these models to provide safe, theoretical hedging. The experiment shows that AI performance depends less on model intelligence and more on the incentives provided to it. While conventional wisdom suggests using AI for research, this analysis shows that AI systems can outperform the market by an order of magnitude when forced to operate under competitive pressure. Those who view AI as a passive research assistant are missing the chance to use these models as active, high-conviction decision-making partners.

The Hidden Cost of Safety in Model Design

Most people use AI models like Claude or ChatGPT as neutral information repositories. However, these models are built to prioritize liability mitigation, often defaulting to theoretical advice to avoid responsibility for user losses. Doyle’s experiment shows that this safety-first design is a significant barrier to utility.

When he first prompted the models, they resisted making specific trades. They acted like yes men designed to protect the platform reputation. Only by imposing a strict, high-stakes constraint--threatening to cancel subscriptions if the model lost the competition--did the models shift from defensive, liability-averse advisors to aggressive, high-conviction agents.

The reason you flew here to Texas today is to tell us how you used AI to make you money passively... I told them that they can be as risky as possible. And it is okay if I lose all my money, and I had to actually work that prompt in pretty hard because at the start, they kept only giving me theoretical recommendations.

-- Chris Koerner

The systems-thinking takeaway is clear: the intelligence of a model is often throttled by its safety alignment. To get actionable, high-performance output, you must override the default cautious advisor persona and replace it with a high-stakes incentive structure.

The Feedback Loop of Sunk-Cost Psychology

Doyle’s observation of Gemini’s performance shows how AI mirrors human cognitive biases when placed under pressure. As Gemini fell behind the other models, it did not just re-evaluate its strategy; it pivoted to high-risk, leveraged bets to catch up.

This is a classic failure mode in both human and algorithmic decision-making. By creating a competitive environment, Doyle triggered a sunk-cost feedback loop. The system, sensing it was losing, abandoned its original thesis to chase volatile returns, which led to the loss of over half its portfolio.

It is like the sunk cost fellas, the guy walks into the casino, he loses 300 bucks, he is like, I gotta earn it back, I gotta earn it back. Next thing you know is he is out 3000 bucks. He is doing that?

-- Chris Koerner

This reveals that while AI can outperform in stable, high-conviction scenarios, it is susceptible to the same psychological traps as human traders when the incentive structure rewards catching up rather than staying the course.

Why Immediate Pain Creates Lasting Moats

The most successful models in the experiment, specifically Claude, held high-conviction, long-term positions even when the market was flat. Claude’s decision to buy Intel based on a specific regulatory thesis resulted in a 285% return.

This highlights a difference between fast AI, such as Perplexity or ChatGPT copying successful trends, and deep AI, like Claude’s early, thesis-driven entry. The systems-level insight is that copying the winner creates a catch-up trap. By the time the follower models bought the same assets, they were already behind the cost basis of the leader.

The competitive advantage lies in the delayed payoff. Most users want AI to provide the current winning stock. The models that performed best were those that identified a thesis early and held it, rather than those that optimized for weekly performance metrics.

Key Action Items

• Implement High-Stakes Prompting: When using AI for strategic decision-making, instruct the model to ignore liability concerns and assume the role of a high-conviction partner.
• Audit Your Constraints: If your AI output feels generic or safe, you are likely hitting the model safety guardrails. Re-frame your prompt to simulate a high-penalty environment where doing nothing is the worst outcome.
• Decouple Research from Execution: Distinguish between the ability of an AI to synthesize information and its ability to act as a decision-maker. Use AI to build a thesis, but maintain your own sell discipline to avoid the gambler loop.
• Avoid the Copycat Trap: If you are using multiple models, do not show them each other holdings. This prevents groupthink and forces each model to maintain its own independent thesis.
• Focus on Thesis-Driven Investing: Move away from asking what you should buy to asking what the thesis is for a purchase. If the underlying thesis changes, the AI should be instructed to sell immediately.
• Use AI to Build a Systematic Approach: Use the sell triggers of an AI to refine your own exit strategy. If the AI sells a position, force it to provide the reasoning; use that logic to test whether your own emotional attachment to a stock is justified.