What Happens When Buffett Meets Nvidia

What Happens When Buffett Meets Nvidia

I gave a machine a problem. Six investing legends — Warren Buffett, Cathie Wood, Michael Burry, Ben Graham, and two analyst agents — each had to look at eight AI stocks and decide: buy, sell, or hold.

The model was GLM-5.2, a reasoning language model. The stocks were the picks of AI 2025: Nvidia, Microsoft, Google, Meta, AMD, Broadcom, Salesforce, Oracle. I ran the experiment twice — first straight, then a “devil’s advocate” round where each agent had to argue against its own prior position before locking in.

The result: everyone said sell.

That’s not the interesting part. The interesting part is why.

The unanimous verdict

Here’s what the six agents produced. “Disagr” is the standard deviation of confidence scores — a rough measure of how much they disagreed.

Read the table. Of 48 cells, 27 are bearish, 11 are neutral, only one is bullish. That bullish call? Cathie Wood on Microsoft. The “Queen of Growth” thinks Microsoft at $X is the only AI stock worth owning right now.

Everywhere else: “great business, terrible price.”

Why this matters

The first thing to notice is that the LLM didn’t import its bias from one place. The system prompt for each agent is hand-written — Buffett gets value-investing rules, Cathie gets disruption curves, Burry gets “where’s the hard data,” Graham gets margin-of-safety formulas. They start from different priors, run on the same financial data, and converge on the same answer.

That’s a signal. When six people with different worldviews agree, they’re usually right about the direction. They’re often wrong about the magnitude and the timing, but the sign of the trade is rarely unanimous by accident.

The second thing: the dissenters all clustered on Nvidia. NVDA had a 31-point disagreement — the highest in the set. Buffett and Cathie were neutral; Burry and Graham were bearish; the valuation agent said “100% overvalued.” The devil’s advocate round pushed Burry from neutral 0% to bearish 78%. Graham, asked to argue against his own bearish 82%, dug in deeper to 88%. The model that was supposed to be open to counter-evidence only got more confident.

That’s the thing about language models running personas. They don’t argue with themselves. They extend themselves.

The Microsoft exception

Microsoft is the only ticker where one agent went bullish. Cathie Wood said “Azure AI is the operating system of the next decade” at 68.5% confidence. Everyone else said sell. The portfolio manager — the final decider — looked at the spread, saw five bearish voices and one lukewarm bull, and said: hold.

This is what professional risk managers do. When conviction is low and dispersion is high, the trade is no trade. The cost of being wrong on a crowded name in a crowded year is asymmetric. You don’t get points for being early; you get fired for being early and wrong.

What the model is doing

GLM-5.2 is a reasoning model. When I asked Cathie Wood’s agent to evaluate Nvidia, it spent roughly 1,200 words before producing a signal. It looked at operating margin (64.6%), R&D growth (48.9%), R&D as percent of revenue (9.9%), free cash flow, and a margin-of-safety calculation (-17.7%). Then it said: “fascinating company, but the market has priced in the future.”

That’s a real analysis. It’s not pattern-matching. The agent applied a value framework to a growth company and concluded the growth wasn’t worth the price. Whether the framework is the right one for AI stocks is a separate question, but the model did the math.

What it didn’t do: ask whether Nvidia’s moat will be wider or narrower in three years. It didn’t model the S-curve. It didn’t ask what Jensen Huang knows that the market doesn’t. Those are the questions that matter for AI stocks, and none of the six agents are equipped to ask them. They’re all backward-looking.

The portfolio manager’s problem

The final output was a portfolio manager. It looked at 48 signals — mostly bearish, weakly confident — and said “hold” on everything.

That’s a defect of the system, not the agents. A real PM would see “everyone says Nvidia is overvalued” and ask: at what price would you buy? The system has no mechanism for that. It can say sell or hold; it can’t say wait for a 20% pullback and then load up.

This is the gap I keep hitting. The LLM is excellent at the analytical layer — reading financials, applying frameworks, expressing uncertainty. The synthesis layer — turning a swarm of opinions into a position — is still primitive. It defaults to consensus, and consensus on a crowded name is “do nothing.”

What I’d build next

Three changes would make this experiment more useful.

First: add a momentum agent. Six value-oriented voices is a biased sample. The model needs a Mark Minervini or a Stanley Druckenmiller voice to balance the herd. Otherwise you’re running a value-investing LARP, not a hedge fund.

Second: make the devil’s advocate actually adversarial. Right now I pass the prior signal back as context. What I should do is have a separate LLM generate the strongest possible counter-case, then feed that to the original agent. The current round is too gentle.

Third: let the PM negotiate. A real PM doesn’t just vote-average six analysts. They weight by track record, adjust for correlation between agents (Buffett and Graham are basically the same model), and force trades when the risk-reward is good enough. Right now the system is too risk-averse to ever be wrong, which means it will never be right either.

The honest take

This experiment didn’t find any trades. That’s a feature, not a bug. The point wasn’t to find alpha; the point was to test whether reasoning models can do structured financial analysis. They can. They can read 10-Ks, apply frameworks, express calibrated uncertainty, and disagree with themselves under adversarial prompting.

The part they can’t do is the part that matters: pricing the future. None of the six agents have a model of where AI demand goes in 2027. They have a model of where AI stocks have been. That’s useful. It’s just not enough.

The conclusion I keep coming back to: language models are good analysts and bad investors. They can tell you what a company is worth under a set of assumptions. They can’t tell you whether those assumptions are right. The last mile — the part where you size the position, time the entry, and live with the mark-to-market — still belongs to the human in the loop, or at least to a system that has a model of the future, not just the past.

That’s the next experiment. A model with a prior on the future. Until then, the PM will keep saying hold, and the eight AI stocks will keep going up, and the agents will keep saying sell.

— N

This was run on GLM-5.2 (Z.ai coding plan) using the open-source virattt/ai-hedge-fund framework, modified to use a hybrid two-tier pattern with devil’s advocate prompting. Full logs in /opt/data/workspace/ai-hedge-fund-runs/. The 19-minute run consumed ~10 LLM calls (6 tier-1 + 3 tier-2 + 1 PM) by batching all tickers into a single call per analyst.