A veteran analyst of our acquaintance recently ran an experiment that should unsettle anyone whose paycheck depends on stock analysis. He asked four leading LLMs whether Microsoft is a buy, using a two-step prompt that first requested technical chart analysis and then layered in fundamentals, earnings, and margins. ChatGPT, Grok, Claude, and Gemini all said BUY. The entire exercise took minutes, cost nothing, and produced reports containing information that even a seasoned reader of equity research found new. His conclusion was blunt: the job of equity analyst is dead. A fund that once needed ten or twenty specialists can now run with two or three generalists.
He is probably right about the headcount. But this observation opens a much more interesting question: if AI can do the analytical work, what exactly is left for the humans to do? The answer turns out to be more surprising — and more ancient — than you might expect.
The Weighing Machine and the Voting Machine
The value of any asset comes down to two things: fundamentals and psychology. Buffett's famous formulation — that the market is a voting machine in the short run and a weighing machine in the long run — captures this neatly, but the clean separation understates how entangled the two really are. Psychology can trump fundamentals not only in the short run but in the long run too, when the psychology itself changes the fundamentals. A stock that stays irrationally elevated long enough enables a company to raise cheap capital, acquire competitors, and recruit talent — transforming narrative into reality.
AI is extraordinarily good at the weighing machine. It can synthesize financial statements, model earnings trajectories, compute relative valuations, and scan thousands of documents for material signals, all faster and more reliably than any team of human analysts. When the LLMs in our acquaintance's experiment were polled, they were doing weighing-machine work. They did it well. But the voting machine — the psychology — is a different matter entirely. It's not just that AI can't predict market psychology. Humans can't either, not reliably. The difference is that the best human investors develop an intuitive sense for crowd dynamics that isn't reducible to pattern-matching on historical data. This isn't mysticism; it's closer to the way some marketers just see what will resonate before the focus groups confirm it. Business schools teach the frameworks — market surveys, segmentation, the 4Ps — and these are genuinely useful, just as fundamental analysis is useful. But it's undeniable that some practitioners consistently outperform others in ways that can't be fully explained by the frameworks. It's not purely random. Something else is going on.
The Analyst Layer Is Dissolving
Whatever that something else is, it does not reside in the equity analyst's job description. What the experiment demonstrated is that the marginal value of a human equity analyst is collapsing toward zero for the work that constitutes the bulk of their output: synthesizing public data, running standard valuation models, and producing recommendations that cluster around consensus.
We shared this analysis with Andrew Parlin, a fund manager at Washington Peak Investment Advisors. His response was direct: "I, personally, use AI to do what it used to take 100 analysts to do." He confirmed what the data has been saying for years — active managers have gone from running circles around broad indices in the 1980s and 1990s to trailing them, often by big margins, over the past fifteen years. The vast majority, he agreed, cannot justify their fees.
Today's agentic AI systems chain together multi-step research workflows — technical charts, SEC filings, earnings transcripts, news sentiment, macro data — in automated pipelines that run continuously. The two-prompt exercise described above is already quaint. A properly configured agent could execute that analysis across an entire portfolio, updating in real time. The analytical grunt work that once required floors of junior analysts is being commoditized at breathtaking speed.
The Contrarian Trap
Andrew identified two edges that still matter in his practice. The first is short selling, where the research infrastructure has largely evaporated and the reflexive instinct of sell-side analysts — and of any LLM trained on their output — is to recommend buying the dip. Finding cases where a setback is structural rather than temporary requires pattern recognition that runs against the grain of the training data. The second is macro regime recognition: Andrew posed the sharp question of whether AI would have flagged the subprime crisis in 2008, and suspects the next major dislocation — possibly in U.S. sovereign debt — will similarly lack precedent in any training set.
Both edges involve contrarian thinking, and it's tempting to conclude that contrarianism is the fund manager's enduring advantage over AI. But we should be honest about how narrow that edge is. As Bezos reportedly said about Peter Thiel: contrarians are usually wrong. For every investor who correctly identified Peloton's terminal decline, dozens made equally contrarian bets that simply lost money.
Worse, contrarian skill — even when genuine — is self-defeating at scale. Tetlock's research on superforecasters confirms that forecasting skill exists and is persistent. But if you could bottle that skill into a perfect superforecaster AI and distribute it to every hedge fund, you'd have an index fund with extra steps. The alpha evaporates precisely because everyone acts on the same signal. This isn't a limitation of AI specifically. It's a property of markets.
Crises Aren't Predictions — They're Decision Chains
There's a deeper problem with the "predict the next crisis" framing, whether applied to AI or humans. Crises are not single events to be predicted. They are cascading chains of decisions, any of which could reverse or amplify the preceding ones.
Smoot-Hawley gets blamed for the Great Depression, but the Depression was made by a long series of follow-on decisions by the Fed, by Congress, by trading partners — choices that deepened what could have been a severe but recoverable downturn. The Iraq invasion is called a mistake, but the invasion itself was a military success; it was the occupation-phase decision-making that determined whether the outcome would be catastrophe or transformation. The subprime crisis wasn't simply "housing prices declined." It was Bernanke choosing to rescue Bear Stearns but not Lehman, Congress debating TARP while markets cratered, and a hundred smaller decisions by executives and regulators at each node in the cascade. A different Fed chair, a different Congress, a different political moment, and the dominoes fall entirely differently.
This matters for the AI question because even perfect identification of initial conditions tells you almost nothing about outcomes. The outcome depends on which specific humans are in which specific chairs, operating under which specific pressures, when the stress arrives. An AI trained on the 2008 crisis would learn that the Fed intervenes aggressively — but only because the 2008 Fed was led by a scholar of the Great Depression. That's not a generalizable pattern. It's a biographical accident.
The Information That Isn't in the Reports
So if analytical skill is being commoditized, and contrarianism is self-defeating at scale, and crisis prediction is inherently unmodelable — what's actually left?
Here we arrive at something the finance industry rarely discusses honestly. The real edge possessed by the best investors was never primarily analytical. It was social. It was what the Japanese, borrowing from the philosopher Michael Polanyi, call tacit knowledge (暗黙知: anmoku chi) — the information you absorb by being physically present, reading body language, sensing the energy in a room, catching the micro-hesitation when a CEO answers a question about margins.
When a seasoned analyst, during the years of buy-side fieldwork, travels to visit companies in person, the point is not to duplicate the work of reading the 10-K. It is to gather a completely different information stream — one that was never digitized because it was never explicit in the first place. Does the CFO make eye contact when discussing receivables? Is the factory floor genuinely busy or staged for visitors? Do the mid-level employees seem energized or beaten down? This is real information with real predictive value. It just doesn't exist in any database.
The great macro investors may not have been better analysts. They may have been better readers of people and situations. The gregarious fund manager who accumulates a dense network of informal signals — who picks up at Davos that a certain finance minister looks nervous, who hears an offhand detail over dinner about credit conditions in an emerging market — is operating on information that no LLM can access. Not because the technology isn't advanced enough, but because the information was never recorded.
We should be clear-eyed about the limits of this kind of edge. Good investor relations teams are skilled at constructing Potemkin villages for visiting analysts, and the ability to deceive scales with the ability to detect deception. Social intelligence is messy, subject to its own biases, and impossible to verify statistically. But it is real in a way that is categorically distinct from what AI can access.
The Implementation Genius
Andrew pushed us further with an observation drawn from Sebastian Mallaby's More Money Than God, a book that, more than any other, illuminates how the great investors actually operated. The key insight: their edge was as much in implementation as in market insight.
George Soros's hit ratio, Andrew notes, was good but not great. What set him apart was that he capitalized massively on his winners and was eerily swift to cut losers. Paul Tudor Jones was, by Andrew's frank assessment, a "fairly mediocre thinker." But he made a fortune by spotting a few big trends and riding them with enormous leverage, putting on trades again and again, absorbing small losses repeatedly, until one worked — and then aggressively feeding the winner for staggering gains that dwarfed the accumulated nicks.
This is a genuinely distinct category of edge, separate from both information and analysis. Certain personality structures seem to confer an ability to make highly selective bets and ride them very hard. It takes a specific kind of courage — or, more precisely, a specific perception of risk.
There's a well-known finding in entrepreneurship research with a counterintuitive conclusion: successful entrepreneurs tend to be more risk-averse than the general population. The reason is that they don't perceive what they're doing as risky. The opportunity is so clear to them that not acting would be the greater risk. We suspect the same applies to the Soros types. What looks like insane leverage to the rest of us feels as obvious and safe as a Treasury bill to them. They're not braver; they're more certain, because they're integrating all of that tacit, social, embodied information into a felt conviction that the analytical framework alone cannot produce.
Could you train an AI to replicate this? You could certainly optimize position-sizing and loss-cutting rules through reinforcement learning. But the crucial feature of these investors is that their implementation style is inseparable from their perception. They cut losses fast not because they follow a rule, but because they feel the wrongness of the position. They size up not because a model tells them to, but because their conviction is genuinely high. An RL agent would learn these as rules — brittle, backward-looking, and vulnerable to regime changes. The human investors are doing something more like continuous real-time calibration of conviction against incoming signals, much of it drawn from the tacit and social information streams that no model can access.
The Uncomfortable Consensus
All four of the LLMs in the experiment said BUY. To any experienced investor, unanimous agreement is not confirmation — it's a signal to be cautious. If every AI system is trained on the same public data and produces the same recommendation, the recommendation contains no alpha. It is consensus, repackaged in a faster and cheaper wrapper.
This is not a temporary limitation that will be solved by better models. It is structural. LLMs are trained to predict the most probable next token given existing text. Applied to financial analysis, this means they converge on whatever the preponderance of existing analysis suggests. The weighing machine gets very precise. But precision on the weighing-machine side tells you nothing about where the voting machine is headed next — and it is the voting machine that creates the dislocations where real money is made and lost.
What Remains
Alpha was always scarce, fragile, and self-defeating at scale. AI doesn't eliminate it — it reveals that alpha was never what the fund management industry was primarily selling. What the industry sold was the appearance of analytical rigor justifying fees. AI strips that away completely. When a two-minute prompt produces what a team of twenty analysts once delivered, the justification for 1% annual fees on active products that underperform year after year — and Andrew expressed genuine amazement that firms like BlackRock and Fidelity still get away with this — simply collapses.
What survives, if anything does, operates on three layers that AI cannot reach. First, tacit social intelligence: the information gathered not from databases but from being in the room, reading people, accumulating informal signals through networks that took decades to build. Second, regime recognition amid cascading human decisions: the ability to read an unfolding crisis not as a data pattern but as a chain of choices by specific people under specific pressures. And third, implementation personality: the rare temperament that perceives opportunity where others see risk, and acts on that perception with a conviction that no model can simulate because it arises from the integration of all the information — quantitative, social, embodied — that the investor has absorbed.
Buffett was right all along. Most investors should buy the index. The few who shouldn't are the ones playing a different game entirely — one that no machine, however intelligent, can join. Not because the machine isn't smart enough, but because you have to be in the room, you have to read the people, and you have to feel it in your bones. Those are the oldest human skills there are, and they may be the last ones standing.