A recent paper, Investing Is Compression (Stiffelman), is persuasive and well grounded in mathematics: investing is a compression problem. Reduce a vast quantity of data about a universe of stocks to a handful of decisions, and earn returns by disagreeing with the market and being right. If that's the game, the question that strikes me: what happens when compression becomes effectively free? And it has. With a good prompt and the right context, a language model can ingest and synthesize more data than any team of analysts ever could. Point a set of coordinated agents at a bounded universe of 2000 stocks and they weigh every scrap of data on every single name with extraordinary efficiency and nuance.

That displaces much of the moat in sophisticated investing and relocates it toward the top of the information funnel. A language model is a brilliant compressor and solver, not a source. Give it a thesis and a set of facts and it can identify the best portfolio far faster and at least as precisely as any human. But it cannot hand you a better view of the world, tell you something only a few people know, or sense when a human relationship matters more than the market realizes. The compression is what's being automated; the judgment of what to compress toward is not, yet. It exposes that the analysis was never the real edge, only the cost of entry. Now, the price of entry to sophisticated investing has collapsed to four hours on a weekend and $1600 of tokens.

What survives is the information going in and the worldview that reads it. Information edge comes in two flavors. One is clever synthesis of public data into a non-obvious understanding, the mosaic most fundamental funds live on, and that is exactly what models now do for free. The other is proprietary data, known before the market prices it, that no model can produce. AI commoditizes the first and leaves the second untouched, so the bar rises: deriving decisions from data better than others is no longer enough; returns will increasingly require privileged information about today, or a view of the world that is both different and right. Some understanding accrues only from being there, from years inside an industry and the rooms where decisions get made; it lives in neurons, not text, so it never reaches a model's weights. The cheaper raw intelligence gets, the more returns rely on asymmetric information and lived experience.

This is why the wave of AI-driven funds will be more paradoxical than it first appears. When everyone runs the same models over the same data, their outputs converge, and any answer a whole market reaches at once is consensus, already priced. Different mandates, horizons, and risk tolerances will still pull prices apart, but any view commoditized models derive decays in value almost as soon as it exists. Today's LLMs are not erasing investing edge; they are rapidly relocating it. Strip away the analytical labor that used to separate professionals from everyone else, and what remains is the oldest thing in the business: knowing something true before the market does, whether a fact about today, a view of tomorrow, or an intuition that cannot be modeled. The commoditization of intelligence will reveal who actually knows where the ball is going, and who was only ever good at producing analysis.