I Was Wrong About AI. Twice.

Let me start with an admission. The technology now reorganising my industry crossed my radar late. Not in 2017, when the foundational research was published. Not in 2020, when the first genuinely capable model appeared. It was late 2022, when a chat box did. I had spent years working with quantitative methods, close enough to this field to feel confident about it, and I still did not see it coming. Worse: once I had seen it, I went on to underestimate it twice more.

I think that sequence is worth your time. Not because the technology is interesting, though it is, but because the mistake I made is one investment professionals are structurally prone to. And because what you do after making it is, in the end, the whole game.

The ground I was standing on

For most of my career, machine learning meant something specific and well-bounded. Regularised regression, Lasso and Elastic Net, was the tried and tested tool for financial prediction, where you have many correlated features, limited data, and a real need to keep the model from overfitting. Gradient-boosted trees, XGBoost and the like, were what you reached for on structured tabular data: predicting the next song on a streaming platform, the next video in a feed. Deep learning was the heavier instrument, justified only when the dataset was enormous, deciding which ad to serve across a network with billions of daily events.

Every one of these shared a single assumption: one model per problem. You defined the task, curated the data, trained the model, deployed it. To do something new, you built something new. And generative AI, to the extent it registered with me at all, meant deepfakes at the consumer end and synthetic data generation at the serious end. A specialist tool tied to a particular kind of output. Useful, bounded, easy to file away.

That confident, working mental model of what AI was is precisely what set me up to miss what it was becoming.

The moment, and the first thing I got wrong

Then, in late 2022, I typed a question into a chat box and got back something that stopped me. It was not the novelty of a machine writing fluent prose. It was the realisation that I was interacting with a machine in plain language, and that this was a genuinely new way to retrieve and work with written information.

We started experimenting within days, front end and back end. The applications that paid off were the back-end ones, the invisible kind. At the time we were managing exposure across a large number of holdings held through investment partnerships, where the intelligence you actually need is buried in unstructured documents. Language models let us extract that intelligence at scale, in a way that simply had not been practical before.

And here is where I got it wrong the first time. I looked at what we had built and saw a friendlier query language, a more forgiving way to ask a database questions. I did not see that the same capability would, before long, call into question the way software itself is built and used.

What I had failed to register is that these models are not merely retrieving and rephrasing. Train one purely on sequences of moves from the board game Othello, with no rules, no image of the board, just notation, then look inside it, and you find it has built a working model of the board, including which pieces flip when a position is captured. It inferred the structure of the game from the statistics of the notation alone. That is not lookup. It is closer to comprehension. And I had it filed under "better search."

Wrong again, the same way

The second miss took the same shape. As the models became capable coders, work that used to take my team months, building an interface, wiring up a data pipeline, began taking days. I registered this correctly as a large efficiency gain. What I underestimated, again, was the scale of it. I saw faster feature extraction. I did not yet see that an entire knowledge industry, research, analysis, the synthesis of written information as a profession, was being reordered, not merely sped up.

Twice, then, I treated a change in kind as a change in degree. Both times I had every qualification needed to know better.

Where it actually bites

Which brings me to where this is landing in our own industry, and where I think it goes next. The effect divides cleanly along the line between discretionary and systematic management.

For discretionary managers, the change so far has been the largest, and it amounts to democratisation. A fundamental manager who never had quantitative infrastructure can now, with a capable model and a reasonable data subscription, do work that used to require a dedicated quant or a standing call to a sell-side desk: cross-asset screening, scenario analysis, comparative valuation across large security sets, parsing dense filings at speed. The technical floor has dropped. Sharp views can now be expressed with more rigour and far less lead time. Non-quant investors have, in effect, been handed a meaningful slice of quant capability.

For systematic managers, the shift is different and less mature, and it is the frontier we are working on now. Consider what quant research actually involves: read the relevant literature, take a promising method, implement it inside your own simulation and backtesting engine, test it under honest assumptions, decide whether it earns production. Every stage is slow, and the number of ideas a team can run through that pipeline in a year is a real constraint on how fast it moves. Current reasoning and code-capable models are close to handling meaningful parts of it. Not the judgment of which hypothesis is worth pursuing, nor the final call on whether a backtest is real or overfitted, but the mechanical core of reading a paper, extracting the method, and producing a first implementation a researcher can then refine. The distance between idea and first result compresses. Cycles get faster.

But the binding constraint is not the model. It is the infrastructure around it. To actually accelerate this work, three things have to be in place and properly connected: a clean, structured data environment with adequate history; a simulation engine that models execution cost, market impact and portfolio constraints honestly; and a reliable bridge from simulation to live trading that preserves the assumptions made along the way. None of these is a given. Most firms have a gap in at least one, or hold the three as systems that talk to each other badly. Model capability is now commoditised. Everyone can reach the same frontier. Infrastructure readiness is not, and that is where the difference will sit.

The real lesson isn't about AI

If there is a thread running through all of this, it is not about technology. It is about being wrong. I had every advantage in seeing this clearly, and I still misjudged the magnitude, repeatedly. That is not a comfortable thing to commit to writing. But it is the most useful thing I can offer a fellow investor.

You will be wrong. Often, and sometimes about the things you are best qualified to judge. The edge was never in being right the first time. It is in noticing the error quickly, saying so plainly, and adapting before the cost compounds. That is true of a misjudged manager, a crowded trade, a thesis that has quietly stopped working. It turned out to be just as true of the most consequential technology shift of my career.

I am on my third reading of it now. I fully expect I am still underestimating it.