AI for Everyone
What it is, how it learns, where it breaks, and how we use it to pick stocks
Kunstig intelligens til udvælgelse af aktier
IDA Vestjylland and Fagteknik hosted this evening at Aarhus Universitet Herning on 7 October 2026. IDA is the Danish Society of Engineers, so the room was engineers rather than finance people, which is a good audience for a talk that spends most of its time on the mathematics.
The evening was split. Mikkel Petersen, cofounder and CEO of AI Alpha Lab, presented the fund itself from his own slides. These are mine, and they cover the AI: what it is, how it learns, and where the learning goes wrong.
What the talk covers
It opens with a working definition: a program whose behaviour comes from data, instead of from rules a person wrote down. From there the talk builds the learning loop, guess, measure, nudge, repeat, and what it needs to work, which is a target, examples, and a score that matches what you actually want. Every failure later in the talk traces back to one of those three.
The middle is language models, because that is the AI everybody has met. There is a live Eliza from 1966 to start, then tokens and embeddings for “I love you very much”, a heatmap of those embeddings, the same treatment applied to a whole Shakespeare poem, and word arithmetic where Queen falls out as King plus Woman minus Man. Then two architectures side by side: the autoregressive loop that produces a chatbot, and a JEPA world model that predicts in latent space instead. One predicts the next word, the other predicts what happens next, and only one of them can be asked what if.
Then uncertainty, which is the part I care most about. Two rain forecasts with the same headline number and very different distributions, the split between noise in the world and the model’s own ignorance, and why a model that cannot tell you how sure it is cannot be used to size a bet.
The worked example is a stock. A live geometric Brownian motion simulation shows why direction is close to unpredictable while volatility is forecastable, which is the whole reason a probabilistic selection model is built the way it is.
The deck carries a backup section at the end, holding the AI Alpha Lab fund slides and a longer set of examples of where AI already works and where it breaks.
The slides include a live Eliza, a physics simulation, and a word-embedding video, so they are best viewed in a browser rather than printed.
AI Alpha Lab (Afdeling AI Alpha Lab Globale Aktier KL) is a UCITS-regulated investment fund. This communication is provided for informational purposes and does not constitute personalised investment advice. Past performance is not indicative of future results. Equity investments can lose value, and tracking error of approximately 15-20% means short-run deviations from any benchmark are expected. For personalised advice, please consult a licensed financial advisor.