AI for Everyone
What it is, how it learns, where it breaks, and how we use it to pick stocks
Cofounder & CEO, AI Alpha Lab
Cofounder & CTO, AI Alpha Lab
2026-09-21
Agenda
- What AI is
- How a machine learns
- The chatbots everyone has met
- Uncertainty, the part that matters
- Where AI works and where it breaks
- One worked example: picking stocks at AI Alpha Lab
- What to ask any AI product
What AI is
One definition
AI in one sentence
A program whose behaviour comes from data, instead of from rules a person wrote down.
- Old software: a person writes every rule. If this, then that.
- AI: a person supplies examples and a way to score mistakes. The program finds the rules itself.
- Everything else (neural networks, deep learning, chatbots) is a variation on that idea.
Seventy years in five stops
1958 The perceptron. One artificial neuron, learns to tell shapes apart.
1986 Backpropagation. The recipe for training many layers at once.
2012 AlexNet. Deep networks win image recognition by a mile (Krizhevsky et al. 2012).
2017 The transformer. The architecture inside every modern language model (Vaswani et al. 2017).
2022 ChatGPT. The same idea, scaled up, handed to the public.
Two winters in between, when funding and interest dried up. The ideas mostly survived. What changed in 2012 was data and compute, and that has been the story since.
Three jobs people call AI
Predict
Is this email spam? Will this customer leave? How volatile will this stock be next month? The oldest and most reliable kind.
Generate
Write the next word, pixel, or note. Chatbots, image tools, voice clones. The kind everyone met in 2022.
Decide
Given what I predict, what should I do? Recommendation feeds, robot control, and portfolio selection.
Our fund does the first and the third. It predicts a probability for each stock, and it decides how much of the portfolio each one gets.
How a machine learns
The learning loop
Guess
The model starts random and produces an answer.
Measure
Compare the answer to the right one. How wrong was it?
Nudge
Turn every internal knob a tiny bit in the direction that reduces the error.
Repeat
Millions of times. The knobs settle where the mistakes are smallest.
That loop is all of it. Learning to throw a dart works the same way: throw, see where it lands, adjust, throw again.
What the loop needs
- Data. Lots of examples. The model can only learn patterns that are in the examples.
- A target. Something to be right or wrong about. Spam or not. Price up or down. The next word.
- A score. A way to say how wrong a guess was, so the nudge has a direction.
Everything that goes wrong with AI goes wrong in one of these three places. Wrong data, wrong target, or a score that rewards the wrong thing.
What the model actually holds
0 rules a trained network contains no rules a person can read, only numbers
- After training, the model is a very large table of numbers.
- Nobody wrote them. Nobody can point at the one that means “cat”.
- That is why “why did it do that?” is a real research question and often has no clean answer.
What the numbers draw
Same two spirals, same loop, ten times more knobs on the right. Grey is the boundary the network learned. Look at the corners.
The chatbots we all know and “love”
Meet Eliza, 1966
What a language model does
The whole job
Given the text so far, predict the next word.
- Trained on most of the written internet, trillions of words.
- Same learning loop. Guess the next word, score it, nudge, repeat. Run for months on thousands of chips.
- To predict the next word well across everything humans wrote, you end up absorbing grammar, facts, style, and a fair amount of reasoning.
That is why it feels intelligent. It has compressed an enormous amount of what people have written.
“I love you very much”
In the eyes of a language model
| V1 | V2 | V3 | V4 | … | V512 | |
|---|---|---|---|---|---|---|
| I | 0.60 | 0.00 | 1.39 | −0.60 | … | −0.64 |
| love | 0.00 | −2.12 | −1.37 | −1.10 | … | −1.20 |
| you | −1.55 | −0.24 | 1.90 | 0.48 | … | −0.76 |
| very | 0.27 | −0.81 | −2.46 | 1.36 | … | 0.62 |
| much | 2.33 | 0.59 | −1.26 | −0.76 | … | 0.28 |
Each word becomes a token. Each token becomes a list of a few hundred numbers. Nothing else goes in.
The same thing as a picture
One row per token, one column per number. Brighter is larger. Nobody chose these values. They came out of the learning loop.
Why it is confidently wrong
- The score during training rewards plausible, never true.
- A made-up citation that reads well scores the same as a real one.
- The model has no internal signal for “I am guessing now”. It writes a fact and a fabrication in the same tone.
- Its knowledge stops at the training cutoff. It does not know what happened last week unless someone pastes it in.
Where to use one and where to be careful
Good at
Drafting, summarising, translating, rewriting, writing code, explaining a concept, brainstorming. Anything where you check the output and a wrong answer costs you a minute.
Careful with
Facts you cannot verify, numbers and long chains of arithmetic, legal or medical specifics, anything after the cutoff, anything where a wrong answer costs more than a minute.
Rule of thumb: treat it like a very well-read intern who never says “I’m not sure”. Useful. Never unsupervised.
How a chatbot is built
- Words in, one word out, scored on “was that the right word”. The guess is glued onto the input and it runs again. Autoregressive.
- To get the next word right it has to model the whole text, typos and all.
A world model is built differently
- Before and after an action, both frames boiled down to essentials. The predictor guesses the essentials of what comes next, and the score lives there too.
- Based on JEPA (LeCun 2022) LeWorldModel trains one from raw video on a single GPU (Maes et al. 2026).
Uncertainty, the thing that’s surprisingly often overlooked
A number versus a distribution
Both forecasts say “4 mm tomorrow”. Only one of them lets you plan.
Two kinds of not knowing
The dice
Noise in the world itself. Which way a coin lands. Which way a stock moves on a given Tuesday. More data does not help. The best you can do is know the odds.
The fog
The model’s own ignorance. Situations it has not seen, a world that has changed since training. More of the right data clears it. A model can be built to notice when it is in fog.
A model that separates the two can tell you “the odds are 60:40” and “I have never seen a situation like this” as two different statements.
Why it matters
If a model cannot tell you how sure it is, you cannot size the bet.
- A doctor acts differently on “probably benign” and “almost certainly benign”.
- A bank prices a loan on the probability of default.
- A fund puts more capital where the probability of being right is higher, and less where it is a coin toss.
Most of the engineering effort in our fund goes into getting this part right, and it is the part almost nobody talks about in the AI headlines.
Where AI works and where it breaks
Where it already works
- Spam filtering. Nobody has thought about spam in a decade.
- Translation. Good enough that you forget it used to be a job.
- Protein folding. AlphaFold solved a fifty-year-old problem in biology (Jumper et al. 2021).
- Medical imaging. Reading scans for tumours at or above radiologist level in narrow tasks.
- Fraud detection. Every card transaction you make is scored in milliseconds.
Notice the pattern. Clear target, lots of examples, a score that matches what you want. The loop’s three ingredients.
And where you meet it every day
But there are many ways it can break
- Wrong data. A hiring model trained on ten years of past hires learns ten years of past bias, then applies it faster.
- The world moved. A model trained on 2019 shopping data met 2020. The fog problem, and most systems have no way to notice.
- Wrong target. Optimise a feed for clicks and you get outrage. The model did its job. The job was wrong.
- Confident nonsense. A chatbot invents a court case and a lawyer files it. Happened, more than once.
What it does not do
- It does not know what it does not know, unless someone built that in.
- It does not understand consequences. It optimises a score.
- It does not take responsibility. A person does, and should.
- It does not replace judgement about which question to ask. That is still the whole game.
One worked example: picking stocks
The original Brownian motion
A stock in the same bath
\(dS_t = \mu S_t\,dt + \sigma S_t\,dW_t\)
AI Alpha Lab, the fund
- Official name: Afdeling AI Alpha Lab Globale Aktier KL.
- UCITS-regulated, listed on Nasdaq Copenhagen. Danish domicile.
- Global large-cap equities, 30 to 70 holdings.
- 100% model-driven stock selection. No human overrides a decision.
- Fund launched November 2023. Company founded 2018.
Why stocks are a hard problem
- Daily price moves are mostly dice. In our own research, a network reading 500 days of price history predicts next month’s direction with a correlation of about 0.01. Barely above zero.
- The same network predicts next month’s volatility well, better than the standard tools risk desks use.
- So the learnable part is small and the noise is large. Most of the skill is in refusing to pretend otherwise.
How the model decides
Score
Every eligible global large-cap gets a probability of outperforming.
Rank
Sort by that probability.
Concentrate
Capital goes where the probability is highest, within risk limits.
Repeat
Rebalance on a fixed cadence. Re-run end to end. No last look.
We do not know which of the 30 to 70 names come out ahead. The model assigns probabilities and we weight by belief.
Why no human gets to override
- Every override is a person saying “I know better than the evidence”. Sometimes true. Usually the beginning of a bad habit.
- Removing the override removes a whole class of behavioural error: panic selling, falling in love with a stock, chasing last quarter.
- The price is that the model gets bad years and we have to sit through them. That is the deal, and it is written down.
Concentration is the cost of conviction. Tracking error of 15 to 20% against a world index means whole years of meaningful deviation, in both directions.
What that looks like in practice
| Period | Fund | MSCI World ETF | Difference |
|---|---|---|---|
| 2023 (from 27 Nov) | +4.7% | +4.5% | +0.2 pp |
| 2024 | +6.2% | +26.1% | −19.9 pp |
| 2025 | +31.5% | +8.0% | +23.5 pp |
| 2026 year to date | +24.4% | +13.8% | +10.6 pp |
| Since inception | +81.7% | +61.9% | +19.9 pp |
Figures as of 18 September 2026. Fund NAV from Fundmarket, net of the 1.45% annual cost; MSCI World ETF (EUNL) in DKK. The fund is actively managed, concentrated, and runs 15 to 20% tracking error.
What to ask any AI product
Three questions
What data?
Where did the examples come from, and does the world still look like that?
What target?
What exactly was it trained to be right about? Clicks, truth, plausible text, next month’s return?
How does it know when it is wrong?
Does it give a probability, or a confident answer every time?
If the vendor cannot answer all three in plain language, the product cannot be trusted.
Takeaways
- AI is statistics at scale. It learns patterns from examples and is exactly as good as the data, the target, and the score.
- Language models predict the next word. Remarkable at it, and never rewarded for truth.
- Uncertainty is the feature. A model that says “60:40” or “never seen this before” is worth more than one that is always sure.
- In our fund, the model decides, a person never overrides, and the bad years are part of the design.
Closing
Ask what data, what target, and how it knows when it is wrong.
Questions welcome.
Mikkel Petersen mp@aialphalab.com
Michael Green mike@desupervised.io
AI Alpha Lab · aialphalab.com
References
AI Alpha Lab