In one sentence: Danielle's husband Nuno, who leads strategy and AI at a consulting firm, explains what AI and machine learning are, how ChatGPT-type models work and fail, and how businesses actually use prediction, as a first step toward a later episode on investing uses.
Key ideas
- Opening claim. Phil says Buffett-and-Munger-style investing has made more fortunes than any other strategy. He cites a respected investor but also notes quant firms such as Renaissance have beaten it on returns using algorithms. This is the host's claim. [01:00–04:00]
- AI is a concept. It means a machine (an "agent") doing human-like tasks. It can be rule-based (if-then, like simple chatbots) or machine learning, which learns patterns from data. [12:00–14:00]
- Training and test data. A learning algorithm produces a predictive model. Data is split into training and test sets, and the model is judged on data it hasn't seen. After training the model is static, which is why there is a cutoff date. [29:00–35:00]
- What generative models do. They predict the next word given the earlier ones, so results depend on the training text and domain. [34:00–38:00]
- Three uses so far. (1) A new way to search the web that merges steps into one answer, (2) companies paying to fine-tune a model on their own documents, (3) people automating parts of writing and research. Accuracy isn't guaranteed. [16:00–24:00]
- Not real time. A trained model always has a cutoff. Nuno says up-to-the-minute answers won't come from the model alone. [21:00–23:00]
- Google's dilemma. If answers replace search results, the ad-based search model is under pressure. The hosts speculate. Bard shows sources, which can keep ads viable. [24:00–26:00]
- Fear: emotional and rational. Nuno says people react to a machine that sounds human and to movie images. The real risks are misuse (such as malware), misinformation and bias in training data. [26:00–30:00]
- Jobs. His view is that tools mostly augment tasks and rarely replace whole roles. Phil counters with bookkeepers and farming: small gains per person still shrink headcount. Open question. [39:00–47:00]
- How companies use ML. Either to reveal patterns that inform a decision (segmenting clients, one-off) or to make recurring predictions (for example flagging a client with an 83% chance of leaving within three months). People still make the final call. [47:00–54:00]
- Blind spots. Algorithms can pick up behaviour people miss. Next episode will cover how an investor can use this. [54:00–58:00]
How it maps to RuleOne
- The "transparency versus prediction" split is useful when designing the agent stack: reports that explain a business (transparency) versus repeating alerts (events).
- Train/test discipline applies to any screen: don't tune a screen on the same history you use to judge it.
- Evaluate AI-exposed companies like any other: Understand the business, check the moat, and don't pay a story price.
Buffett, Munger and Graham links
- Buffett's circle of competence applies: if you don't understand how an AI company earns money, skip it (see 420).
- Munger on incentives and "man with a hammer" bias: tools shape how you see problems.
- Graham on analysis before purchase (The Intelligent Investor ch. 1).
Words to know
- Machine learning: algorithms that learn patterns from data instead of following hand-written rules.
- Training data: the examples a model learns from.
- Fine-tuning: further training a general model on a narrower body of text.
- Cross-validation: testing a model on data it hasn't seen to check it generalises.
Try this
Write down a stock-screen rule you use. Apply it to data from five years ago only and see which companies it would have picked, then check how they did. That is a hand-made train/test split.
Check yourself
- What is the difference between rule-based AI and machine learning?
Answer
Rule-based AI follows fixed if-then logic. Machine learning learns patterns from data. - Why do chat models have a cutoff date?
Answer
The model is trained once and then stays static. - Why use a test set?
Answer
To measure accuracy on data the model hasn't seen before deploying it.
Short quotes
"The only thing they do is predict what's the next word." (Nuno, ~35:00, auto-transcribed)