In one sentence: Nuno explains how large language models, quantitative models and no-code "AutoML" platforms could change research for ordinary investors, and why the human choice of the question and the inputs still decides whether a model is useful.
Key ideas
- Language models can be fine-tuned on your own reading. Open-source models can be trained further on a private body of text, such as your own research, to give answers tailored to your area. [03:03–05:00]
- Quant firms came first, with numbers. Phil cites Renaissance's famous returns; Nuno notes they began with structured data before language models existed. "More quant" doesn't have to mean trading. [05:03–08:05]
- A three-part recipe. (1) Get the data. (2) State the problem precisely, such as "how large will this company be in 10 years?" (3) Train algorithms to answer it. [07:02–09:01]
- AutoML tools hide the coding. You choose inputs and a target and the platform tests algorithms and reports the best performer. Currently mostly aimed at businesses; consumer versions are likely later. [09:01–11:03, 40:04–42:02]
- Democratization pattern. Phil's story of Steve Jobs describing digital editing in 1988, and home video editing less than a decade later: tools move from institutions to individuals. [11:03–16:04]
- The creative step is the human one. Turning a business or investing question into a data problem, and inventing good inputs, is the valuable bit. Nuno's example: about half the predictive power in a client-attrition model came from features he engineered, such as trend and variance in behavior. [17:03, 34:00–38:04]
- Structured vs unstructured data. Numbers and categories are more mature territory for algorithms than free text. Competitors have to be modelled too: your growth depends on theirs. [19:01–26:04]
- Can AI predict a crash? Phil points to Taleb and Spitznagel, whose "insurance" approach buys puts all the time and doesn't predict. Nuno thinks machine learning could, in principle, use many more inputs and non-linear relationships. This remains speculative. [28:05–34:00]
- Nobody has a go-to investing AutoML tool yet. Phil asks for one and Nuno says it probably doesn't exist yet. [40:04–42:02]
How it maps to RuleOne
- The agent stack in this repo is a small version of the idea: let software gather and structure the data, keep the judgement (circle of competence, moat, price) with you.
- Verify any AI output against primary filings, as in 415 on hallucination.
Buffett, Munger and Graham links
- Rule #1 and Buffett rely on understanding a business and estimating a value range, not forecasting indexes. The Taleb example is the opposite stance (a hedge, not a forecast). Buffett's 2008 letter on derivatives is a useful skeptical counterpoint.
Words to know
- LLM (large language model): software trained on lots of text to predict the next word.
- Fine-tuning: further training of a model on your own text.
- AutoML: tools that train and compare models without writing code.
- Feature engineering: building new inputs from raw data (for example a trend slope).
Try this
Write one precise question you could answer with data (for example "has this company's owner earnings grown every year for 10 years?"). Open a company from All stocks and see which of the numbers you need are already on its page.
Check yourself
- What are the three parts of Nuno's recipe?
Answer
Get data, define the problem precisely, train an algorithm to answer it. - Where does the human still matter most?
Answer
Posing the problem and designing the inputs (features). Platforms automate the coding, not the creativity. - How did Taleb and Spitznagel protect against crashes, according to Phil?
Answer
By constantly buying put options as insurance, without trying to predict timing.
Short quotes
"The creativity that needs to happen to unlock the potential of AI is absolutely critical." (Nuno, ~17:30, auto-transcribed)