In one sentence: Phil and Danielle try Google's Bard, watch it invent facts about Phil's past, consider how AI might help investors screen for companies, and conclude that the circle of competence is still the guard against trusting a machine you can't check.
Key ideas
- Why now. ChatGPT (Microsoft and OpenAI) prompted a "red alert" at Google, and Bard shipped early. [01:00–03:00]
- Made-up facts. Phil asked Bard for a photo of Jonas Salk and himself at West Bridge. Bard first said Salk was never there. After Phil pushed back it agreed and then invented a wife, Jenny, and a wrong description of the fund. Phil compares it to a helper who must give directions. [06:00–11:00]
- What it says it is. Google's FAQ calls it a language model predicting reasonable text and does not promise facts. The hosts say it works for creative tasks but not for fact work. [11:00–15:00]
- Conversation, not search. The big difference from search is follow-up questions, which keep context. [05:00–07:00]
- Useful as an SEC-data fetcher. ChatGPT returned SEC data fine, though it was nothing you couldn't get elsewhere, and was out of date past 2021. [13:00–15:00]
- Where it may go. Bloomberg is working on a natural-language search. Phil describes an app (RocketTrade) that would take your interests and match companies. Coding help is another use: they mention a video where ChatGPT wrote trading backtest code, though the huge returns shown are not evidence of anything. [15:00–21:00]
- Black box risk. A screen's results are still worth reading, but you don't see what it missed. Danielle wants to know how a tool built its list. [22:00–24:00]
- A screen for good companies in a bad patch. Phil points out that many screens use recent data, so good businesses with a rough two or three years drop out. An AI-built screen could look at a past window. [23:00–25:00]
- Stay in your circle. Phil shows why with Michael Burry's picks. Burry's expertise lets him own things Phil doesn't understand. If an AI hands you ideas outside your circle, you may trust it too much. The Will Rogers line: what gets you is what you think you know that isn't so. [25:00–28:00]
How it maps to RuleOne
- This repo's agents should be treated the same way: cite filings, and check each number against the source before relying on it.
- The "good company, rough years" screen is a natural feature for the screen page: look at 10-year history and ask which names fell out of rank recently.
- 13F-based idea lists are Radar inputs only. Run the Understand step before believing any one of them.
Buffett, Munger and Graham links
- Buffett's circle of competence (1996 letter, as in 001) is the safeguard against an overconfident tool.
- Munger's "inversion": ask how the tool could lead you wrong before using it.
- Graham's Intelligent Investor ch. 1 separates investing from speculation by requiring analysis you can explain.
Words to know
- Hallucination: when a model states something false in fluent language.
- Black box: a system whose inner workings you can't inspect.
- Backtest: running a strategy on past data to see how it would have done.
Try this
Ask any chatbot a factual question about a company you know well, such as who its CEO was five years ago, and check every claim against the filings or /stock/TICKER/.
Check yourself
- What did Bard do when pushed back?
Answer
It agreed and then added made-up details, including a non-existent wife. - Why might a screen miss good companies?
Answer
It uses recent data or averages, so firms with a rough few years drop out. - Why does the circle of competence matter with AI?
Answer
Outside it you can't tell when the tool is wrong.
Short quotes
"It's the things you don't know that'll get you." (Danielle, ~27:00, auto-transcribed)