In one sentence: Phil and Danielle try ChatGPT for stock research, discuss "hallucination" (confident, invented answers) and Bloomberg's finance-tuned model, and conclude that AI can speed up background research only if you verify everything.
Key ideas
- Phil's experiment. He asked ChatGPT for small caps with growing return on equity and free cash flow plus a durable niche. It had no financial data after September 2021. [04:00–05:00]
- Hallucination. Phil describes a case where the tool listed books that did not exist, and one where a reporter's chat turned strange. These are anecdotes from news he recalls. [05:00–07:30]
- Google's caution. Phil says Google's CEO worries about convincing but false output and is holding back its own tool (Bard). [08:00–10:00]
- Possible disruption. Phil expects knowledge work, including coding, to change. [10:00–11:00]
- A back-test warning. A demo turned $10,000 into $17 million in a back-test with AI-written code; Phil says that proves nothing because the data already existed and the future doesn't. [10:30–12:00]
- BloombergGPT. A finance-trained model from Bloomberg; Phil notes it uses off-the-shelf methods on proprietary data. [11:00–15:30]
- Headline-prediction studies. Phil says two university papers found a chatbot could judge the direction of stocks from news headlines about as well as a human; both were back-tests, so treat with caution. [12:00–13:30]
- Use it for context, verify the rest. Danielle worries checking every claim removes the speed; Phil says use it as a starting point and discard what you can't verify, as with Seeking Alpha articles. [17:00–21:30]
- Test results. Asked for companies with high and growing ROIC, low debt and a durable niche, it named Alphabet, Mastercard, Intuitive Surgical, Adobe and Waste Management; Phil says all are on his wish list but only Alphabet was near a buying price. A second list of mid-caps was shown (HubSpot, Chegg, MongoDB, Trade Desk, AppFolio). These are not recommendations. [26:00–31:00]
- Phil's 26% logic. Buying at half of intrinsic value and seeing it double in three years is about 26% a year; the limit is that once price reaches value, growth is just the business growth rate. [30:00–32:00]
- Reader input. An outdoor-industry listener, Stephen Housley, wrote that forecasts for 2022–23 were cut and inventories were full, which bears on 413. [24:00–26:00]
How it maps to RuleOne
- The agent stack in this repo is a version of "AI for research": the agents draft and summarise, but every figure must trace to a filing. Treat chat output like any unverified tip.
- The /stock/TICKER/ pages give sourced numbers you can check against a chatbot's claims.
Buffett, Munger and Graham links
- Buffett's well-known skepticism of analysts' forecasts is behind Phil's point; Buffett's 2008 letter and shareholder meetings discuss forecasts and models. I'm not quoting any passage.
- Graham's Intelligent Investor (ch. 1) separates investing from speculation; short-term headline-trading tools are on the speculation side.
- Munger's "man with a hammer" (the 1995 Harvard talk on worldly wisdom) cautions against letting any tool drive every decision.
Words to know
- Large language model (LLM): software trained on lots of text to produce likely next words.
- Hallucination: a fluent output that is false or invented.
- Back-test: applying a strategy to past data; flattering because the past is known.
Try this
Ask any chatbot for a company's free cash flow for the last five years. Then check each year against the 10-K cash-flow statement or the stock page on /stock/TICKER/. Count the errors.
Check yourself
- What is a hallucination?
Answer
A confident-sounding answer the model made up. - Why is a spectacular back-test unconvincing?
Answer
It uses known past data, so it shows what would have worked, not what will. - How should an investor use a chatbot?
Answer
As a source of leads and context, with every fact verified against filings.
Short quotes
"Go use it for what it's useful for. Check everything." (Phil, ~39:00, auto-transcribed)