Frank Long of Coatue presents a framework for turning investment research into a testable network of code, models, and human judgment in this 14-minute AI Engineer talk.
Give the system the right context
- Model the investment firm as a function: external and internal information go in; what to buy, how much, and when come out.
- Connect news, prices, filings, current positions, risk limits, and institutional notes through programmatic interfaces.
- Long argues that model intelligence has limited value when the agent cannot access the information its human colleagues use.
Break decisions into inspectable steps
- Start with gathering relevant evidence, analyzing its meaning, and deciding how to act.
- Korean-language semiconductor news illustrates the smaller choices hidden inside gathering: collection, filtering, translation, extraction, and cross-checking internal views.
- Chunk size, image handling, summaries versus structured records, and the order of operations all change the resulting evidence.
Choose who or what runs each step
- Use deterministic code for exact rules and routine collection, statistics for learned relationships, and language models for tasks such as summarization.
- Keep people where judgment and real-world relationships matter, including meetings with company management.
- Combine these approaches within a step when that produces better work.
Evaluate the workflow, not its automation rate
- Treat branching logic, model choice, reasoning effort, tool-running software, skills, and hooks as choices to test.
- Define a quality target for each understandable component before optimizing it.
- Long’s prediction that investing will change more than coding remains a thesis; this talk supplies no comparative study or measured investment-return improvement.
“There is no glory in automating things for the sake of automating things.” — Frank Long, 9:07