Nota Bene

A Few Thoughts on AI | Last updated: October 13, 2026

1. In thinking about the ramifications of AI, it helps to focus on three of its features. First, AI is alien: it never tires, it triggers an aversion to things that seem almost human, and it bears no legal liability. Second, AI is encyclopedic: it is trained on nearly everything people have written (errors included), so knowing which question to ask matters more than ever. Third, AI is generative: it extracts patterns and reproduces them with high fidelity.
2. AI tools fall into four groups.
   • Chatbots: ChatGPT, Claude, Gemini, and Grok.
   • Search and deep research: Perplexity, ChatGPT Deep Research, and Gemini Deep Research.
   • Plugins for office software: Claude for Excel, Word, and PowerPoint, ChatGPT for Excel, Microsoft Copilot, and Google Workspace AI.
   • Agents: Claude Cowork, Claude Code, and Codex.
3. There is no theory of how to use AI yet. Most of what is known came from trial and error, so the choice of what to experiment with should be deliberate.
4. Judging AI models runs into the joint hypothesis problem (Fama, 1970). The output depends on both the input and the model's ability, so a bad answer does not reveal which of the two failed.
5. AI is most useful where its output can be checked: formatting, extraction, citation checks, code with tests, and summaries of documents at hand. That is where to use it most aggressively.
6. The most common complaint about AI models is that they lack "taste." Part of the reason is that people mean very different things by taste. A model cannot meet a standard that has not been described to it.
7. AI calls take time, so the person becomes the router. While one task runs, start another in a different window: code in one, a literature search in a second, a draft in a third. As AI writes more of the output, understanding that output becomes the bottleneck (Litt, 2026).
8. Most advice on AI is about the prompt, which is a variable cost: each prompt produces one answer. Less is said about the fixed cost of getting good output. Examples include a written guide to how models, comps, and memos should be built at your desk, a library of past deals, filings, and research notes that the AI can draw on, and a description of what a good investment memo looks like. These take weeks to set up, but every later request starts from them.
9. As analysis gets cheaper to produce, taste, breadth, and judgment become more valuable.
10. A few habits make AI output more reliable.
    • Tell the model who it is acting as, what problem you are solving, what output you want, and what rules it must follow.
    • Work in durable files. Use Markdown as the default format, attach source documents, and ask for citations, page numbers, and short quotes when the output rests on evidence.
    • Keep generation and verification separate. Use AI to draft, extract, summarize, or code, then check important output with tests, the source documents, a second model, or a separate critic pass.
    • Split the work across agents: drafting, critique, editing, search, and coding. Coordinate them through clear Markdown handoff files.
    • Manage input and attention. Use voice input for long prompts, keep context small and structured, and close loops before opening new ones.
11. When everyone has the same tools and the same information, everyone gets the same answer. An edge has to come from somewhere else: different data, different questions, or different judgment.
12. For jobs that use AI, the best signal is something a candidate has built with AI and can show: a hackathon project, a side project, a tool made at work. A walk-through of that project, including where it broke and how the output was checked, reveals more than a course certificate.
13. People have always hired partly for "aura": whether someone is good to work with, can be trusted with a client, and makes the team better. AI makes this matter more, not less. As the analysis gets easier to produce, what sets people apart is what AI cannot supply: access to clients or deal flow that others do not have, a reputation for numbers that hold up, the ability to defend a recommendation in front of a skeptical investment committee, deep knowledge of one industry or market, and the willingness to start things without being asked.
14. In an ideal world, writing would be judged only on its quality. But reading takes time and effort, so readers look for cheap screens, and AI detectors have become one: a fast way to decide whether a piece deserves a careful read. Like any cheap screen, a detector makes mistakes in both directions. Careful human writing gets flagged as AI (false positives), and edited AI writing passes as human (false negatives). A screen like this can help decide what to read first, but it is a poor basis for punishing the writer. See also Mahoney (2026) on writing as a costly signal that AI makes cheap.
15. Most people building frontier AI think alike, for two reasons. The first is the race: when every single employee counts, labs hire people who already believe in the mission, and disagreement slows things down. The second is selection: the sector attracts people from a narrow set of communities, such as effective altruism and the rationalists, who arrive with similar priors. The cost appears when the consensus is wrong: few people on the inside are in a position to say so.
16. Blockchain, arguably the hype before AI, spent years rediscovering why finance looks the way it does: why custodians, clearinghouses, and know-your-customer rules exist. AI will likely go through the same exercise. Many of the problems AI agents now run into (whom to trust, how to verify work, who bears the loss when something goes wrong) are old problems in finance and economics, with a large body of work behind them.
17. The financing side of AI gets far too little attention. Most of the conversation is about models and what they can do. But data centers are being paid for with bonds, private credit, and leases that sit off balance sheet, and someone ends up holding that risk. If the payoff arrives later than expected, who holds the risk will matter more than which model is best. See also Van Nieuwerburgh (2026).
18. People reach for metaphors to make sense of what AI will do.
    • "The best metaphor for AI right now is the Western garment industry collapse, 1970 to 2000." The cost of high-quality clothing stayed about the same, while the cost of low-quality clothing fell about tenfold. People now own far more clothes, most of them mediocre. They are much better off, and they complain constantly.
    • "Mathematicians will become samurai." After firearms ended the sword's role in battle, swordsmanship survived as an art and a mark of status. Proving theorems by hand may go the same way.