How Investors Are Using AI [Business Breakdowns: Episode 240]

Summary

• David Plon — ex-Baupost and Slate Path generalist, now founder of Portrait Analytics — argues AI should attack the investor’s information bottleneck, not the friction that builds conviction. His three target workflows: idea generation, initial context building, and monitoring — the last being where, as a generalist, “there was more than one earning season when I got smacked” for being last to notice end-market demand weakening.
• Monitoring’s wider-net problem is now a particularly tangible AI win: it can put “a smart filter on top of all that data” across a holding’s ecosystem. Own Expedia and you don’t want every Marriott headline — just the data points on travel demand, pricing, and OTA distribution and share shifts, a mosaic historically available only to dedicated sector analysts.
• A major pre-buy edge is killing ideas faster and pulling deep-dive analyses up the pipeline. Mapping five years of proxy comp metrics is “now on Portrait a click of a button,” and reconstructing 3-4 years of hard and soft guidance builds a management credibility profile — a team that promises margin expansion every year and never delivers can kill a turnaround thesis before the 12-hour reading marathon starts.
• Idea generation works for trend exposure (tariffs, second-order supply-chain effects) and for encoding a nuanced mental model — but articulating the model is the hard part. Plon’s own: previously high-performing franchises revalued after a potentially temporary hiccup — “a lot of times it’s more of a feeling, right? You know it when you see it.” When it lands: “clear my calendar, I’m going to spend the next week just figuring this out.”
• Prompt like you’re emailing a smart overnight analyst overseas: background, task and why, output shape, guidelines, and domain wisdom — including “take a skeptical eye” to always-positive management commentary. Calibrate accuracy tolerance by task: “there are certain tasks where if you’re 99% accurate, you’re 0% useful” (model building), while triage-stage industry surveys can absorb a stray error.
• Spend ~15% of your time experimenting: keep a suite of ten tasks the models can’t yet do reliably and rerun them on every new release, because the capability frontier is “quite jagged.” On raw models, don’t overload the context window; at around 70% usage, “you will see a degradation in any complex task” (Gemini ~1M tokens, GPT ~400K, Claude/Opus ~200K).
• The compounding bet: document thinking, decisions, and research now, because model usefulness “rises exponentially with the amount of context,” while day-to-day adoption must come bottom-up rather than by mandate alone. He’s near-term bearish on memory (“a bit of a shortcut” for pasted context) but long-term imagines a model that “has lived every single one of a firm’s investments”; agentic research is following coding’s lead — “a lot more meat behind that buzzword than there was maybe a year ago.”

Deep dive

1. An investor-turned-builder targets the information bottleneck, not conviction

• Plon’s path: Barclays special situations, generalist at long/short fund Slate Path Capital, then Baupost’s public markets team across equities and distressed credit; the itch dates to business school at Stanford from 2015-2017 — deep learning was “very much in the zeitgeist,” still pre-transformer, but AI in research “just seemed inevitable.”
• His governing distinction: the output of research is “hopefully a high quality decision,” and some friction builds conviction — “I could never really outsource model building” — so AI belongs where hours-in-a-day limited him, not where the struggle creates ownership.
• The three buckets he names: idea generation (there were probably dozens of sweet-spot ideas at any time; it’s hard to know if you’re working on one), initial context as a generalist (“table stakes” on how an industry works), and monitoring the surrounding ecosystem of customers, suppliers, and competitors.

2. Monitoring: a smart filter over the ecosystem, not just the ticker

• There are many good solutions for watching one name; the unlock is “casting a wider net” for relevant data in a much sparser stream. His Expedia example: you don’t care about Marriott’s unit growth versus Hilton — only data points feeding the mosaic on consumer travel demand, pricing, OTA distribution strategy, and market-share shifts. Historically, “there was no smart filter on top of all that data”; now AI makes this much simpler.
• Matt’s corroboration from his transport-analyst days: late Thursday nights “Ctrl+F-ing” CPG transcripts for freight-cost commentary while the rest of the world enjoyed Manhattan happy hour.

3. Pre-buy work: kill ideas quicker and pull the deep dive forward

• Plon’s year-end accounting as an investor: “a remarkably low number” of ideas ever reached deep research, and many could have died far earlier — existential risks or comp misalignment surfaced before committing “12 hours reading through all the historical content.”
• Two categories of pulled-forward analysis: templatizable screens (proxy comp mapping over five years, aggressive revenue-recognition flags) and pattern hunts — reconstructing 3-4 years of hard and soft guidance (“we expect revenue to accelerate sometime in the second half”) to profile credibility. If a turnaround management has promised margin expansion three straight years and it never happens, “that would be enough to maybe kill the idea.”
• The subtlety worth keeping: a Bloomberg screen says management “beats guidance every quarter,” but dig in and the Q1 full-year guide gets revised down all year — “you probably should shade whatever management is saying.” Matt’s flip side: kitchen-sinking all the pain in one quarter “can be better” than serial guide-downs.

4. Idea generation: turning “a feeling” into a query

• Mode one is trend exposure: when tariffs were first announced — “I guess in April of last year” — investors asked which companies had predominantly U.S.-based supply chains while competitors had international supply chains. That second-order mapping is an area where AI is useful, with modern models bringing substantial world knowledge about who might be affected.
• Mode two is harder: encoding a nuanced mental model. Plon’s own — previously high-performing businesses hit by some sort of hiccup, such as a macro issue, bad product cycle, or execution mess, where the market is revaluing the franchise. The headwind may be temporary, while “the headline numbers are going to look bad whether it’s temporary or not.”
• The real obstacle is articulation: some investors list attributes A, B, C, but “a lot of times it’s more of a feeling.” Portrait works backward from trading history and firm context to define “what a 10 out of 10 idea looks like for you” — and when it works, “there are few feelings as exciting in this business… clear my calendar.”

5. Prompting: write to the smart analyst overseas

• The durable mental model — even as effective prompting changes every three months — is an email to someone smart working overnight who lacks your context: background, the task and its why (“build this cost curve because I think it might be shifting”), optional output format, task guidelines, and domain knowledge.
• Sometimes withhold the output spec: constraining format can hurt, the same way you’d give a human analyst “some slack on the rope” to synthesize creatively.
• His favorite domain-wisdom line: management commentary is “always biased positively… it’s important you take a skeptical eye” — necessary because helpfulness training skews models optimistic, like “an eager college student who’s excited and believes in the good.”
• Calibrate by stakes and structure: “if you’re 99% accurate, you’re 0% useful” applies to high-precision tasks such as model building; early triage surveys can tolerate a factual error here or there. Structured tasks — upload the documents, since off-the-shelf web crawling is imperfect and can produce hallucinations; exploratory ones — instruct it to “pull on threads, chase down leads.” And iterate: Matt notes that responses are instant and query cost is trivial; Plon calls the process “a two-way dance.”

6. Treat capability as a jagged, moving frontier

• Spend ~15% of time experimenting: the frontier holds “undiscovered capabilities you can pick up on well before others do.” Plon reruns a suite of ten tasks that models cannot yet do reliably on every new release to gauge the leap and reposition the frontier’s edge — once one works, such as a cost curve, save the template and reuse it in the research process.
• Matt’s own prompt experiment: write the same prompt at escalating specificity and watch outputs shift with question ordering and context loading — this technology “doesn’t have many buttons that tell you, oh, this is four-wheel drive.”
• Context-window mechanics: sizes have been flat in roughly the past year (Gemini ~1M, GPT ~400K, Opus/Claude ~200K), but usage improved — models once handled needle-in-a-haystack lookups yet struggled to “build a three-statement model” across five 10-Ks. In managed tools like Portrait or NotebookLM, “I wouldn’t really hold back”; on raw models, using around 70% of the window can cause degradation on complex tasks.

7. The compounding bet: document now, let the agents arrive

• Top-down edicts such as “you must use such-and-such tool” can backfire. What works: firmwide initiatives requiring no forced process change — bespoke idea-generation screens run “like an outsourced analyst,” thesis monitoring off a ticker — while day-to-day usage earns trust bottom-up, individual by individual.
• The sports-analytics lesson applied: model usefulness “rises exponentially with the amount of context,” so memos and short paragraphs on why a trade happened become valuable intellectual property — “hard to imagine two or three years from now that every piece of data isn’t being used within a model that is operating within the context of a fund.” Matt’s onboarding anecdote: quick post-earnings blurbs tracking an emerging Amazon threat taught him more than polished buy and sell memos.
• On memory, an honest hedge: near-term it’s “a bit of a shortcut” for repeated context — he’s “a little bearish” versus just pasting it in — but long-term imagines a model that “has lived every single one of a firm’s investments,” capturing the experience and “scar tissue” behind great investors’ intuitive judgment and potentially becoming more powerful than any individual human. His proposed AGI test: “can it predict the future?”
• Agentic AI — reason, reflect, take action, and reflect on those actions toward a goal — has crossed from buzzword to working: Portrait’s original GPT-4 agent needed a roughly 30,000-token system prompt because it could not reliably self-correct; now Claude Code and Codex can be left alone for multiple hours. Code is the perfect proving ground — local context, “the code either runs or it doesn’t” — and the underlying iterative-reasoning capability now works, though adapting it to investment research remains an engineering problem: “it’s just a matter of the engineering work.”