What this covers

Matt talks with David Plon about how working investors are folding AI into research. The conversation starts with definitions: what agentic AI means, why code is a friendlier environment for agents than investment research (a contested claim), and why context window sizes have stayed roughly flat over the past year even as models have gotten better at using them. The recurring example is mapping how CEO pay metrics changed across five years of proxy statement filings, once a painful manual job and now a matter of minutes. Matt adds a sports analogy: teams that kept data before analytics went mainstream benefited decades later.

The middle of the discussion is practical. David describes templatizing pre-buy checks for aggressive revenue recognition, tracking every piece of management guidance over three to four years in a Management Credibility Audit, and shading forward commentary when full-year guidance keeps being cut despite quarterly beats. He advises telling models that management is biased positive, and uploading source documents rather than letting a model crawl the web. He also separates needle-in-a-haystack lookups from multi-document synthesis, a gap he says has narrowed.

The more speculative stretch covers continual learning, and a firm-wide memory that has lived through every investment and pass, which David thinks could outdo any individual's judgment. He also suggests that forecasting is a strong test of general intelligence, and the discussion turns to whether off-the-shelf models can already trace the effects of tariffs. On habits, he recommends about 15 percent of time spent experimenting, a personal suite of ten tasks that models cannot yet do, and bottoms-up adoption rather than mandates.

Sharpest takeaway

David Plon argues AI's near-term value for investors lies in cheaply triaging information, surfacing patterns, and killing weak ideas earlier, while conviction-building stays human. Returns depend on learned skills (prompt framing, experimentation, documenting decisions) more than on the tools themselves.

  • AI expands the information an analyst can triage, but the decision and conviction must still come from the investor
  • Prompt quality, iteration, and a regularly re-tested exercise suite determine how much value investors extract from models
  • Firm-wide AI adoption works best when it does not force changes to individual research processes; trust is built bottoms-up
  • Documented decisions and reasoning become proprietary context that makes AI more useful over time

The argument · threads7 threads · 65 claims
0.81

AI screening kills weak ideas early and widens research coverage across names and sectors.

6 pointscentrality 5/5
  • Most ideas that investors look at never reach deep research, and AI can surface enough baseline context to kill an idea earlier, such as spotting an existential risk or misaligned management compensation before spending many hours on historical reading.

    “how many ideas did I really look at? A remarkable low number that made it through to the deep research stage.”

  • Analyses that once required deep research, such as mapping CEO compensation metric changes across five years of proxy statements, can now be done in minutes, letting analysts turn over more rocks and spend creative time on pre-screened ideas.

    “go through the last five proxies and try to map out what are the metrics that the CEO is being comped on ... now on Portrait it's a click of a button.”

  • Historically, monitoring a stock like Expedia required reading everything about adjacent companies because there was no smart filter; AI now lets investors pull only the data points relevant to their thesis, such as hotel demand and OTA distribution shifts.

    “There was no smart filter on top of all that data.”

  • The most useful AI applications in investing are high-level information processing tasks that triage new data points across three stages: idea generation, initial context building on new names, and monitoring existing positions.

    “high-level information processing to help triage new data points – broadly speaking, are the types of things where AI can be really useful.”

  • Generalist investors struggle to track customers, suppliers, and competitors across a sector, and were repeatedly caught last to realize end-market demand was weakening or suppliers were pushing through price increases, which AI monitoring can now address.

    “I was the last person to realize that end-market demand was weakening or suppliers were pushing prices through”

  • The acceptable accuracy of an AI output depends on where the task sits in the research process: for tasks where 99 percent accuracy yields no usefulness, errors are unacceptable, while early triage or industry surveys can tolerate factual errors.

    “There are certain tasks where if you're 99% accurate, you're 0% useful, and that's like self-driving cars.”

0.63

Investors should prompt models iteratively, adding context and reasoning, and allow output latitude.

5 pointscentrality 4/5
  • Because LLM queries are nearly instant and cheap, investors should start with a simple prompt, add complexity as needed, and give real-time feedback, treating the process as a two-way dance rather than waiting overnight for analyst output.

    “The cost of sending a single query is trivial. Start simple, start adding complexity as it's helpful.”

  • For exploratory or creative research, investors should upload relevant documents but explicitly instruct the model to pull threads, follow leads, and perform extensive web searching, since the answer's direction is unknown.

    “pull on threads, follow it, chase down leads.”

  • Telling a model why it is building an analysis, such as a cost curve that might be shifting and implying future pricing changes, produces better output than giving the bare task.

    “build this cost curve because I think it might be shifting and that could imply something about future changes in the pricing”

  • Prescribing exact output formats for an LLM can limit its creativity in synthesis, just as over-constraining a human analyst can, so investors should sometimes give the model latitude.

    “sometimes that can be hurtful in that you're constraining the model's output the same way you would a human analyst.”

  • A useful mental model for writing LLM prompts is to imagine emailing a smart overnight analyst who lacks context, and include the background, task, requested output, guidelines, and domain knowledge a human would need.

    “imagine you are writing an email to somebody maybe overseas who's going to work overnight and is going to be doing a task for you.”

0.63

AI adoption must leave analysts' conviction-forming process and personal judgment intact.

5 pointscentrality 4/5
  • The combination of qualitative reasoning and the difficulty of articulating investing criteria makes idea-finding hard, but when AI matches a pitch to an investor's historical 10-out-of-10 idea profile, it produces a uniquely exciting result.

    “even enunciating that clearly is hard”

  • Investors often find attractive ideas as previously high-performing businesses with durable economics that were hit by a temporary headwind, and the market revalues them; the hard part is that the criteria are gestalt and difficult to translate into a machine query.

    “a lot of times it's more of a gestalt feeling, right?”

  • Friction in the investment research process is valuable because it builds the analyst's personal conviction, so AI should accelerate information gathering without replacing the steps where the investor forms conviction, such as building the model personally.

    “that friction was helpful because it helped me build conviction personally. I was one of these guys that could never really outsource model building.”

  • Top-down mandates to use specific AI tools often backfire, whereas firms succeed by providing firm-wide initiatives that do not change individual research behavior and letting people adopt AI for daily tasks when they trust it.

    “I think that adoption has to happen bottoms-up, and people need to feel comfortable with it”

  • Firm-level AI initiatives such as bespoke idea-generation screens and thesis monitoring work well because they deliver insights without asking analysts to change their process, and they can be ignored when irrelevant.

    “we just run those like an outsourced analyst. And that's really value-add, because all we're doing is, we're not asking anyone to change the process.”

0.57

Agentic AI is viable for multi-hour work but remains constrained by context and self-correction limits.

7 pointscentrality 3/5
  • Agentic AI means a model can reason, reflect, take actions, and reflect on those actions in pursuit of a goal, with varying degrees of scaffolding and constraint on what it can do.

    “the model has the capability to reason, reflect, take action, reflect on those actions, ultimately in the pursuit of a goal.”

  • Early GPT-4 agent attempts at Portrait only worked when the model was tightly constrained, with a system prompt of around 30,000 tokens, because the agent could not self-correct when it went down a wrong path.

    “I think our system prompt was 30,000 tokens or something crazy.”

  • Modern coding agents such as Claude Code and Codex can be left alone for hours, dynamically understanding codebases, making changes, testing, and debugging, showing that multi-hour agentic work has become viable.

    “people are leaving these models alone for multiple hours and it is dynamically operating like a software engineer”

  • Frontier context windows have been roughly flat over the past year, at about one million tokens for Gemini, around 400,000 for GPT, and about 200,000 for Claude, while models have improved at using large context intelligently.

    “Gemini at around a million, GPT I think around 400,000, and Opus or Claude around 200.”

  • When using a raw model rather than a managed tool, one should not fill most of the context window, since using around 70 percent of it causes degradation in complex tasks, so tasks should be broken into smaller pieces.

    “you do not want to overload the context window and use 70% of the context. You will see a degradation in any complex task.”

  • Models historically handled needle-in-a-haystack lookups well but struggled with tasks requiring attention across many documents simultaneously, such as building a three-statement model from several 10-Ks; this gap has narrowed.

    “They call that a needle-in-a-haystack problem. They would struggle if you said, "Build a three-statement model using these."”

  • Code is an ideal environment for AI agents because it is text-based, locally contained in a repository, and verifiable since code either runs or does not, whereas investment research lacks these properties.

    “it's super verifiable – the code either runs or it doesn't.”

0.53

Documented firm decisions fed into AI memory could eventually build judgment beyond any individual.

6 pointscentrality 3/5
  • Firms should document their investment thinking, decisions, and short rationales in real time, because this data will become valuable context for AI systems that operate within the firm and is hard to load in otherwise.

    “that data can be captured in real-time, that forms an enormously valuable source of IP that will make AI unique to you.”

  • Because models store large amounts of knowledge in their weights from pre-training, the same paradigm should extend to continual learning on the job, which David expects to be a major development.

    “it's just a matter of time before that same paradigm applies to continual learning on the job.”

  • The practical near-term use of memory is as a shortcut for context that must be repeated in every interaction, such as system-level knowledge or the history of about 20 queries on a single name.

    “It's a bit of a shortcut for just loading in context that you don't have to repeat.”

  • In purpose-built tools that manage context, investors should provide the entire document corpus and let the model choose which content to use, much as one would give a junior analyst agency over sources rather than a fixed list.

    “I wouldn't really hold back. I would give it pretty much the entire corpus.”

  • If AI memory is implemented well, a model that has lived through every investment a firm made and passed on could develop intuitive judgment from experience that might exceed any individual human's.

    “imagine having a model that has lived every single one of a firm's investments and maybe things they even passed on.”

  • Current AI memory implementations are far worse than human memory because they are narrow, expensive in context, and lack the abstract pattern-forming quality that makes human memories useful for investment judgment.

    “today, as it's implemented, it's far worse than a human's memory. It is super limited.”

0.45

Management guidance track records reveal credibility, so forward commentary needs skeptical reading.

4 pointscentrality 3/5
  • Headline data showing management beats guidance every quarter can be misleading when full-year guidance is revised down each quarter, so investors building year-ahead models should subtly shade management's forward commentary.

    “they end up revising it down quarter after quarter after quarter”

  • Because models are trained to be helpful, they tend to be overly positive about management commentary, so prompts should explicitly tell the model that management teams are biased positively and that it should read them skeptically.

    “the models have been trained to want to be helpful, and that usually influences them to be more positive than they otherwise should be.”

  • Tracking every piece of management guidance, including soft qualitative statements, over three to four years reveals a management team's credibility, such as repeatedly promising margin expansion that never arrives, which can be enough to kill a turnaround thesis.

    “over the last three years every year they've said they expect margins to expand and it never happens”

  • Building a record of every piece of management guidance, including qualitative statements, over the last three to four years is essential for judging management's guidance style and credibility, and an investor new to a name lacks this context that long-time followers have intuitively.

    “building up a picture of management's guidance style and credibility is really important”

0.39

Investors should continually test models against their own tasks and ground them in uploaded documents.

6 pointscentrality 2/5
  • In a fast-evolving technology with a jagged capability frontier, investors should spend roughly 15 percent of their time experimenting with models, because this surfaces undiscovered capabilities before others and builds intuition about where models are heading.

    “spending some 15% of your time on experimentation is really important”

  • Off-the-shelf state-of-the-art models already contain enough world knowledge to identify first- and second-order effects of events such as tariffs, including which companies have US supply chains while competitors rely on international ones.

    “which companies have a predominantly US-based supply chain and their competitors have an international supply chain”

  • For structured, quantitative tasks, investors should upload source documents rather than rely on an off-the-shelf model crawling the web, because web crawling is inefficient and tends to produce hallucinations.

    “generally speaking, an off-the-shelf model is going to have to go crawl the web, and that's an imperfect, inefficient way to gather a bunch of structured information, and you'll most likely get hallucinations.”

  • Investors should maintain a personal suite of around ten tasks that current models cannot yet do reliably, run it against every new model release, and save each solved task as a reusable research template.

    “I run my suite of 10 different exercises across it.”

  • Predicting the future is a strong test of general intelligence in a model, because it requires a robust world model and understanding of many interacting variables, which is exactly what an investor does.

    “can it predict the future? You need a really good world model and understanding of lots of variables”

  • Some pre-buy checks can be templatized, including looking for aggressive revenue recognition and changes in accounting assumptions that inflate figures, plus recurring patterns that automatically turn an investor off.

    “There's certain accounting things I would look for, with respect to like aggressive revenue recognition”