The violinists and the sunk thesis
In 1993 K. Anders Ericsson, working with Ralf Krampe and Clemens Tesch-Römer, published a study of violin students at the Berlin Hochschule der Künste. Sorted by teacher rating into the merely good, the very good and the exceptional, the groups differed in one measurable way: accumulated hours of solitary, effortful, corrected practice. Not talent as reported, not years enrolled, not stage exposure. The specific hours in which a task just beyond current ability was attempted, scored against a standard, and repeated with the error named.
The problem Ericsson was solving was older than 1993. Psychology had a talent puzzle: experience predicted skill weakly and inconsistently. Chess masters studied by Adriaan de Groot in 1946 showed the same pattern later formalised by Herbert Simon — expertise tracked structured exposure to positions with known outcomes, not raw hours at the board. Ericsson's contribution was to relocate the explanation from the person to the loop. Skill is not what accumulates in someone who shows up. It is what accumulates in someone whose actions are scored against outcomes fast enough, and specifically enough, to correct the next action.
Venture capital is a useful place to watch this finding fail to arrive.
The partner who has seen everything and learned little
A general partner with fifteen years in the seat has sat through several thousand pitches, read several hundred data rooms, and sits on a dozen boards receiving monthly decks. By the exposure account of expertise, this is a formidable amount of learning. By Ericsson's account it is mostly noise, because almost none of it closes a loop.
The pitch is heard once. The decision to pass is rarely revisited against what the founder actually built. The decision to invest generates a number — marked-up, marked-down, written off — on a cycle of seven to ten years, contaminated by a fund-level narrative that resists correction because admitting the thesis was wrong two years in is expensive in ways that admitting it eight years in, quietly, in a footnote to LPs, is not. The partner accumulates exposure at extraordinary volume and receives almost no scored feedback on the specific judgement that mattered: was this founder, this market, this structure, going to work.
The characteristic failure of the profession names itself precisely: a thesis is defended for a year after the market it assumed has dissolved. Someone underwrote a company on the belief that a category of infrastructure spend would keep compounding, or that a regulatory gap would stay open, or that a competitor's product was structurally inferior. The filings that would have shown the category decelerating arrive quarterly and are read. The hiring signals that would have shown the competitor scaling engineering headcount are visible on public job boards. The product telemetry that would have shown user retention cratering is, in many cases, sitting in a data room the partner has access rights to and does not open between board meetings. None of this is secret information. It is unattended feedback, arriving continuously, scored against nothing, because the partner's belief was fixed at underwriting and the ritual of the annual board deck asks "how's it going" rather than "was I wrong."
This is the radiologist who has viewed a million films and never learned a diagnosis, transposed into term sheets.
Where the corpus, the room and the open stream diverge
The lineage from Large Language Model to Large World Model to Large Universe Model is a lineage in what a system is permitted to observe, and the venture case shows exactly why the boundary matters, because each position corresponds to something a fund's process actually does today.
A fund's memo library, pattern libraries and "we've seen this movie before" instincts are Large-Language-Model-shaped: a frozen corpus of prior deals, read once, generalised into pattern-matching heuristics, and never updated by what happened to the deals it was trained on unless someone deliberately goes back and scores it — which is rare, expensive, and organisationally unrewarded. The corpus contains other people's closed loops, encoded as case studies, not the fund's own loop on the decision now in front of it.
A single board meeting, a single data-room review, a single diligence sprint is Large-World-Model-shaped: a bounded scene with real feedback inside it. The founder's answer to a hard question, the reference call that contradicts the deck, the cohort chart shown live — these close a loop, but only about what is inside the room, for as long as the room lasts. The founder's retention six months after the round, the competitor's Series B eighteen months later, the second-order effect of a rate move on the category's cost of capital — all mature after the scene has ended and are invisible to it.
What the profession's failure mode actually calls for is something structurally different from both: filings, hiring signals, product telemetry and market structure kept open as live, attributed, decaying beliefs against every thesis on the books, for as long as the position is held — years, not the length of a board meeting. That is the condition under which the partner could, in principle, learn from the specific bet rather than from the folklore of the asset class. It is also not a product any fund runs today. It is the shape the failure mode is pointing at.
| position | what it holds about a thesis | when it learns it was wrong |
|---|---|---|
| Large Language Model | prior deals as memo, frozen | never, unless manually revisited |
| Large World Model | this meeting's evidence | only while the meeting lasts |
| Large Universe Model | live filings, hiring, telemetry, structure | as the market moves, with provenance |
Two objections a serious reader will raise
The corpus is not naive. The memo library already encodes outcomes — this is what happened to the last five companies that looked like this one. That is feedback, already closed by someone.
This is the strongest defence of the memo, and it should be granted its full weight. A well-kept post-mortem is feedback that a previous partner already scored. The failure is selective, not total: the library contains outcomes for the bets that were made and written up, not for the bet in front of the partner now, in this market structure, against this competitor set, at this valuation. Daniel Kahneman and Gary Klein's account of when intuition can be trusted turns on exactly this distinction — expert judgement transfers only where the environment is regular enough that past feedback resembles the decision now being made. Venture markets are precisely where that regularity breaks down fastest: the category that supported the last five comparable exits can dissolve in the eighteen months between term sheet and Series B, which is the entire content of the domain's characteristic failure. Borrowed feedback is calibrated to somebody else's market. It says nothing about whether this one is still there.
Streams of filings and telemetry do not make a good investor. Ericsson's finding required a coach who chose the task at the edge of ability and named the specific error. A partner drowning in dashboards without curation gets the plateau — the twenty-year GP who is no sharper than the five-year one — not expertise.
This is the sharper objection and it should not be waved off. Continuous intake is not sufficient; undirected volume produces exactly the plateau this objection names, and funds that mistake more data for better judgement will get more noise, not more skill. But the objection assumes curation and intake are separable, and in this domain they are not. A partner cannot decide which thesis to re-underwrite this quarter, or which specific assumption to test against new hiring data, without a current, dated record of what the thesis actually claimed and what has since become observable against it. The coach Ericsson describes is a function applied to a feedback record. Without an open, attributed record — this filing, dated, against this line of the memo — there is no task to select and no error to name. The frozen memo cannot supply that record because it was never designed to be checked against the present. The stream is the precondition for the coaching, not a replacement for it.
The top rung, not the top of intelligence
None of this claims that continuous filings and telemetry make someone a better investor by volume alone — Macnamara's meta-analytic finding that deliberate practice explains a modest share of variance in professional judgement, well under half, applies here as everywhere, and talent, timing and market luck still carry enormous weight in any individual outcome. What the venture case demonstrates is narrower and does not depend on that share being large. It demonstrates that a thesis fixed at underwriting, defended by a mind that stopped receiving scored information about it, has no mechanism by which it could be corrected before the capital is gone. Enlarging the memo library does not fix this. Sitting through more board meetings does not fix this. The only structural remedy is intake that stays open long enough to see the consequence and carries enough provenance to say which line of which thesis it falsifies. Once a fund's beliefs are held that way — live, dated, revisable, decaying on schedule — there is no further category of observation to add to the loop. What is left after that is discipline, trust in the record, and the willingness to act on what it says.