Every young investor starts in the same place: the case studies.

The founder interviews. The origin myths. The stories about getting in early, back when nobody else understood it. Read enough of them and venture starts to look learnable — as if the people who made the great bets simply saw, a year or two ahead of the room, what was always going to be obvious in hindsight.

Every one of those stories was written by someone who survived.

Picture CES, January 2024. A small orange box under the lights.

The Rabbit R1, a hundred and ninety-nine dollars. You talk to it and it orders your food, hails your cab, books your flight. A marquee Valley fund led the round to thirty million, and its founder promised that within a decade there would be more agents loose on the network than there are people on Earth — one in every pocket.1 Microsoft's CEO called it the best demo since the first iPhone.2

VCs lost their minds. Real technology, heavy names, a pain point that could not hurt more.

A hundred thousand units sold. Twenty million in revenue. Five months in, five thousand people still opened it on a given day.3

Same wave of cheap money: Stability AI, the most beautiful name in the room. Stable Diffusion overnight, valuation talk climbing toward four billion.

Then the round didn't close. No takers.4

AutoGPT, the same story told faster. Top of GitHub's trending page within days, a well-known venture firm twelve million in.5 For a few weeks everyone believed the agent that does your job for you had finally arrived, and then they watched it spin in place, burning GPT-4 tokens at three to six cents a thousand, looping patiently back to where it had started.5

All three got told and retold, mostly as jokes, mostly over drinks.

None of them ever made it onto anyone's track record.

A warplane comes home with bullet holes across the wings, the fuselage, the tail. Map enough of them and the lesson looks obvious: that is where the metal gets hit, so that is where you bolt the armor.

The map lies. Those holes are on the planes that made it back. A plane hit through the engine or the cockpit does not come home with holes to count — it does not come home at all. The clean spots on the survivors are the kill spots.6

The map of damage is a map of every place a bomber can be shot and still survive.

That is the whole mechanism, and it is worth saying once, plainly: venture is a prediction problem, and the industry trains on the planes that came back. The case studies, the founder interviews, the track records — all of it is damage data drawn from survivors.

But the failures aren't merely hard to see. They get destroyed, three times over, each pass cleaner than the last.

First, the founder never tells it. A company dies and the story simply stops. There is no post-mortem for a thing that vanished, no stage, no festival panel — nobody books the founder of a startup that no longer exists. The primary record is never written down at all.

Second, the investor never records it. Whoever wrote the losing check keeps it off the track record. Missing a good deal is the only mistake this industry will let you confess in public; nobody ever raised a fund by owning up to a loss. So the loss moves as a story over drinks, gossip, and never hardens into data.

Third — and this is the one nobody counts — the investor who actually learned the lesson leaves.

The people who come to understand, in their bones, how brutal the base rate is are disproportionately the ones who got burned learning it. And getting burned is how you fail to raise a third fund. Most firms never raise one; an emerging manager, by definition, is a firm that has not yet raised three.7 So the field quietly sheds the exact people carrying the correction. They walk out with the lesson still in their heads, unwritten, and what stays in the room is whoever wasn't burned — or whoever buried it.

Run the count and you can see why. Take a fund with thirty companies in it. Even in the rare fund that returns more than five times its money, fewer than six of those thirty carry almost all of it.8 The other twenty-four come back middling, or come back nothing.

So most people in this business never touch the deal that would have kept them in it. They make reasonable bets, watch them die, and run out of road before the one that mattered ever shows up — if it was ever going to show up. They are the ones who learned the most. They are the least likely to still be here.

And the few who stayed can't pass it on either. From inside a win, skill and luck look identical. You wait seven, eight, ten years to find out whether the one bet that mattered was right, and by the time the answer lands you can't tell which it was. The survivors keep the trophies and lose the reason.

So the record empties out at every level. No public story. No institutional memory. No one left in the chair who paid for the knowledge. The next cohort doesn't inherit a biased dataset.

It inherits a scrubbed one.

When a company finally goes down — shot through the engine, gone before it could be measured — it leaves no mark on anyone's map.

So they turn back to the hangar and keep studying the planes that came home. The ones whose pilots lived to be interviewed.

Strip everything else away and venture capital is a prediction problem.

Starve a prediction problem of data and it overfits. There's nothing moral in it — just what happens to any estimator trained on a censored sample.

This industry has spent its whole history cutting half of its own training set: the half where the plane went down. It trains the next cohort of investors on the bullet holes — and now it trains the machines the same way, on the runs that paid off, the failed ones filtered out before anyone looks.

The people are smart enough. They were always smart enough. Intelligence was never what this industry ran short on.

What it runs short on is the other half of the data — the planes that never landed — and that half is the one nobody will ever hand over.

Notes

[1] R1 priced at $199; a major venture firm led the round bringing total funding to $30M; the firm's founder made the "tens of billions more agents than people on the planet" prediction. rabbit inc. press release, Business Wire, 2023-12-21.

[2] Microsoft's CEO, in a Bloomberg interview at Davos (WEF), called it one of the most impressive demos since the iPhone launch — a remark about the quality of the presentation, not a declaration of an "iPhone moment." Decrypt / Bloomberg, 2024-01-17.

[3] Over 100,000 units sold, ~$20M revenue; founder Jesse Lyu disclosed ~5,000 daily active users and 16 OTA updates at the Fast Company Innovation Festival. The device shipped in April 2024 and the figure was disclosed in September — hence "five months after it shipped." Fast Company, 2024-09 (via 9to5Mac, TechTimes).

[4] Stability AI failed to raise at a ~$4B valuation; the round collapsed for lack of takers. TechCrunch / Bloomberg, 2024-03; VentureBeat, 2024-03.

[5] AutoGPT released 2023-03-30, top GitHub trending within days; parent company Significant Gravitas raised $12M in October 2023, led by a venture firm with another fund participating — no valuation disclosed. GPT-4 API ran roughly $0.03 per 1K input tokens and $0.06 per 1K output. Wikipedia "Auto-GPT"; founder's OpenUK interview, 2023.

[6] The warplane-armor problem is the classic illustration of survivorship bias: the failures (downed planes) are absent from the data, so reasoning from the survivors points exactly the wrong way. Origin — Abraham Wald, Statistical Research Group, Columbia University, WWII; popularized in Jordan Ellenberg, How Not to Be Wrong (2014).

[7] "Emerging manager" is typically defined as a firm with fewer than three funds; most struggle to raise a second or third without exits. PitchBook / TechCrunch, 2024.

[8] Horsley Bridge, across ~7,000 of its investments (1975–2014): in funds returning more than 5x, fewer than 20% of deals produced roughly 90% of returns. Via Toptal, "Venture Capital Portfolio Strategy." (The "thirty companies, fewer than six" framing applies this ratio to an illustrative portfolio.)