Intelligence Was Never One Thing: The Paradox That Shaped How We Built Vantage

The hardest problems in intelligence turned out to be the ones a toddler solves before breakfast.

In May 1997, a computer beat the greatest chess player alive. Chess was supposed to be the summit of human intellect, and Deep Blue climbed it, summited and drove a definitive stake into the ground, “machines win at this game”. Yet no machine on Earth that year could have walked into the room and moved a pawn without knocking over the clock. Nearly three decades later the pattern holds. AI passes the bar exam and still cannot reliably fold laundry.

This inversion has a name: Moravec's paradox. And it's the most useful map we have for deciding where AI belongs in a firm, a workstream, or a life.

The hard problems are easy and the easy problems are hard

In 1988, the roboticist Hans Moravec observed that getting computers to perform at adult level on intelligence tests or board games was comparatively easy, while giving them the perception and mobility of a one-year-old was somewhere between difficult and impossible. Picking up a penny is still a computationally intense task for an AI-powered humanoid robot, yet a toddler can do it with relative ease. Marvin Minsky made the companion observation: we are least aware of what our minds do best. Steven Pinker later compressed the whole field's experience into one line: "the hard problems are easy and the easy problems are hard."

In plain English: the things that feel difficult to us, like calculus, chess, or drafting a benchmark analysis, feel difficult precisely because they're evolutionarily new. We perform them with slow, conscious, effortful attention. The things that feel effortless, like recognizing a face, reading a room, or catching a falling glass, feel effortless because they run on machinery evolution has been polishing for half a billion years.

The paradox isn’t random. The difficulty of automating a human skill tracks the evolutionary age of that skill, not how impressive it looks on a resume. Perception and motor control have been under continuous optimization since the first vertebrates, on the order of 500 million years. Language is perhaps 100,000 years old. Writing is about 5,000. Formal logic is roughly 2,500.

Imagine these capabilities laid out over the compressed timeline of a single year. Vision and movement start on January 1. Language shows up around 10 pm on December 31. By most estimations, writing arrives five minutes before midnight. Formal reasoning, the crown jewel of white-collar work, occupies the final two and a half minutes.

If we zoom in on the final minutes, they're not uniform. One of the hardest tasks for a firm is fusing dozens of weak, noisy, correlated signals into a single calibrated probability. Weighing burn rate against cohort retention against churn against market timing against a thousand faint echoes of deals that came before, and outputting one number you'd stake capital on. Firms have tried to staff analysts against it, but the human mind was never built for the job. This kind of probabilistic reasoning didn't exist as a formal skill before the 18th century. On the calendar, it's the last couple of seconds before midnight.

The skills evolution perfected are intuitive, unconscious, and massively parallel, while the skills we acquired last are taxing, conscious, and serial. What's more, the recent skills are the easy ones to replicate precisely because we can articulate them. We can write down the rules of chess and the details of the share classes of a cap table. Nobody can write down how a great investor reads a founder, because that skill predates writing by 100,000 years.

Perceived difficulty is a terrible guide to computational difficulty. It is almost perfectly inverted.

The second act nobody expected

For 30 years, the paradox was a robotics story. Then large language models gave it a second act, and four things changed at once.

Language fell to the machines. Reading, writing, summarizing, and pattern-matching across text moved decisively to the machine side of the ledger. Fluent prose is now abundant.

Context stopped being scarce. Human working memory holds roughly four items at the same time. A modern model's context window holds on the order of a million tokens, roughly 3,000 pages, not skimmed and summarized but held simultaneously, every clause available at all times. That's not a better analyst's memory. It's a different kind of memory.

Attention became free. A machine can run scheduled jobs against every news source, every filing, every secondary signal, every portfolio company, every morning, without fatigue and without a Friday afternoon. Human attention is the scarcest asset in any firm. Machine attention is now effectively unlimited.

The temptation arrived. Because machines now speak fluently, it became easy to believe they can do everything, and an entire generation of AI software is being built on that belief.

The second act nobody expected

For thirty years the paradox was a robotics story. Then large language models gave it a second act, and four things changed at once.

Language fell. Reading, writing, summarizing, and pattern-matching across text moved decisively to the machine side of the ledger. Fluent prose is now abundant. As we have written before, fluency is not judgment, but fluency at scale is real and it is new.

Context stopped being scarce. Human working memory holds roughly four items at once, a limit cognitive scientists have measured for decades. A modern model's context window holds on the order of a million tokens, roughly 3,000 pages, not skimmed and summarized but held simultaneously, every clause available at all times. That is not a better analyst's memory. It is a different kind of memory.

Attention became free. A machine can run scheduled jobs against every news source, every filing, every secondary signal, every portfolio company, every morning, without fatigue and without a Friday afternoon. Human attention is the scarcest asset in any firm. Machine attention is now effectively unlimited.

The temptation arrived. Because machines now speak fluently, it became easy to believe they can now do everything, and an entire generation of AI software is being built on exactly that belief.

"Replace the humans" is a category error

Almost every industry is now fielding some version of the full-replacement pitch. Payroll is the largest line item in most professional firms. The demos are dazzling. And some tasks do automate end to end, cleanly and completely. The gap between replace-everyone and what it actually delivers is worse than vaporware, because there might actually be some output and it will be garbage.

If intelligence were a single axis, a dial that goes from mouse to human to superhuman, then replacement would just be a matter of waiting for the dial to turn. But Moravec's point is that intelligence isn't a single axis. It's terrain: radically different from one region to the next, with machines and humans holding the high ground in different places.

Build as if the dial exists and the mistake gets poured into the foundation. No model upgrade fixes a foundation.

Approach	How it works	Where it breaks
Full replacement	Automate the entire function and remove the humans from the loop	Puts AI in the judgment seat, the one region where it holds no advantage, and removes humans from exactly the seats where they do
Autonomous decision engines	Model output executes directly: the AI trades, allocates, decides	Produces conviction without an owner. When it is wrong, and it will be, nobody can say why and nobody can answer for it
One synthesized assistant	A single model reads everything and returns one smooth, blended answer	Averages away disagreement. The most valuable signal in any real team is where the specialists diverge, and synthesis erases it

Each of these draws the boundary between human and machine according to economics or enthusiasm. None of them draws it according to where the actual advantages sit.

The Moravec allocation

There's a better rule: give each kind of intelligence the ground where its advantage is asymmetric, and treat the boundary between them as the most important design decision in the system.

On the machine side of the boundary, the advantages aren't marginal. They're categorical. Holding an entire data room, a decade of portfolio history, and every prior memo in working memory at once. Watching every relevant source on a schedule that never slips. Recognizing patterns across corpora no human career is long enough to read. Benchmarking against reference sets that update continuously. Maintaining a competitive map that evolves every single day instead of every quarter. These aren't jobs machines do better; they're jobs the human mind was never built to do at all.

On the human side, the advantages are just as categorical, because they tap into age-old parts of the human experience. Judgment under irreducible uncertainty, where the data ends and the decision still has to be made. Trust built across a table, and the thousand unconscious reads that inform it. Intuition that this founder will run over obstacles and through walls, again and again. Taste, including the contrarian conviction to act against the pattern the data suggests. Relationships that compound over decades. And accountability, which isn't a skill at all but a property only a person can have: a decision needs an owner, and a model can't be one.

This rule refuses to treat the human as a fallback for whatever the machine can't do yet, and it refuses to treat the machine as a junior employee.

How we built Vantage around the paradox

Vantage is what happens when you build an investment firm’s operations around this allocation.

Vantage does not make investment decisions. It's not an AI trading platform, and it never will be, not because the technology is immature but because the architecture is wrong. Decisions live on the human side of the boundary.

On the machine side, we didn't build one assistant. We built more than 30 individual agents, each a specialist in its own discipline: conviction analysis, benchmarking, valuation, comparable events, memo work, and the rest of the disciplines a firm actually runs on. These agents disagree with each other, and the platform surfaces that disagreement instead of blending it into a smooth consensus.

Memo Analyst vs. Financial Model Analyst “the Story & the Spreadsheet”

An investment memo arrives making the case for a Series B, and two specialists read it. The Memo Analyst, built to attack load-bearing assumptions, comes back almost complimentary: the narrative is internally consistent, the founder's claims are evidenced, the story holds. The Financial Model Analyst, which never reads narrative at all, runs the same plan forward and returns a different verdict: hitting these numbers requires every account executive to close at a rate no cohort in the company's own history has produced. Both are right. The story is coherent, and the story is impossible, and those are different tests. The platform doesn't split the difference into a cautious maybe. It puts the disagreement on the table, because deciding which test matters more, for this company, at this moment, is exactly the judgment the partner is paid for.

Deal Terms Analyst vs. Thesis Checker and the “2x Mirage”

A portfolio company closes new financing at a headline number well above the last mark, and the Thesis Checker reads it the obvious way: an outside investor just re-underwrote the thesis at a higher price, and conviction should move up. The Deal Terms Analyst never sees the headline. It reads the documents, and the documents say participating preferred with a 2x stack ahead of the common. Its conclusion: on these terms, the price of the shares the fund actually owns is a third below the number in the press release. Both agents read the same round. One read the signal, the other read the security, and the question they hand the partner is the one a headline never asks: validation at what price and in what paper?

When you put all this together the result is the best team we could imagine placing around yours: tireless, specialized, comprehensively read, and pointed at exactly the terrain where machine intelligence dominates, so your partners can spend their finite attention on the terrain where human intelligence does.

Alongside this operating system, the fleet of specialists on the machine side that never sleeps, is the operating team: a bench of expert operators your partners tag in when the work demands it. Both sit underneath the people you already trust, freeing them for the terrain that was always theirs: the judgment, the relationships, the reads no model can make.

The deepest implication of Moravec's paradox isn't about software. It's that intelligence was never one thing. A toddler's glance, a grandmaster's calculation, a founder's instinct, and a million-token context window are different shapes of intelligence, each extraordinary in its own terrain and unremarkable outside it. For 70 years we ranked them on a single scale, and the paradox is the evidence that the scale was always wrong.

The real question of the next decade isn't how much human work machines can absorb. It's whether we have the discipline to notice which shape of intelligence each problem actually calls for, and the wisdom to put every gift, human and machine, on the ground where it's strongest.

That's the future we're building toward. It starts with a 40-year-old idea and a toddler who can do something no machine can.

Vantage puts AI where it holds the asymmetric advantage and keeps humans where they are irreplaceable. The judgment and the decision stay yours.

Conviction Made Citable.