The Untrainable: A Simple Map of AI Value

Coding AI got roughly six times better on its tests in a year — from solving 13% of benchmark tasks to the high eighties. Real-world software shipping went up about 30%.

Sit with that gap. The machine got dramatically smarter, but the work barely moved. That gap is the whole story.

Investor Sarah Guo calls what lives inside it “the untrainable”: work that AI cannot eat no matter how smart it gets. It maps to a simple 2x2 grid. This post explains that grid in plain language, so you can find where your own work sits on it.

The Two Questions That Draw the Map

The matrix is built from two yes/no questions about any piece of work. Answer both, and you know which box you’re in.

Question 1: Is “correct” public or private?

Public correctness means anyone can check the answer against a shared standard. A math problem, a chess move, a coding test — there’s a right answer out in the open, and everyone agrees on it.

Private correctness means “right” only exists inside one company’s walls. What counts as a good M&A deal memo at this law firm, for this client, under this set of confidential facts? Nobody outside can grade it. The answer key is locked away.

This is the axis that matters most. The sharpest way to put it: “Anything you can put on a leaderboard, you can train against.” If an answer is public, an AI lab can practice against it until the model wins. If the answer key is private, the lab never even sees the test.

Question 2: Is the work frontier or saturated?

Frontier work is hard and cutting-edge — the stuff at the edge of what’s possible today.

Saturated work is routine — done a million times, cheap, and well understood.

That’s it. Two questions, four boxes.

The Four Boxes

Here’s the full grid. Each box behaves differently as AI improves.

Walk through them one at a time.

Saturated + public: commodity tokens. Routine work with an open answer. Summarize this email, translate this paragraph, write a generic function. Free and open-source models already do this well enough. The price falls toward zero. Nobody builds a lasting business here.

Frontier + public: the labs win. Hard work, but with an answer anyone can check — and coding benchmarks are the headline example. Because the test is free and public, owning it is worthless; whoever trains the smartest model takes the prize. That’s the foundation labs themselves, not the companies building on top — when the test is free, owning it counts for nothing.

Saturated + private: walled but routine. Locked away, but not hard. Real value exists, but it’s local and limited — there’s not a huge prize in automating routine work only your firm can see.

Frontier + private: the untrainable. Hard work where the answer key lives inside one organization. This is the prize corner. The lab building the model never sees your private ground truth, so it can’t train against it — no matter how clever the model gets.

Why the Untrainable Corner Is the Prize

The smartest model in the world still hits a wall here: “A smarter model does not make private ground truth public. It does not hold the license, sign off on the liability, or own the firm’s files.”

Three things keep this corner safe.

The answer key is private. A doctor decides if a diagnosis is right. A partner decides if a deal memo holds up. That judgment lives in people and proprietary files, not in any dataset a lab can buy. The AI can’t practice against a test it can’t see.

Someone has to sign off. In banking, healthcare, and law, a licensed human is accountable when things go wrong. You can’t delegate liability to a model. That accountability is a moat made of trust, not intelligence.

Trust is a lock and a deadbolt. Getting into a hospital’s or bank’s workflow needs security and integration — that’s the lock. Getting people to actually change how they work is the deadbolt, and it only opens through relationships built over time. The medical-AI tool OpenEvidence didn’t win on raw compute; it won by earning physician trust. Money can’t buy that quickly.

The takeaway in one line: a token spent answering a generic question is worth almost nothing, but a token spent reasoning over your company’s private data is worth much more. Or, more simply — the untrainable is value with history.

Where Does Your Work Sit?

Put your own work on the map. Ask the two questions.

• If your answer is public and routine, AI is coming, and fast. Move up or move private.
• If your value is a thin layer over a public model, it gets absorbed. The lab pulls your trick into the base model.
• If your value lives in private judgment, accountability, and trust — the relationships, the files, the sign-off only you can give — that’s the durable corner. Lean into it.

The doom story says AI eats everything. The doom is only half right. Generic wrappers do get eaten. But value keeps pooling in the one place models can’t reach: behind organizational walls, inside human trust, where judgment still beats measurement.

Find your untrainable corner — and build there.