The Untrainable

Thesis Abstract🔗

The central argument of The Untrainable is that the AI industry's despair — that only compute and frontier weights hold value — is half right but wrong about what it leaves. Intelligence keeps getting cheaper, and value keeps sliding toward the few places a model cannot reach.

Sarah Guo identifies a structural dynamic she calls the absorption frontier: the boundary where measurable work gets pulled into model weights. What is legible (benchmarkable, trainable) gets commoditized. What survives is illegible work — whose correctness is private, grounded in history, and established inside organizations through trust and relationships.

"The most cited benchmark score of the year is a map of territory about to be worthless, and a notice of who is about to lose the right to say what counts as good."

The Thesis Framework🔗

Guo proposes evaluating any knowledge work domain along two axes:

Key Insight: Private Correctness

Private correctness is the article's central concept. It means truth that exists only inside a specific firm's data and context — the kind that cannot be read off a leaderboard. You find out whether a complex system works by running it in the world long enough to learn. As Noam Brown of OpenAI noted, evaluating an agent over a one-year horizon may require running it for a year. The clock cannot be skipped.

The Absorption Frontier🔗

The absorption frontier describes how the labs are trying to get models to swallow their own scaffolding — the retrieval, routing, tool use, and reasoning policy that used to wrap the model are being pulled into the weights, until the wrapper is the model. Margin pressure cuts both ways: a general agent must be ready for anything (expensive), while a focused application tunes one workflow until it runs on a fraction of the token spend and keeps the difference.

"What's legible is what's leaving. The valuable work is illegible by construction: anything you can put on a leaderboard, you can train against, so anything measurable is already on its way to commodity."

Thin wrappers — shallow applications over foundation models — are being absorbed. But absorption is not total. On a narrow task with private data and custom evals, a focused company can train to the frontier and beat the general models where it counts. That specialized model becomes part of the moat.

Permission and Accountability🔗

Guo argues that intelligence is not the bottleneck. Permission and accountability are. The door to private-correctness domains has two barriers:

The Economics of Value🔗

Matt MacInnis of Rippling provides the value framework: a token spent answering a generic question is worth almost nothing (anyone's model can answer it), while a token reasoning over a company's specific data is worth much more (it does the thing you actually want).

Outcome-based pricing is the natural mechanism for this: Sierra charges when its agent resolves a customer issue, nothing when it escalates. Cognition's Devin offers a performance guarantee in software. You can only price outcomes in a system you are trusted inside.

"A company that brings the translation is tough to copy — and the translation never ends. Integration and maintenance run as long as the relationship does."

Competitive Dynamics🔗

Guo directly addresses the objection that labs will run first-party products below cost to bleed out application-layer companies. Her rebuttal: this only works if the model layer is a single-player game. It is not — it looks like a three-and-a-half-way death match with international players six months behind and a development league 5x the size it was last year.

In consumer chat, the best model has never simply won. ChatGPT is losing share to Gemini on the strength of Android and Search, not a better model. Anthropic, rated as having the best model by prediction markets, is barely a factor in consumer chat. If the frontier remains crowded, the layer above will be valuable.

The Innovator's Dilemma applies: incumbents keep the ground they have, and the next thing comes from someone who finds a use before the rest. Competing on the general model is a capital war you lose to whoever owns the most compute.

Knowledge Graph Explorer🔗

Interactive visualization of the thesis knowledge graph. Drag nodes to explore relationships, scroll to zoom, click nodes or edge labels to open entity pages.

Frequently Asked Questions🔗

Glossary of Key Terms🔗

The Untrainable

Value that cannot be captured by AI benchmarks because its correctness is private, grounded in history, and established through relationships inside organizations.

Absorption Frontier

The boundary where measurable work gets pulled into model weights as labs incorporate retrieval, routing, tool use, and reasoning policy directly into the model.

Legible Work

Work that can be measured, benchmarked, and trained against. Being eaten from two directions: task saturation from below and model absorption from above.

Illegible Work

Work whose correctness is private and expensive to establish. The valuable work that survives after legible tasks are commoditized.

Private Correctness

Truth that exists only inside a specific firm's data and operational context. Cannot be verified by public benchmarks; requires time and trust to establish.

Permission Moat

The double barrier of environment (security, integration, liability) and user (habit, trust, acquiescence) that protects private-correctness domains from AI absorption.

Outcome-Based Pricing

A pricing model where the fee is the evaluation: charge only when the AI agent produces a verified good outcome. Only possible inside a trusted system.

Thin Wrapper

A shallow application over a foundation model that is easily absorbed as the model incorporates the wrapper's functionality into its own weights.

Commodity Token

A token spent answering a generic question that any model can answer. Worth near zero because the buyer can substitute any provider.

Vertical Mapping

The strategic process of mapping knowledge work domains by correctness type (public vs private) and saturation level to identify where defensible AI value can be built.

Strategic Framework: How to Build Defensible AI Value🔗

Derived from Sarah Guo's thesis, seven steps for building AI companies that resist commoditization by the absorption frontier.

Explore the Knowledge Graph Using SPARQL🔗

Run SPARQL queries against the thesis analysis graph. Each query targets the named graph <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>. Click a query to expand, then Run on URIBurner to execute it live.

PREFIX schema: <http://schema.org/>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>

SELECT ?concept ?definition
FROM <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>
WHERE {
  ?concept a skos:Concept ;
           skos:definition ?definition .
}
ORDER BY ?concept

PREFIX schema: <http://schema.org/>

SELECT ?entity ?type ?name
FROM <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>
WHERE {
  { ?entity a schema:Person ; schema:name ?name . }
  UNION
  { ?entity a schema:Organization ; schema:name ?name . }
  BIND (IF(EXISTS { ?entity a schema:Person . }, "Person", "Organization") AS ?type)
}
ORDER BY ?type ?name

PREFIX schema: <http://schema.org/>

SELECT ?question ?answer
FROM <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>
WHERE {
  ?q a schema:Question ;
     schema:name ?question ;
     schema:acceptedAnswer ?a .
  ?a schema:text ?answer .
}
ORDER BY ?question

PREFIX schema: <http://schema.org/>
PREFIX thesis: <https://saranormous.substack.com/p/the-untrainable%23>

SELECT ?domain ?name ?correctnessType ?saturation
FROM <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>
WHERE {
  ?domain a thesis:WorkDomain ;
          schema:name ?name ;
          thesis:hasCorrectnessType ?correctnessType .
  OPTIONAL { ?domain thesis:hasSaturationLevel ?saturation . }
}
ORDER BY ?name

PREFIX schema: <http://schema.org/>

SELECT (SAMPLE(?s) AS ?EntityID) (COUNT(*) AS ?count) (?o AS ?EntityTypeID)
FROM <https://linkeddata.uriburner.com/dav/home/demo/docs/the-untrainable-thesis-analysis-big_pickle-1.ttl>
WHERE {
  ?s a ?o .
}
GROUP BY ?o
ORDER BY DESC(?count)
LIMIT 50