@prefix : . @prefix schema: . @prefix owl: . @prefix rdfs: . @prefix rdf: . @prefix skos: . @prefix dcterms: . @prefix foaf: . @prefix xsd: . # ── Article ──────────────────────────────────────────────────────────────────── :article a schema:Article ; schema:name "The Untrainable" ; schema:headline "The Untrainable" ; schema:description "A thesis-driven investment essay arguing that durable AI value lies not in model capability but in private ground truth, domain trust, and illegible correctness that no benchmark can reach." ; schema:url ; schema:datePublished "2026-06-10"^^xsd:date ; schema:dateModified "2026-06-10"^^xsd:date ; schema:author :sarahGuo ; schema:publisher :saranormousSubstack ; schema:about :thesis_untrainableValue, :concept_absorptionFrontier, :concept_illegibleCorrectness, :concept_privateTruth ; schema:keywords "AI investment, untrainable, private ground truth, absorption frontier, illegible correctness, benchmarks, AI moat" . # ── Author ───────────────────────────────────────────────────────────────────── :sarahGuo a schema:Person ; schema:name "Sarah Guo" ; schema:url ; schema:sameAs ; schema:jobTitle "Venture Capital Investor" ; schema:description "AI-focused venture investor and author of the Saranormous Substack." . :saranormousSubstack a schema:Organization ; schema:name "Saranormous (Substack)" ; schema:url . # ── Central Thesis ───────────────────────────────────────────────────────────── :thesis_untrainableValue a schema:CreativeWork ; schema:genre ; schema:name "The Untrainable Thesis" ; schema:description "Durable AI value concentrates in work whose correctness is private, expensive to verify externally, and locked inside systems requiring trust and permission — the 'untrainable corner' no benchmark or general model can colonize." ; schema:hasPart :claim_benchmarkEatsItself ; schema:hasPart :claim_intelligenceNotBottleneck ; schema:hasPart :claim_absorptionFrontierRises ; schema:hasPart :claim_thinWrapperAbsorbed ; schema:hasPart :claim_offenseHarder ; schema:isPartOf :article . # ── Core Claims ──────────────────────────────────────────────────────────────── :claim_benchmarkEatsItself a schema:Statement ; schema:name "Benchmarks are self-defeating" ; schema:description "A thing you can measure is a thing you can train against. Once work is checkable cheaply, it commoditizes toward open models at lowest cost. The measurable is always on its way to commodity." . :claim_intelligenceNotBottleneck a schema:Statement ; schema:name "Intelligence is not the bottleneck — permission and accountability are" ; schema:description "A model smarter than any person still needs to be let in the door. Capability does not confer license, liability, or ownership of private data. Trust is built slowly on relationships, not gradient descent." . :claim_absorptionFrontierRises a schema:Statement ; schema:name "The absorption frontier keeps rising" ; schema:description "Labs pull scaffolding (retrieval, routing, tool use, reasoning policy) into weights, narrowing the app layer. The untrainable ground shrinks under whoever stands on it; companies must continuously step toward what can't yet be scored." . :claim_thinWrapperAbsorbed a schema:Statement ; schema:name "Thin wrappers are being absorbed" ; schema:description "Companies built merely on top of a frontier model with no proprietary data or trust integration are at severe absorption risk. The despair about AI investability is 'half right' on this point." . :claim_offenseHarder a schema:Statement ; schema:name "Choosing what to build is harder than defending what you have" ; schema:description "The model cannot tell you what is worth building. Identifying unaddressed use cases before the rest of the market is a uniquely human judgment — perhaps intent is an even scarcer input than compute." . # ── Key Concepts ─────────────────────────────────────────────────────────────── :concept_untrainableCorner a schema:DefinedTerm ; schema:name "The Untrainable Corner" ; schema:description "Frontier work whose correctness exists only in private data — the 2x2 quadrant where tasks are not saturated and answers are not publicly verifiable, making it resistant to general model commoditization." ; owl:sameAs . :concept_absorptionFrontier a schema:DefinedTerm ; schema:name "Absorption Frontier" ; schema:description "The expanding boundary at which model labs pull application-layer scaffolding (retrieval, routing, tool use, reasoning) into the base model weights, eroding the margin of app-layer companies." . :concept_illegibleCorrectness a schema:DefinedTerm ; schema:name "Illegible Correctness" ; schema:description "A form of truth that cannot be read off any leaderboard and is only discovered by running a complex system in the world over time. Decade-old codebases, hospital workflows, and legal matter flows exemplify this." ; owl:sameAs . :concept_privateTruth a schema:DefinedTerm ; schema:name "Private Ground Truth" ; schema:description "Correctness that exists only inside a firm's data, workflows, and accumulated judgments. Neither trainable from public data nor verifiable by external benchmarks." . :concept_twoByTwo a schema:DefinedTerm ; schema:name "Saturated × Public/Private 2×2" ; schema:description "A strategic framework: (1) Saturated + public answers = commodity tokens for open models; (2) Frontier + public answers = lab territory (free eval); (3) Saturated + private = partial moat; (4) Frontier + private = the untrainable corner, where durable value lives." . :concept_tokenEconomy a schema:DefinedTerm ; schema:name "Token Value Gradient" ; schema:description "As articulated by Matt MacInnis at Rippling: a token answering a generic question is worth almost nothing; a token reasoning over proprietary company data is worth much more." . # ── Evidence & Supporting Data ──────────────────────────────────────────────── :evidence_mitStudy a schema:ScholarlyArticle ; schema:name "MIT Study on Coding Agent Productivity" ; schema:author "Mert Demirer et al." ; schema:description "Across more than 100,000 developers, the latest coding agents lifted code written by ~180% but code actually shipped by only ~30%. Writing got cheap; the rest still runs through a person." ; schema:about :concept_illegibleCorrectness . :evidence_devinBenchmark a schema:Dataset ; schema:name "SWE-Bench Devin Trajectory" ; schema:description "Devin shipped in 2024 solving 13% of SWE-Bench tasks. By mid-2026 the best agents hit the high 80s. The benchmark was conquered but the 'wrong lesson' was drawn — engineering resists full measurement." . :evidence_noamBrownQuote a schema:Quotation ; schema:name "Noam Brown on Evaluation Horizon" ; schema:description "OpenAI's Noam Brown wrote that the only sure way to evaluate an agent over a one-year horizon may be to run it for a year — the clock cannot be skipped." ; schema:author :noamBrown . :evidence_openEvidence a schema:Observation ; schema:name "OpenEvidence Physician Adoption" ; schema:description "A majority of American doctors now open OpenEvidence every day. No amount of compute buys physician habit or decision-flow integration; trust is built slowly through relationships." ; schema:about :org_openEvidence . # ── People Mentioned ────────────────────────────────────────────────────────── :gabePereyra a schema:Person ; schema:name "Gabe Pereyra" ; schema:description "Observed that real automation requires the product, model, workflow, and firm to move together — and three of those four move at the speed of an organization." . :mattMacInnis a schema:Person ; schema:name "Matt MacInnis" ; schema:description "Rippling executive who articulated the token value gradient: generic-question tokens are near-worthless; proprietary-data-reasoning tokens carry real value." ; schema:affiliation :org_rippling . :noamBrown a schema:Person ; schema:name "Noam Brown" ; schema:description "OpenAI researcher, pioneer of reasoning models. Observed that a one-year-horizon agent may require a full year of real-world operation to evaluate." ; schema:affiliation . # ── Organizations & Products ────────────────────────────────────────────────── :org_rippling a schema:Organization ; schema:name "Rippling" ; schema:url . :org_openEvidence a schema:SoftwareApplication ; schema:name "OpenEvidence" ; schema:description "AI medical reference platform used daily by a majority of American doctors." ; schema:url . :org_sierra a schema:Organization ; schema:name "Sierra" ; schema:description "AI customer service company that charges only when its agent resolves a customer issue — outcome-based pricing that works because Sierra owns the definition of 'resolved'." ; schema:url . :org_cognition a schema:Organization ; schema:name "Cognition (Devin)" ; schema:description "AI software engineering company offering a 'performance guarantee' on Devin's coding output — only possible because they're trusted inside the customer's system." ; schema:url . :org_harvey a schema:Organization ; schema:name "Harvey" ; schema:description "AI legal company that publishes a benchmark for law — earning the right to define what 'good' means in legal work through real adoption." ; schema:url . :org_baseten a schema:SoftwareApplication ; schema:name "Baseten" ; schema:description "Inference serving provider; best AI-native companies concentrate serving on one or two providers for reliability and guaranteed compute access — not just cheapest token price." ; schema:url . :org_fireworks a schema:SoftwareApplication ; schema:name "Fireworks AI" ; schema:description "Inference serving provider alongside Baseten; reliability under real traffic and guaranteed access to scarce compute are non-commodity differentiators." ; schema:url . # ── HowTo / Strategic Playbook ──────────────────────────────────────────────── :howto_findUntrainableCorner a schema:HowTo ; schema:name "How to Build in the Untrainable Corner" ; schema:description "A strategic playbook derived from the article's thesis for companies seeking durable AI value." ; schema:step :step1_getInside ; schema:step :step2_arrangePrivateReality ; schema:step :step3_writeDownWhatGoodMeans ; schema:step :step4_priceOutcomes ; schema:step :step5_avoidCapitalWar ; schema:step :step6_keepSteppingForward . :step1_getInside a schema:HowToStep ; schema:position 1 ; schema:name "Get inside one" ; schema:text "Do the unglamorous work of passing the security review, signing the integration contract, and becoming the party with your name on outcomes. Permission precedes capability." . :step2_arrangePrivateReality a schema:HowToStep ; schema:position 2 ; schema:name "Arrange the customer's private reality" ; schema:text "Connect the model to the firm's private data, internal tools, and workflows. The translation of a company's private reality into model-actionable form is hard to copy and never ends." . :step3_writeDownWhatGoodMeans a schema:HowToStep ; schema:position 3 ; schema:name "Write down what good means" ; schema:text "Accumulate enough in-domain judgments to define evaluation criteria. Companies that own the definition of 'resolved' or author the legal benchmark earn standing no frontier lab can buy." . :step4_priceOutcomes a schema:HowToStep ; schema:position 4 ; schema:name "Price outcomes, not inputs" ; schema:text "Charge for resolution, not tokens. Outcome pricing is only credible if you own the definition of success and are trusted inside the system. It signals depth of integration." . :step5_avoidCapitalWar a schema:HowToStep ; schema:position 5 ; schema:name "Avoid the general-model capital war" ; schema:text "Train specialized models on your private data and evals for narrow tasks. Companies that try to out-train frontier labs on general swaths usually end in a sale to someone compute-rich." . :step6_keepSteppingForward a schema:HowToStep ; schema:position 6 ; schema:name "Keep stepping toward the not-yet-scored" ; schema:text "The untrainable ground shrinks continuously. Don't find a defensible spot and rest. Re-underwrite constantly, moving toward whatever work cannot yet be benchmarked." . # ── FAQ ──────────────────────────────────────────────────────────────────────── :faq1 a schema:Question ; schema:name "Why is investor despair about AI investability only 'half right'?" ; schema:acceptedAnswer :faq1answer . :faq1answer a schema:Answer ; schema:text "The despair correctly identifies that thin wrappers over frontier models face absorption. It is wrong about what survives: private ground truth, domain trust, and illegible correctness create durable value the model layer cannot reach." . :faq2 a schema:Question ; schema:name "Why can't a smarter model win in markets like consumer chat?" ; schema:acceptedAnswer :faq2answer . :faq2answer a schema:Answer ; schema:text "ChatGPT held its lead through years of real competition not on model quality alone. Gemini is gaining share via Android and Search, not a better model. Trust, habit, and distribution are non-model advantages the public weighs." . :faq3 a schema:Question ; schema:name "What makes the law firm M&A example archetypal?" ; schema:acceptedAnswer :faq3answer . :faq3answer a schema:Answer ; schema:text "A top white-shoe M&A practice runs ~1,000 deals per year with practice-area-specific deal shapes (NDA → term sheet → diligence → purchase agreement) and confidentiality constraints. Transformation requires an operator with ambiguous intermediate goals, incomplete feedback, and a multi-year horizon — no single eval can capture it." . :faq4 a schema:Question ; schema:name "What is the 'absorption frontier' and why does it keep rising?" ; schema:acceptedAnswer :faq4answer . :faq4answer a schema:Answer ; schema:text "The absorption frontier is the expanding boundary at which model labs pull application scaffolding (retrieval, routing, tool use, reasoning policy) into base weights. It rises because more of the measurable work is continuously being benchmarked and then trained against, narrowing the app layer unless companies move into private, illegible territory." . # ── Glossary ────────────────────────────────────────────────────────────────── :gloss_moat a schema:DefinedTerm ; schema:name "Moat" ; schema:description "In competitive strategy, a durable barrier protecting a business from rivals. In the AI context, the untrainable corner argues that moats come from private data and trust relationships, not model quality." ; owl:sameAs . :gloss_verifier a schema:DefinedTerm ; schema:name "Free Verifier" ; schema:description "A check that costs nothing to run — a compiler, a test suite — enabling rapid grinding against a benchmark. Its absence is what makes a task illegible and resistant to training." . :gloss_groundTruth a schema:DefinedTerm ; schema:name "Ground Truth" ; schema:description "The authoritative correct answer to a question. 'Private ground truth' exists only inside a firm's systems and cannot be accessed or verified externally." ; owl:sameAs . :gloss_inference a schema:DefinedTerm ; schema:name "Inference (AI Serving)" ; schema:description "Running a trained model to produce outputs. While token price commoditizes on schedule, reliability, and guaranteed compute access during high traffic do not." ; owl:sameAs . :gloss_distillation a schema:DefinedTerm ; schema:name "Model Distillation" ; schema:description "Technique to produce a smaller, cheaper model from a larger one. Enables focused apps to run a workflow on a fraction of the token spend of a general agent." ; owl:sameAs .