Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497]

What It Covers

Cliff Sosin, a concentrated investor with ~80% of his portfolio in Carvana, argues that intelligence has fundamental limits in complex dynamic systems, making the true investment opportunity not in AI supply infrastructure but in owning the complements to intelligence: data, attention, physical proximity, and scale.

Key Questions Answered

• •Intelligence as a ratio, not a scalar: Treat intelligence like the roundness of a ball rather than an unbounded number. For complex dynamic systems — weather, markets, human behavior — even perfect intelligence cannot produce closed-form answers. Small errors compound through simulation steps, making prediction fundamentally unreliable. This means AI cannot simply "think harder" to solve most real-world problems; it must touch reality through experiments, which take time regardless of compute.
• •Owning complements to intelligence: As intelligence commoditizes toward near-zero cost, the assets that retain value are those intelligence cannot replicate: proprietary data, user attention, physical proximity, regulatory position, and scale. Meta exemplifies this — its entire cost base (data centers, inference, engineering, content moderation) is essentially purchased intelligence, meaning falling AI costs structurally reduce expenses while data and attention assets remain intact and compounding.
• •The diamond-shaped demand curve for AI: AI addressable markets initially expand as models improve from inadequate to capable, unlocking larger use cases. But demand eventually narrows at the frontier — very few tasks require 300 IQ reasoning. Investors should expect rapid commoditization behind the frontier wave, with frontier tokens commanding premium pricing only in a shrinking set of genuinely hard tasks, making AI lab valuations structurally difficult to underwrite.
• •AI agents accelerate low-cost-producer advantages: When consumer AI agents continuously optimize spending decisions, price stickiness disappears. A garbage collection company with two customers on one street can now efficiently offer lower prices to the whole street, winning volume. Businesses that are already low-cost producers in markets with high consumer inertia — logistics, local services, financial products — gain disproportionate share as agents remove switching friction and route demand to the cheapest qualified provider.
• •Corporate AI adoption sequencing: The practical path for any company to capture AI value runs in four stages: (1) give every employee, including the CEO, read-only AI access to all internal systems; (2) drive usage until demand for write access emerges organically; (3) establish write permissions; (4) build change-management processes for AI-driven system modifications. Companies starting from zero today have a structural advantage over incumbents burdened with legacy system integration debt across each of these stages.
• •Education in an AI-abundant world: Children need two parallel tracks. First, foundational skills learned without models — multiplication before matrix multiplication, writing before prompting — because directing AI effectively requires domain knowledge sufficient to communicate at a non-crayon level. Second, ambitious model-assisted projects where students must justify and own the output. The goal is not AI avoidance but building the conceptual vocabulary needed to extract frontier-level value from models rather than surface-level outputs.