ARTIFICIAL INTELLIGENCE

Artificial intelligence is a field of computer science concerned with building systems capable of behavior that would require intelligence if a human performed it.

What AI Is Not

AI is not automation in the traditional sense, and conflating the two causes most of the confusion in mainstream coverage. A rule-based system that routes customer service tickets based on keyword matches is automation; it has no ability to generalize beyond the rules it was given. A search engine indexes and retrieves existing content using statistical ranking — it does not synthesize, reason, or generate. Modern AI systems, by contrast, use machine learning to learn representations directly from data and apply those representations to novel inputs they were never explicitly programmed to handle. The distinction matters because it determines what breaks when conditions change and what new possibilities open up.

Why It Matters for Technology

AI is becoming infrastructure. The same way cloud computing became a default substrate for software in the 2010s, AI capabilities are being embedded in development tools, business software, and consumer applications. The current generation of systems — large language models, diffusion models, reinforcement learning agents — represents a qualitative shift from earlier rule-based approaches. Capabilities have improved dramatically as compute and training data scaled, and that trajectory has not flattened. Understanding AI is now a prerequisite for making informed technology decisions, not an optional specialization.

Why It Matters for Economics

The economics of AI are structurally unusual. Training costs are enormous and front-loaded; inference costs fall rapidly as hardware and optimization improve. This creates a winner-take-most dynamic in foundation model development while enabling a long tail of application-layer businesses built on top of commodity APIs. The capital allocation implications are significant: companies that control foundation model training have large, durable advantages, while companies that build on top of those models compete on distribution and workflow fit rather than model quality. For investors and operators, the key variable is not which model is best today but which layer of the stack captures durable margin.

What Most Coverage Gets Wrong

AI coverage tends to oscillate between hype and panic, and both framings distort the actual picture. The more useful frame is industrial: AI is a general-purpose technology in the early adoption phase, analogous to electricity in the 1900s or the internet in the 1990s. The transformative effects are real, but the timeline and distribution of those effects is uncertain and uneven. Early general-purpose technologies typically improve productivity in the sectors that adopt them first — those with the least friction, the most digitized operations, and the clearest ROI — while barely touching other sectors for a decade or more. AI is following that pattern precisely.

The Capability-Deployment Gap

AI capabilities are advancing faster than deployment, and the gap is not technical — it is organizational, regulatory, and economic. Changing a hospital's clinical workflows requires FDA clearance, physician retraining, liability restructuring, and EHR integration; none of those are solved by a better model. The industries where AI deployment is moving fastest are those with the fewest regulatory constraints, the most digitized data, and the clearest ROI on reduced headcount. The industries moving slowest are not technically resistant — they are structurally resistant, and the constraint is institutional rather than computational. This gap explains why AI investment and AI productivity gains diverge so sharply in the near term: the capital is deployed years before the workflow and institutional changes catch up.

Open Questions

Whether scaling laws continue to hold as models grow larger; whether current architectures can achieve reliable multi-step reasoning or whether new architectures are required; how liability and intellectual property law will adapt to AI-generated outputs; whether the productivity gains from AI deployment will be broadly distributed or concentrated; and how labor markets will adjust to the displacement of knowledge work tasks that were previously assumed to be automation-resistant.

Part of the knowledge graph at The Best Blog Ever — reference definitions for ideas that matter.