Why AI Investing Needs Good Explanations, Not Predictions

Progress happens when people guess solutions to problems, criticize those guesses, and replace worse ideas with better ones. That is the argument at the heart of David Deutsch's The Beginning of Infinity, and it has direct consequences for how investors should think about businesses, moats, and artificial intelligence.

The question of why progress suddenly accelerated in some societies and not others comes down largely to culture. Deutsch distinguishes between static societies, which suppress criticism and preserve tradition, and dynamic ones, which allow existing ideas to be questioned and replaced. The Enlightenment established a more durable tradition of criticism through open debate, competitive markets, and constitutional democracy. These mechanisms let people replace bad laws, bad businesses, and bad ideas without overturning the whole system.

Businesses are institutions that apply knowledge to solve problems for customers. Competition continually tests those solutions, and profits and losses provide feedback about whether customers value them. A durable business must embody knowledge that competitors cannot easily reproduce, while staying capable of developing new solutions as conditions change.

The trucking industry shows how similar-looking services can have very different economics. A full-truckload carrier moves one shipment directly from origin to destination, which requires no dense terminal network. Barriers to entry are low, competition is intense, and profitability is modest. A less-than-truckload carrier combines shipments from many customers across a network of terminals and routes. Its advantage depends on shipment density, route efficiency, pricing knowledge, and the coordination of thousands of daily decisions. A competitor is free to enter, but it cannot easily build that network or reproduce the operating knowledge needed to run it efficiently.

Much of the knowledge behind a durable business may be inexplicit, residing in accumulated routines and practical know-how. Recognizing an attractive opportunity does not provide that knowledge. Overcoming a moat may require creating a materially better solution rather than copying what already exists.

A moat earned through customer choice does not prevent competition; it makes the company difficult to displace. That differs from an advantage sustained by political privilege. Even a genuine moat is not permanent. Newspapers held dominant local positions for decades, supported by economies of scale and network effects between readers and advertisers. The internet lowered the cost of distributing content and shifted advertising toward digital platforms. Newspapers retained their journalism knowledge, but that was no longer enough to sustain the old competitive position.

Amazon (AMZN stock page) offers a contrast. Jeff Bezos recognized that the internet could give people access to a much larger selection of products. The retail business expanded from books while staying focused on enduring customer wants: greater selection, lower prices, and faster delivery. Those goals have no obvious endpoint. They give Amazon a reason to keep investing in fulfillment, software, and logistics while developing new knowledge about serving customers. That culture does not make Amazon immune to disruption, but it encourages the company to revise existing methods in pursuit of the same wants.

Moats protect existing earning power; adaptability helps a company improve existing solutions or develop new ones. Warren Buffett has generally preferred businesses with durable moats that do not require continual reinvention. Berkshire's experience with newspapers shows that even a genuine moat can erode when innovation changes the economics supporting it.

Every investment thesis is a conjecture, but not every conjecture is equally useful. Deutsch argues that a good explanation is hard to vary: its details are constrained by what it explains and cannot be altered arbitrarily without weakening it. A bad explanation can be changed freely while still seeming to account for almost any outcome.

The value of a business depends on the cash it can distribute to owners over its remaining life. A thesis that does not connect the price paid to those future cash flows is speculation. Buying a stock because it appears undervalued means judging that your expectations about the business are less wrong than those reflected in the market price. Because that judgment remains conjectural, a thesis's key assumptions must be explicit enough to be criticized as new evidence emerges.

Changes in stock price should be considered separately from evidence about the underlying business. A rising share price does not validate the thesis, just as a falling price does not invalidate it. What changes directly is the prospective return at the new price. When results are weaker than expected and the price also falls, two questions matter: how much should the new information change the estimate of future cash flows, and how much has the lower price changed the prospective return? A business can become less valuable while its stock becomes more attractive if the price falls by more than the estimated decline in value.

This approach is especially useful during periods of rapid technological change. Recognizing that a new technology may transform society is not enough to identify a good investment. Buffett made this point in a 1999 Fortune article: at least 2,000 companies entered the automobile business in the United States, yet by the 1990s only three U.S. car companies remained. Henry Ford's first automobile company failed. General Motors' rapid acquisition spree left it financially strained and cost founder Billy Durant control in 1910. Autos transformed society, but identifying which companies would survive was far harder.

Artificial intelligence presents the same challenge. NVIDIA (NVDA stock page) developed GPUs for video game graphics and later introduced CUDA for general-purpose computing. AlexNet demonstrated their value for training deep neural networks, Google researchers introduced the transformer architecture, and OpenAI used that architecture to create increasingly capable large language models. Few people anticipated that chips designed for video games would become central to training large AI models, let alone the range of tasks those models would perform. ChatGPT, released in November 2022, brought AI to a mass audience and helped drive a surge in demand for computing infrastructure.

One unresolved question is whether continued development will produce machines capable of creating explanatory knowledge, the threshold for artificial general intelligence. Current models do not yet demonstrate that ability in an open-ended sense. A system trained only on information available before Darwin or Einstein would have to conjecture explanations not contained in prior observations. But AI does not need to cross that threshold to be economically transformative. These systems can assist with or automate parts of coding, analysis, experimentation, and communication. By making existing knowledge cheaper to apply, they can lower the cost of producing goods and services.

The harder question for investors is how the resulting economic value will be divided among customers, companies supplying the technology, and businesses applying it. AI may improve one part of an existing service without eliminating the need for the broader product and the distribution network, customer relationships, and accumulated knowledge required to deliver it.

Investing in a company simply because it involves AI is not an investment thesis. The same questions apply: What problem does the company solve? How does AI improve its solution? What makes that solution difficult to replicate? Who captures the resulting value? What expectations are already reflected in the stock price?

NVIDIA shows that an AI-related thesis can rest on knowledge and advantages that already exist. Its hardware, CUDA software ecosystem, and relationships with developers embody knowledge accumulated over decades. Those advantages may help explain why NVIDIA has captured substantial value from AI growth, but any thesis would still need to address whether alternative chips and computing architectures could weaken its position, and whether expected future earning power offers an attractive return at the current price.

The more a thesis depends on capabilities or industry structures that do not yet exist, the more it rests on assumptions that cannot be tested in the market. Those assumptions should affect whether to invest at all, the price to pay, and the amount of capital to commit. The practical question is whether the thesis offers an attractive prospective return without requiring a chain of technological and competitive developments to go right.

Active investing does not require identifying every eventual winner as a technology emerges. Early in a technology's development, companies are still discovering which products and business models will work. Many will fail. Venture capital portfolios spread capital across many companies before their models have been extensively tested. Public-market investors can wait until customer adoption, competitive advantages, and cash generation become easier to evaluate.

AI does not overturn this framework. Its economic consequences will depend on the problems it solves, which businesses use it effectively, who retains the value created, and what expectations are already reflected in market prices. The future cannot be predicted in detail because it will be shaped by knowledge that has not yet been created. That uncertainty does not prevent sound investing. It makes good explanations, attractive prices, and a willingness to correct mistakes all the more important.