Part X: Frontiers and Limits

"For a successful technology, reality must take precedence over public relations, for nature cannot be fooled."

Richard P. Feynman, "Personal Observations on Reliability of Shuttle"

Part IX ended with machines that break. The frontier is not only where capability grows. It is where constraints become visible. This part asks about a different kind of limit: not what the substrate can supply, but what the supply converts into. Compute is an input and capability is an output, and between them sit three constraints that no amount of hardware removes: the stock of learnable signal, the instruments that report how far the frontier has moved, and the cost of turning a produced claim into an accepted one.

These questions end the book's account of capability itself. What follows takes capability as given and asks what is done with it: first the markets that price and distribute it, then the practice of operating it. Grouping the three limits here makes a point the earlier parts could not: the ceiling on a frontier system is now as often a limit on what can be known as a limit on what can be built.

Chapter 70 starts on the input side, where the stock of high-quality human text is finite and every proposed replacement remains contested: synthetic data, learning from experience, spending compute at inference rather than training, and models that keep learning after deployment. Chapter 71 turns to measurement, where the frontier reads better as the length of task a model can complete on its own than as a benchmark percentage, and where the instruments saturate faster than replacements can be built. Chapter 72 closes on acceptance: once claims are cheaper to produce than to check, proof, replication, adversarial review, and the institutions that grant trust become infrastructure in their own right.

The three limits are not symmetrical. One is about fuel, one about instruments, one about trust, and only the first has much chance of yielding to more spending. A shortage of compute announces itself as a queue and a bill. A shortage of instruments or of trust announces itself as a claim nobody can check, which takes far longer to notice and costs more to correct.