Machine God

Who is assembling the last invention?

In 1965, I. J. Good’s Speculations Concerning the First Ultraintelligent Machine described the intoxicating loop: build a machine better than us at designing machines, let it design a better version of itself, and stand well back while improvement becomes an intelligence explosion. The first ultraintelligent machine might therefore be the last invention we need to make. This is a marvellous sales proposition provided nobody asks whether “last invention” is a product category or a threat.

The modern construction method is less thunderbolt, more procurement spreadsheet. Scaling Laws for Neural Language Models found predictable improvements in language-model loss as model size, data and training compute increased. That result does not prove that adding more accelerators eventually manufactures a deity. It does, however, give an industry with warehouses of accelerators a powerful reason to behave as though the matter has already been settled by mathematics.

Levels of AGI tries to replace the fog machine with two useful questions: how capable is the system, and how broadly can it apply that capability? Superintelligence is the far end of both axes, not a magic property emitted when the server count reaches a sufficiently theological number. Between here and there sit training runs, evaluations, data pipelines, cooling systems, human feedback, safety decisions and executives describing all of this as inevitable so that nobody notices they are making choices.

That is the machine on the shirt: not a god discovered, but one under construction. The stack rises like a cathedral because scale is excellent theatre. The halo is an oscilloscope trace because even the sacred glow came from somebody’s instrument bench. At the bottom, two tiny technicians have the rear panel off and are still wiring the alleged final invention together. MACHINE GOD is the prospectus. SOME DISASSEMBLY REQUIRED is the engineering report. Divinity, at present, remains a project plan with a power budget.

Read it for yourself: Speculations Concerning the First Ultraintelligent Machine, I. J. Good, Scaling Laws for Neural Language Models, Kaplan et al., Levels of AGI for Operationalizing Progress on the Path to AGI, Morris et al.