The 1983 DARPA plan to build machine intelligence

Strategic Computing: New-Generation Computing Technology, A Strategic Plan for its Development and Application to Critical Problems in Defense is DARPA's own founding plan for the Strategic Computing Initiative, published October 28, 1983 and approved for public release. It is not a leaked or FOIA'd document. DARPA wrote it to be public, to justify the program to Congress and industry. It was archived by the Defense Technical Information Center under accession number ADA141982.

The plan runs to roughly 110 pages and lays out a ten-year roadmap, with the first five years budgeted at approximately 600 million dollars, rising from 50 million in fiscal year 1984 to 150 million by fiscal year 1986, with later years marked to be determined by program progress.

The plan does not hedge. Its executive summary states that DARPA plans to create a new generation of machine intelligence technology, with unprecedented capabilities, and frames the effort as a response to advances converging simultaneously in artificial intelligence, computer science, and microelectronics. The document explicitly names expert systems, machine vision, speech recognition, and natural language understanding as the technical pillars, and states that the new machines would exhibit human-like, intelligent capabilities for planning and reasoning.

The plan's conclusion goes further, predicting that if the technology succeeds, Americans will have a significantly improved capability to handle complex tasks and to codify, mechanize, and propagate their knowledge, and that leaders would be able to tap the nation's pool of knowledge through intelligent computers acting as active assistants. This is, in substance, a 1983 description of what would decades later be called an AI copilot or an AGI-adjacent decision support system.

DARPA selected three military applications to drive the technology, each pulling on a different capability. The Autonomous Land Vehicle was intended to navigate up to 50 kilometers of cross-country terrain, planning routes from digital terrain data, identifying obstacles, and reporting reconnaissance findings, all while moving at speeds up to 60 kilometers per hour. The plan estimated this would require computing power on the order of 10 to 100 billion operations per second, compared with roughly 30 to 40 million operations per second in the most powerful conventional computers of 1983.

The Pilot's Associate was conceived as an AI system personal to an individual combat pilot, trained by that pilot, capable of automatically reconfiguring aircraft systems in response to battle damage and offloading routine cockpit tasks so the pilot could focus on tactical decisions. The plan describes it explicitly as not replacing but complementing the pilot.

The Battle Management System, developed in connection with the existing DARPA and Navy program aboard the USS Carl Vinson, was designed to generate hypotheses about enemy intent, propose courses of action, simulate likely outcomes for each option, and explain its reasoning to a human commander, who would retain the final decision. The plan estimated this system alone would require a distributed expert system running roughly 20,000 rules.

Strategic Computing ran from 1983 to 1993 and ultimately spent over a billion dollars, more than the initial plan's five-year estimate. Historical accounts of the program, including the definitive study by historians Alex Roland and Philip Shiman, describe a program that produced real technical advances in parallel computing, VLSI chip design, and speech recognition, but fell well short of its original machine intelligence goals. When new IPTO leadership took over in 1987, funding for the artificial intelligence software components was cut sharply, and the ambitious end-to-end demonstrations envisioned in this 1983 plan were never fully realized in the form the document describes.

This document establishes, in DARPA's own written words from October 1983, that the US military set out with a formal, funded, ten-year plan to build machine intelligence for autonomous weapons systems, pilot assistance, and battle command, using the same conceptual building blocks, planning, reasoning, vision, speech, and natural language, that define AI research today. It establishes specific, quantified technical goals and budgets that the agency committed to in writing.

What it does not establish is what the program actually achieved. This is a plan, not a results report. The historical record, assembled separately from this document, shows the program spent a decade and over a billion dollars pursuing these goals with mixed results, falling short of full machine intelligence while advancing several underlying technologies. Readers should treat the ambitions described here as DARPA's 1983 forecast, not as an account of what was built.