**John McCarthy (1927–2011)** was one of the founders of artificial intelligence. He coined the term *“artificial intelligence”* in the mid-1950s and helped launch the Dartmouth summer project that effectively started the field. He later created **Lisp** (late 1950s). He spent part of his early career at **MIT**, then went on to build Stanford’s AI effort, and from 1965–1980 he directed what became the **Stanford Artificial Intelligence Laboratory (SAIL)**. McCarthy received the **Turing Award** in **1971** for foundational contributions to the field. This **1964** memo reflects a career-long ambition to formalize common sense and logical reasoning within machines, aimed at solving complex real-world problems like autonomous planetary exploration. ### Public interest in Martian life When McCarthy wrote this memo, the notion that Mars might be home to intelligent life—or even thriving “colonies” of canal-building Martians—had only recently faded from mainstream belief. Widely reported newspaper headlines (e.g., New York Times, 1911: “Martians Build Two Immense Canals in Two Years”) and popular culture had fueled decades of public excitement about Martian canals and civilizations. The idea lingered into the early 1960s, even as most astronomers grew skeptical. Just thirteen months later, NASA’s Mariner 4 flyby (July 1965) sent back the first close-up photos, revealing a cratered, dry world and ending the canal myth forever. 
*New York Times cover from 1911 announcing the construction of canals on Mars* John McCarthy had coined the term “Artificial Intelligence” just 9 years earlier (1955) when proposing the Dartmouth AI project. ### Autonomous Space craft in the 1960s In June 1964, when John McCarthy wrote this memo, no autonomous vehicle had ever operated on another celestial body. All prior space probes - such as the Soviet Luna 1–3 flybys that photographed the Moon’s far side in 1959 or NASA’s Mariner 2, which flew past Venus in 1962 - were non-mobile craft equipped only with fixed, hard-wired instruments. These probes collected data during a single pass or crash landing and transmitted it back in real time or via simple recorders; they had no on-board programmability, no mobility, and no ability to navigate or adapt to terrain. Soft landings on any extraterrestrial surface had not yet been achieved anywhere (the first would be Luna 9 on the Moon in February 1966), and the concept of a rover was absent from all mission designs. "computer controlled" as opposed to fixed, hardwired sequencing and simple relay logic. McCarthy requires that the Beagle computer survive errors in individual experiment programs without bricking the whole machine. Without this protection, the flexibility of running many different programs and uploading revisions from Earth after early results arrive would be too dangerous. He draws directly on techniques then emerging in time-sharing operating systems. These systems introduced mechanisms such as user mode versus executive mode, memory protection, and timed interrupts to isolate user programs and prevent one faulty program from crashing the whole computer. In 1964 such safeguards were brand new; most computers still ran unprotected batch jobs where a single error could stop everything. McCarthy, who had proposed time-sharing concepts as early as 1959, saw that the same mechanisms could enable safe, adaptive programming on a distant spacecraft. This insight was far from obvious at the time. In 1963 NASA commissioned General Electric to study a Mars lander as part of the early Voyager program, using Saturn V rockets for a 1969 or 1971 launch. Codenamed “Beagle” (after Darwin’s ship), the study described a stationary 5,000-pound lander with 300 watts of power and a one-year lifetime. Its goal was a search for life, but it included no mobility and no on-board computer control. The projected cost exceeded $1 billion in 1964 dollars. McCarthy uses it as a technical baseline to argue that such a massive payload could and should be mobile. The ambitious Voyager Mars program was cancelled in 1967 due to budget cuts and Apollo priorities. It was later revived and scaled down as the Viking program, which successfully landed two stationary spacecraft on Mars in 1976 - the first soft landings there - carrying life-detection experiments but without mobility or the programmable computer control McCarthy envisioned. McCarthy recommends that the Beagle computer avoid mechanical secondary storage such as magnetic tapes or drums. Moving parts are vulnerable to the extreme vibration, temperature swings, and radiation of a Mars mission and would sharply reduce overall system reliability. In practice, several early spacecraft did fly magnetic tape recorders despite this risk. The Gemini spacecraft (1965–1966) used a tape drive to load programs into core memory. Mariner probes, the Viking landers (1976), and the Voyager spacecraft (1977) all carried digital tape recorders to store camera images and science data for later playback when Earth was not visible. These devices worked but occasionally caused problems (including mechanical wear on Voyager’s recorders decades later). True drum memory, however, was never used in any flown spacecraft computer. McCarthy’s caution proved well-founded once solid-state memory became practical in later missions. McCarthy is pointing out that many “experiments” are really applications on top of the same core capabilities: sensing, manipulation, and vision. That framing mirrors early time-sharing and OS design: build shared system services (I/O, scheduling, device control) once, then let many programs reuse them, instead of hardwiring each experiment as a one-off. McCarthy observes that if the Beagle operates for a full year, early experimental data will naturally suggest changes to later observations. This advantage is possible only with a stored-program computer that can be reprogrammed from Earth. This foresight was fully realized in the Voyager missions launched in 1977. The spacecraft flew pre-set trajectories. Yet ground controllers uploaded thousands of new command sequences after each flyby. Discoveries such as Io’s active volcanoes and Jupiter’s faint ring system led to completely revised imaging plans for the later encounters. Different filters, exposure times, pointing directions, and entirely new targets at Saturn, Uranus, and Neptune were used. For instance, close-up pictures of Neptune were never part of the original mission plan. New logic was uploaded to rotate the entire spacecraft slowly during exposures so that long-exposure images of the planet and its rings would remain sharp despite the spacecraft’s motion. The result was far richer scientific returns than any fixed pre-launch program could have achieved. 
*Close picture of Neptune - only made possible by software uploaded to the Voyager probe after launch* In June 1964 these targets were ambitious yet within the realm of active research. Ground-based scientific computers like the IBM 7090 offered a 2.18 µs memory cycle, up to 32 768 words of 36-bit core memory, and add times in the same range as McCarthy’s 2 µs, but they filled entire rooms and consumed tens of kilowatts. No spacecraft had yet flown a stored-program computer at all. The Apollo Guidance Computer, then in early design, would reach flight with an 11.7 µs cycle time, only 2 048 words of erasable memory plus 36 864 words of fixed ROM, and 55 W power consumption - already considered a major engineering triumph. McCarthy’s goals of a 1 µs cycle, 2 µs add, 10 µs floating multiply, 40 W total power, and 130 144 words of 48-bit memory (roughly 780 KB) were therefore pushing the limits of what seemed feasible for a radiation-hardened, one-year Mars mission. In practice they proved challenging to meet for the era. The Viking landers that finally reached Mars in 1976 flew dual Honeywell 24-bit computers with only 18 KB of plated-wire memory each, smaller and slower than McCarthy had hoped. The gap between laboratory performance and space-qualified hardware would narrow over the following decades. McCarthy’s proposed software structure is a remarkably accurate blueprint for the modular operating-system architecture that would dominate computer science for decades. By separating the core “platform logic” of the machine from the specific “tasks” it performs, it anticipates several foundational principles of modern systems. **Kernel and scheduler:** the “time-sharing executive” acts like an OS kernel scheduler, allocating CPU time across programs and ensuring critical system work (like thermal control and comms) keeps running even if an experiment program misbehaves. **Hardware abstraction and drivers:** putting device operations into shared subroutines with built-in safety checks is an early version of device drivers and a hardware abstraction layer. The goal is to protect motors, sensors, and actuators from being damaged by buggy experiment code. **Separation of concerns:** splitting “housekeeping” (system services) from “experiments” (applications) creates a clean boundary so domain experts can write specialized programs without also managing low-level realities like power budgets, telemetry, or mobility control.