From NES-4021 to moSMB3.wmv: Speedrunning the Serial Interface

Patrick LeMieux's article examines the serial interface of the Nintendo Entertainment System controller as a foundation for understanding speedrunning and tool-assisted speedrunning. The core argument is that games repeat at multiple levels, from industrial structure down to the microtemporal mechanics beneath conscious perception. The NES-4021 shift register converts button presses into a timed stream of electrical pulses, roughly 108 microseconds of serial traffic every sixtieth of a second, and the author treats these hidden serial structures as the terrain that tool-assisted play exposes and manipulates.

The analysis centers on Morimoto's 2003 Super Mario Bros. 3 speedrun video, moSMB3.wmv, which spread virally through torrents and online communities. Early reactions framed his skill in racialized, "alien" terms, a response the author reads as evidence that race functioned as a mode of interface. Viewers later discovered that the run was produced in the Famtasia emulator through frame-by-frame input sequences and save-state rerecording, reportedly exceeding 40,000 iterations over roughly two years. This revelation exposed a gap between human and machine temporality and helped popularize tool-assisted speedrunning through community efforts such as Yliluoma's NESvideos and TASVideos.org.

The study's scope encompasses the NES hardware, the Famtasia emulator and its movie file format, and the speedrunning communities that formed around these tools in the early 2000s. Its analysis is primarily technical and theoretical, drawing on Denson and Jahn-Sudmann's concepts of seriality. The author also documents how tool-assisted runs were demonstrated on physical hardware using automated "NESbots," confirming that emulator-designed inputs produce matching outputs on original consoles. The article further identifies a methodological caveat: TASVideos and SpeedDemosArchive use different timing conventions, which makes tool-assisted records appear roughly 3.27 seconds closer to human records than they are under real-time standards.