← Back to Home
ON THIS DAY 2026-10-29

1969: The first message crosses the ARPANET

On the evening of 29 October 1969, a graduate student named Charley Kline sat at a terminal at UCLA and tried to log into a computer at the Stanford Research Institute, 560 kilometres away. He typed "L", then "O" — and the system crashed before he could finish "LOGIN". A phone call confirmed the two letters had arrived. An hour later, the full message went through. It was an inauspicious but real birth: the first host-to-host transmission on the ARPANET, the packet-switched network that would become the ancestor of the internet.

The project was funded by the Advanced Research Projects Agency (ARPA), the same Pentagon office that, through the 1960s, was quietly underwriting much of American artificial intelligence research — at MIT, Stanford, and Carnegie Mellon. ARPA's visionary program officer, J.C.R. Licklider, had argued years earlier in his essay "Man-Computer Symbiosis" that computers should extend human thought, and had sketched a "Galactic Network" connecting researchers everywhere. For Licklider, networking and intelligent machines were two sides of the same ambition: to let minds, human and eventually artificial, share and amplify information rather than work in isolation.

That ambition is easy to underestimate from today's vantage point. In 1969, computers were room-sized, rare, and incompatible with one another. Getting two of them to exchange a single character required inventing new protocols, new hardware (the Interface Message Processors), and new ways of thinking about computation as something distributed across machines rather than confined to one. The crash on that first attempt was almost poetic: the infrastructure of connection had to be built through repeated, patient failure, much like the neural networks and training runs of later AI research.

The thread from that stuttering "LO" to contemporary AI is direct, if often invisible. The ARPANET's descendants gave AI researchers the means to share code, datasets, and results across continents; later, the World Wide Web and cloud computing made it possible to assemble the vast corpora of text and images that train today's large language and vision models. The distributed, networked architecture pioneered in 1969 is now the circulatory system of machine learning — every model trained on scraped web data, every API call to a hosted AI system, traces back to that October evening when two computers, however briefly, managed to speak to each other.