At about half past ten at night on 29 October 1969, a student programmer at UCLA named Charley Kline sat at a terminal and tried to log in to a computer 560 kilometres away at the Stanford Research Institute.
He typed L. A colleague at the other end, on the telephone, confirmed they had received L. He typed O. They had O. He typed G, and the machine at Stanford crashed.
So the first message ever sent from one computer to another over a packet switched network was LO. They fixed it and completed the login about an hour later. Kline’s entry in the log book for that night reads, in full, that he talked to SRI host to host.
What Was Actually Being Tested
The hardware consisted of Interface Message Processors, or IMPs, built by the Cambridge firm Bolt Beranek and Newman out of ruggedised Honeywell minicomputers. Each one weighed about four hundred kilograms and sat beside the host computer it served, handling the network so the host did not have to.
The first IMP was delivered to Leonard Kleinrock’s laboratory at UCLA at the end of August 1969, and the second to Doug Engelbart’s group at Stanford Research Institute at the start of October. The link between them ran at fifty kilobits per second over a leased telephone line.
By December there were four nodes: UCLA, SRI, the University of California at Santa Barbara and the University of Utah. That was the whole network.
There was no press. Nobody photographed it, nobody had prepared a sentence for history, and the people present were mostly graduate students who wanted to know whether the thing worked. Kleinrock has pointed out since, with some amusement, that the same year produced one small step for man, and his own field’s founding utterance was two letters cut off by a crash.
The Nuclear War Story Is Wrong
The most repeated claim about the ARPANET is that it was built to survive a nuclear attack, with a decentralised design so that the network would keep working after cities were destroyed.
The people who built it have denied this repeatedly and specifically. Bob Taylor, who started the project at ARPA in 1966, said flatly that it had nothing to do with supporting or surviving war. Larry Roberts, who ran the construction, said the same.
What they were actually solving was a resource problem, and it was a mundane one. In 1966 ARPA was funding computing research at a dozen or so universities, each of which wanted its own expensive machine. The machines were incompatible, most of them were idle much of the time, and researchers at one site could not use software that existed at another. Taylor had three separate terminals in his office, one for each remote system he needed, and each with its own login procedure.
The ARPANET was built so that researchers could share computers and programs. That is the whole original purpose.
The confusion has a real root. Paul Baran at the RAND Corporation had studied survivable communications between 1960 and 1964, explicitly with nuclear attack in mind, and he invented distributed packet switching as the answer. His work fed into the ARPANET’s design. But the network that was built, and the money that built it, were about sharing scarce computers, and the engineers doing it said so at the time.
Packet Switching, and Why It Was Not Obvious
The technical idea underneath is the one that mattered, and in 1965 it was regarded as slightly mad by the people who ran telephone networks.
A telephone call is circuit switched. When you dial, the network reserves a continuous path from one end to the other, and that path belongs to you for the duration, whether you are talking or silent. It is a perfect model for a conversation and a terrible one for a computer, because a terminal session is mostly silence punctuated by short bursts.
Packet switching breaks the data into small pieces, each carrying its destination address, and sends them into the network to be forwarded hop by hop. Packets from many conversations share the same lines, interleaved. They can take different routes, arrive out of order, and be reassembled at the far end. If a link fails, subsequent packets go another way.
Three people arrived at this independently. Baran at RAND had it by the early 1960s. Donald Davies at the National Physical Laboratory in Britain worked it out in 1965 and gave it the name packet. And Kleinrock had done the mathematical analysis of message flow in such networks at MIT in 1961 and 1962, applying queueing theory to work out the delays.
The established telecommunications engineers thought it would not work. AT&T, approached by Baran, declined on the grounds that the idea was unsound and would interfere with the telephone network. Their objection was not stupid: everything they knew about carrying traffic assumed a reserved path, and a scheme in which pieces of a message wander off separately and get put back together sounds like a recipe for chaos.
It works because of what Shannon had established in 1948: that information can be treated as a quantity independent of the medium carrying it, and that with the right coding you can transmit reliably over an unreliable channel. Packet switching is that principle turned into plumbing.
The Invention Nobody Mentions: The RFC
There is a second invention from 1969 that has been at least as consequential as the hardware, and it is a document format.
The graduate students working on the protocols needed to agree on how the machines would talk to each other, and nobody had the authority to decide. Steve Crocker, then a student at UCLA, wrote up the first set of notes in April 1969 and chose a deliberately unassuming title to avoid implying that any of it was settled: Request for Comments.
RFC 1 is about host software. There are now over nine thousand of them, and the entire technical specification of the internet, including the protocols your device used to fetch this page, exists as numbered RFCs.
The choice of that name set a working culture in which a specification is a proposal open to argument, adopted because it works and because people implement it rather than because an authority mandates it. That is why the internet has no owner, and it is a governance decision made by a graduate student who did not want to seem presumptuous.
What Came Next, Quickly
- 1971. Ray Tomlinson at BBN sends the first message between users on different machines, and needs a character to separate the user from the host. He picks the one on the keyboard that nobody was using for anything else: @
- 1973. The network reaches Europe, to NORSAR in Norway and University College London
- 1974. Vint Cerf and Bob Kahn publish the design for a protocol to join dissimilar networks together, which becomes TCP and then TCP/IP
- 1 January 1983. The ARPANET switches over to TCP/IP on a single flag day. This is the better candidate for the birthday of the internet, because it is the point at which the thing stops being one network and becomes a way of connecting any network to any other
The ARPANET itself was formally shut down in 1990, by which time it had been made redundant by the networks built on top of its ideas.
The Interesting Thing About the Crash
The Stanford machine fell over after two characters because of a bug in how it handled the arriving login sequence, which was fixed that evening. It is remembered as a charming detail.
It is also the first demonstration of the property that the whole design turned out to depend on. The network delivered the data correctly. The host at the far end failed. Those are separate things, and keeping them separate, so that the network makes no promises about what the endpoints do and the endpoints take responsibility for their own reliability, became the organising principle of internet architecture.
The stored program computer had been specified in 1945 and built in the following few years, and for two decades each one sat alone in its room. Everything that has happened since, including every system trained on text scraped from millions of machines at once, depends on the evening somebody managed to get two letters from Los Angeles to Menlo Park before the other end gave up.
No products in the cart. 