← The canon · AItopiaOrAImageddon?

The ARPANET's first nodes

moment · ARPA's Information Processing Techniques Office — Robert W. Taylor, who started it, and Lawrence G. Roberts, who ran it; the subnet built under contract by Bolt Beranek and Newman; the first four sites UCLA, Stanford Research Institute, UC Santa Barbara and the University of Utah · 1969

Something that happened and changed what people expected next.

Read on: The TCP/IP cutover — federation without a centre.

Filed as moment, and the filing needs an argument, because this entry passes one house test for the kind and strains the other. build.py states the kind as "Something that happened and changed what people expected next" — and against that, 1969 is not even a close call. The stricter test the entries have accreted is narrower. eliza-1966 talked its way out of moment and in doing so fixed the definition as a date on which something visibly happened in public; llama-weights-2023 reused that wording, darpa-grand-challenge-2005 tested every clause of it, and unix-linux-1969-1991 failed the "in public" clause and went to idea instead. The two definitions have quietly drifted apart, and this entry sits exactly in the gap between them. It is worth saying so once, plainly, rather than picking whichever test gives the answer I want: the builder's definition and the entries' definition are not the same definition, and a future entry will have to reconcile them.

On the date and the clock, this is the best-documented event in the canon. At 22:30 on Wednesday 29 October 1969, Charley Kline, a UCLA programmer working under Leonard Kleinrock, sat at an SDS Sigma 7 in room 3420 of Boelter Hall and typed toward an SDS 940 at Stanford Research Institute in Menlo Park, with Bill Duvall on the other end and a telephone open between them. He got the L and the O away and the SRI host crashed. The full login worked about an hour later. The record is a line in the UCLA IMP Log, in Kline's hand: "Talked to SRI Host to Host." There is a room, a clock, two named people, two named machines, and a contemporaneous handwritten log. Nothing else in this canon is documented that tightly.

The "in public" clause is the weak one, and the honest answer is stranger than either yes or no. There was a public announcement. On Thursday 3 July 1969, four months before the first message, UCLA issued a press release written by Thomas Tugend under the headline "UCLA to be first station in nationwide computer network." It named the four sites, named the funder — "The ambitious project is supported by the Defense Department's Advanced Research Project Agency (ARPA)" — and closed with a quotation from Kleinrock that turned out to be one of the better technology predictions of the century. So the thing was announced, in public, on the record, before it happened. And by Kleinrock's own later account there was "absolutely zero interest at the time." Unix in 1969 was invisible because nobody was told. The ARPANET in 1969 was invisible because everybody was told and nobody cared. Those are different failures, and only the second one is compatible with moment, which is why this entry can take the kind that unix-linux-1969-1991 had to refuse.

Why not idea. This is the serious rival and it has two candidate mechanisms, which is one more than it needs to fail. The first is packet switching — but that is not this id's to claim, its priority is genuinely disputed (section 4), and its inventors were at RAND and the UK's National Physical Laboratory, not on the ARPANET. The second is the ARPA funding model itself: a mission agency, program managers with signature authority, money placed on institutions for long horizons with no proposal ritual. That is a real mechanism, it is implementable — it has been implemented repeatedly since, which is what ARPA-E and ARIA are — and an idea entry on it would be legitimate and this canon should probably have one. But it would be a different entry with a different id, and this id names a network coming up. Networks coming up are events. darpa-grand-challenge-2005 reached the same fork over inducement prizes and turned the same way, and consistency between two entries about the same agency is worth something.

Why not prediction. The 1969 nodes are load-bearing for an unusual number of dated, gradeable claims, and section 3 grades six of them. But the event is not a claim; it is the thing the claims were about. The forecasts are graded inside a moment rather than promoted to the file's kind, exactly as darpa-grand-challenge-2005 handles DARPA's own forecasts.

Why not limit. Nothing was proved. Four machines exchanged bits over leased lines. Section 4 exists because that modest fact is routinely made to carry conclusions about survivability, about government, and about the private sector, none of which follow from it.

---

The proposal line, corrected

canon/proposals.md reads: "arpanet-1969 — the first ARPANET nodes, funded by the agency Sputnik created. Cite when: public research funding's role in a capability is disputed — the line from a 1957 panic through ARPA to the network and to most of early AI is one of the few clean causal chains this canon can point at. Descends from sputnik-1957 if that is taken. the owner, 2026-08-17."

The "cite when" is right, is the reason this entry earns admission, and section 2 is built on it. Two things in front of it need correcting, and one of them is the most interesting fact in the file.

"The agency Sputnik created" is true, and "clean causal chain" is not. ARPA was indeed created in the shadow of the launch: Sputnik went up on 4 October 1957; Secretary of Defense Neil McElroy issued DoD Directive 5105.15 on 7 February 1958, four months and three days later; the first director was Roy Johnson, an executive out of General Electric. But ARPA was created to run America's space programme, and it lost that mission almost immediately. Executive Order 10783 made NASA the controlling agency for non-military space on 1 October 1958, taking ARPA's lunar probe with it; TIROS went to NASA in 1959; man-in-space went to NASA in 1959; and during 1960 the remaining civilian space work went to NASA and the military space work went to the individual services. What was left was an agency two years old with most of its founding portfolio confiscated, which then concentrated on ballistic-missile defence, nuclear test detection, counterinsurgency — and, newly, computer processing. The Information Processing Techniques Office was established in 1962, with J. C. R. Licklider as its first director from October of that year.

So the chain is real but it is not clean. It runs: panic → agency → agency stripped of the mission the panic created it for → agency casts about for work → computing. The network exists partly because ARPA needed something to be for. A reading that cites this entry for "national rivalry funds capability" is entitled to; a reading that cites it for a tidy relationship between the fear and the payoff is citing something that did not happen. The best short statement of the real relationship is that the panic bought an institution, and the institution — not the panic — bought the network, thirteen years later, for reasons that had nothing to do with satellites.

descends_from is empty, and that is a finding rather than a gap. sputnik-1957 is proposed in proposals.md and has not been written, so there is no id to point at and I am not inventing one; per the house rule it is named here in prose instead. Two other unwritten proposals are relatives rather than ancestors and are named for the same reason. tcp-ip-1983 is this entry's child, not its parent, and section 4 explains why keeping them apart matters. von-neumann-singularity-1958 shares nothing with it but a decade.

The entry I most want an edge to and cannot have is sputnik-1957, and when that entry is written this one should be revisited, because the correction above belongs on the edge between them rather than inside either file.

---

What it is

The office, and the man who defined the job

IPTO's budget went from about $9 million in 1962 to $23 million in 1970, and sat at $15 million in 1965. Those are not large numbers even for 1965. What made them decisive was concentration and duration: by the early 1970s IPTO accounted for most of the Defense Department's basic research and about half of its applied research in computing, and it spent that money on a short list of places — MIT, Carnegie Tech, Stanford, SRI — for years at a time. Bob Taylor's description of the method is the one worth keeping: "I don't remember a single case where we ever funded a single individual's work." The office bought institutions, not projects.

Licklider had arrived with a thesis rather than a programme. He believed the computer's future was as a partner in thought and, from 1962, he was in a position to pay for that belief. The consequence for this project's subject is recorded in the National Research Council's history: the establishment of IPTO "radically changed the scale of research in AI, propelling it from a collection of small projects into a large-scale, high-profile domain," and "DARPA provided the bulk of the nation's support for AI research" for three decades after. The specific sums matter for section 2: the Speech Understanding Research programme took $3 million a year from 1971 to 1976; the Strategic Computing Program committed $1 billion over ten years from 1983; federal AI funding including SCP ran $105.6 million in 1984 and $274 million in 1988, of which the Defense Department was 73–92% of the applied total. The same office paid for the network and for the field this project watches. That is the entry's central fact and it is why an id about a network belongs in an AI canon at all.

Taylor's decision, and what it was actually for

Taylor took IPTO in 1966 and started the network that February. His own account, written in 2004, is short and has the merit of naming the decision-makers: "In February of 1966 I initiated the ARPAnet project . . . There were only two people involved in the decision to launch the ARPAnet: my boss, the Director of ARPA Charles Herzfeld, and me. The creation of the ARPAnet was not motivated by considerations of war. The ARPAnet was created to enable folks with common interests to connect with one another through interactive computing even when widely separated by geography."

BBN's own history of the decade states the goals in the flattest possible terms: "to interconnecting computers in such a way that a very broad class of interactions are possible" and "to improve and increase productivity through resource sharing." And it states the acceptance criterion, which is the sentence that actually describes what was hard: "The most important criterion for the type of network interconnection desired was that any user or program on any of the networked computers be able to utilize any program or subsystem available on any other computer without having to modify the remote program."

Read that last clause slowly, because it is the part that generalises. Not move the data. Not dial the machine. Use a remote program you did not write and cannot change. Every argument in 2026 about tool use, about model endpoints, about an agent calling a service it has no source for, is a descendant of that acceptance criterion, and it was written down as a procurement requirement before any of it existed.

The manifesto, published while the procurement was running

In April 1968, in Science and Technology, Licklider and Taylor published "The Computer as a Communication Device." It opens: "In a few years, men will be able to communicate more effectively through a machine than face to face." It is nineteen pages long, it has cartoons, and it is the most accurate piece of technological forecasting this canon has yet examined. Section 3 grades it. What matters here is that it was written by the man who founded the office and the man who started the network, eighteen months before the first message, and that it describes what they thought they were buying.

The procurement

BBN's history gives the sequence and I take it as the record. The RFQ "for the network was mailed out" in July 1968. "Twelve proposals were received" and "four bidders were rated within the zone of contention." "One was chosen in the week before Christmas 1968" — Bolt Beranek and Newman of Cambridge, Massachusetts, Frank Heart's group — and "work began the second day of the New Year in 1969." The contract is widely reported at about $1 million; BBN's own history states no dollar value and I have not found the figure in a primary, so it is carried here as secondary.

The machine BBN built was the Interface Message Processor, a ruggedised Honeywell DDP-516 that sat between a site's own computer and the leased lines. Its significance is architectural and is easy to miss: the IMP meant a site did not have to agree with any other site about anything except how to talk to its own IMP. The subnet was a contractor's problem; the hosts stayed heterogeneous. That separation is why the press release could promise to "link together computers of different makes and using different machine languages," and it is the same move as an API boundary.

RFC 1

On 7 April 1969, Steve Crocker at UCLA wrote RFC 1, "Host Software." It begins by stating the split — "The software for the ARPA Network exists partly in the IMPs and partly in the respective HOSTs" — and then, more consequentially, it describes a group with no authority: "During the summer of 1968, representatives from the initial four sites met several times to discuss the HOST software and initial experiments on the network," after which Steve Carr of Utah, Jeff Rulifson of SRI and Crocker of UCLA "met during the fall and winter." Graduate students, writing numbered notes, deciding the protocols, because nobody senior had told them not to. The document series they started is still running.

The nodes

IMP number one arrived at UCLA at the end of August 1969, on a Saturday, and was plugged in on the loading dock; Kleinrock's account is "To BBN's credit, they connected this thing up, and that day we had bits moving back and forth." The host interface that joined UCLA's Sigma 7 to it had been built by a graduate student, Mike Wingfield, for about $4,000 in under six weeks, against an outside contractor's bid of $19,000 and six months — a detail worth keeping because it is the entire economics of the project in miniature. IMP two went to SRI on 1 October, three to UC Santa Barbara in November, four to the University of Utah in December. Four nodes by the end of the year. Fifty-odd by 1975, "spanning Hawaii to Norway."

November 1969, which is the part nobody tells

In the same month the third node came up, Congress closed the funding regime that had produced all of it. Section 203 of Public Law 91-121, the Mansfield Amendment, enacted November 1969 and effective for FY1970, required that Defense-funded research bear "a direct and apparent relationship to a specific military function or operation." IPTO's whole method — buy institutions, wait years, do not ask what it is for — became legally awkward in the exact season its best product was being installed. ARPA became DARPA in 1972. Program managers learned to write military justifications. Some of the people who would have stayed in universities went to Xerox PARC instead, Taylor among them.

The chain that this entry exists to document was already being cut when the first packets moved. Any reading that cites 1969 for the power of patient public funding has to carry that too, or it is citing half a fact.

---

Why a reading would cite it

The cite-when is right: use this entry when public research funding's role in a capability is disputed. It is the strongest case the canon can point at, and the reason it is strong is that the counterfactual is unusually tight. No customer wanted a computer network in 1966. No firm was building one. The demand had to be manufactured by an agency that paid for the machines, paid for the lines, paid the contractor, and paid the four sites to accept a box on the loading dock. darpa-grand-challenge-2005 is the canon's other entry of this shape, and the two together make a better argument than either alone: a government agency created a capability twice, forty years apart, in fields where the private return did not yet exist.

The concrete 2026 case. The dispute is live and it has numbers. Per the Congressional Research Service report of 6 January 2026: NSF was funded for FY2025 at the FY2024 level of $9.1 billion; the Administration's FY2026 request was $3.9 billion — a $5.2 billion cut, −56.9%, with the main research account down 58.4%; the Senate committee marked $9.0 billion on 17 July 2025 and the House committee $7.0 billion on 12 September 2025; and the CHIPS and Science Act had authorised $17.8 billion for that year, "which is $13.9 billion (357%) more than the FY2026 request." Trade coverage reports the final enacted figure at $8.75 billion, signed 23 January 2026 in a minibus, with Research and Related Activities at $7.2 billion — a much smaller cut than requested, and I flag that the enacted number is from secondary reporting because the CRS report predates it.

But the number that makes this entry worth citing is not the topline. It is which line survived. The same request that proposed cutting NSF by 57% named three protected "critical activities": artificial intelligence, quantum information science, and the Directorate for Technology, Innovation, and Partnerships. So the 2026 dispute is not public money versus no public money for AI. It is whether the mechanism that produced the ARPANET — broad, undirected, long-horizon funding of institutions, by an office allowed not to say what it was for — is what survives, or whether only the named-priority, justify-the-application part survives. That is the Mansfield Amendment question, put again, fifty-seven years later, and this entry is where a reading can show that the question has a documented precedent and a documented answer. In 1969 the answer was that the named-priority regime won, and the people who did the undirected work left for the private labs.

The magnitude, recorded and not editorialised. NSF's entire research account for FY2026 is $7.2 billion. Analyst compilations put the combined 2026 capital expenditure of the largest cloud and AI operators somewhere in the $700–800 billion range — a figure I mark explicitly as an analyst estimate rather than a filing, per the house rule, and which should be treated as an order of magnitude and not a measurement. The ratio is about a hundred to one. Whatever role public funding plays in AI capability in 2026, it is not the role it played in 1969, when ARPA was effectively the only buyer in the room.

The honest limit on the argument, which a reading must carry. This entry is an existence proof, not a mechanism, and the difference is the whole of its usefulness. It establishes that public funding has produced a foundational capability that no private actor was going to produce. It does not establish that public funding is sufficient, that it is repeatable on demand, or that the current frontier would exist faster with more of it. The chain also has a genuinely mixed tail: the same office that funded the network funded the AI programmes that ended in the collapses recorded in lighthill-1973 and expert-systems-collapse-1987, and it funded them with the same confidence. An entry that cited the network and skipped the winters would be selecting on the outcome. ARPA's hit rate is a fact this canon does not have, and until it does, 1969 should be cited as a thing that happened and not as a policy.

One more use, for the concentration lens. Licklider and Taylor asked the lens's central question in 1968, in almost the words the lens uses now: "For the society, the impact will be good or bad, depending mainly on the question: Will 'to be on line' be a privilege or a right? If only a favored segment of the population gets a chance to enjoy the advantage of 'intelligence amplification,' the network may exaggerate the discontinuity in the spectrum of intellectual opportunity." When a reading needs to show that the access question is not a recent anxiety invented by the current cycle, that sentence is the citation, and it was written by the two men who were spending the money.

---

What it got right, and what it got wrong

Not required for moment, but this entry sits on top of an unusual density of dated claims by the people doing the work, and grading them is most of its value. All page numbers are from the reprint of Licklider and Taylor read for this entry.

1. "In a few years, men will be able to communicate more effectively through a machine than face to face." (April 1968, p. 21.) They gave themselves a schedule, which is rarer than it should be: "It will take two years, at least, to bring the first interactive computer networks up to a significant level of experimental activity. Operational systems might reach critical size in as little as six years if everyone got onto the bandwagon . . . So let us take six years as the target." Due: April 1970 for experimental activity, April 1974 for operational critical size. Verdict: the first half is a hit, slightly early — the first message came eighteen months later and four nodes were up in fourteen. The second half is a clear miss. By 1974 the ARPANET was tens of nodes of federally funded research sites. "Critical size" in the sense they meant — enough people on line that the medium beat the alternative — arrived in the mid-1990s, roughly twenty years late. The pattern is the standard one for this canon: right about the thing, wrong about the adoption curve by an order of magnitude in time.

2. The cost model. They costed a communicator-hour at "just less than $11" and, correcting the duty-cycle assumption, at $16 per hour in 1968 dollars, then argued it down on the expectation that processing and storage costs halve every two years. Verdict: right on the trend, and — more impressively — right about which input to worry about. They singled out transmission: "the cost of transmission is not falling nearly as fast as the cost of processing and storage . . . at the rate things are going, in six years it will be the dominant cost. That prospect concerns us greatly and is the strongest damper to our hopes." That was the correct worry, it stayed correct for about twenty-five years, and it is the direct reason the 1974 target was missed. A forecast that misses its date but correctly names the variable that will cause the miss is a better forecast than one that gets the date right by luck, and the canon should score it that way.

**3. "They will be communities not of common location, but of common interest."** (p. 38.) Hit, and the most consequential one. They also drew the conclusion: "First, life will be happier for the on-line individual because the people with whom one interacts most strongly will be selected more by commonality of interests and goals than by accidents of proximity." The mechanism was right and the valence was not. algorithmic-monoculture-2022 is the entry where the cost of sorting people by common interest instead of common location is examined; read against it, this sentence is a correct prediction of a structure and a wrong prediction of its effect, which is a distinction worth keeping because it is the failure mode of nearly all optimistic technology forecasting.

4. OLIVER. (pp. 38–39.) This is the passage that makes the paper matter to this project rather than to a history of networking, and it is worth quoting at length because paraphrase makes it sound vaguer than it is:

> A very important part of each man's interaction with his on-line community will > be mediated by his OLIVER. The acronym OLIVER honors Oliver Selfridge, > originator of the concept. An OLIVER is, or will be when there is one, an > "on-line interactive vicarious expediter and responder," a complex of computer > programs and data that resides within the network and acts on behalf of its > principal, taking care of many minor matters that do not require his personal > attention and buffering him from the demanding world. "You are describing a > secretary," you will say. But no! Secretaries will have OLIVERS. > > At your command, your OLIVER will take notes (or refrain from taking notes) on > what you do, what you read, what you buy and where you buy it. It will know who > your friends are, your mere acquaintances. It will know your value structure, > who is prestigious in your eyes, for whom you will do what with what priority, > and who can have access to which of your personal files. . . . > > Some parts of your OLIVER program will be common with parts of other people's > OLIVERS; other parts will be custom-made for you, or by you, or will have > developed idiosyncrasies through "learning" based on its experience in your > service.

Written in April 1968. A network-resident personal agent; a shared base with per-user customisation; behaviour that diverges through learning on your own interaction history; an explicit access-control model over your files. Verdict on content: a hit, and an uncomfortably precise one. Verdict on timing: ungradeable — they attached no date to it, and an entry that pretended otherwise would be manufacturing a score. The part they got wrong is not technical, it is proprietary. Every sentence assumes the OLIVER is yours: your OLIVER, custom-made by you, learning in your service, enforcing your access rules. What shipped is a service, running on someone else's machines, whose value structure is configured by its vendor and whose logs are the vendor's asset. They predicted the artefact and inverted the ownership, and that inversion is the single most citable thing in this file for the concentration lens.

5. "Unemployment would disappear from the face of the earth forever." (p. 40.) The full sentence: "Unemployment would disappear from the face of the earth forever, for consider the magnitude of the task of adapting the network's software to all the new generations of computer, coming closer and closer upon the heels of their predecessors until the entire population of the world is caught up in an infinite crescendo of on-line interactive debugging." Verdict: wrong, and it is only fair to add that it is plainly written as a flourish — the paragraph before it is explicitly hedged as hope rather than plan, and nobody should grade it as a labour-market forecast. What earns it a place here is which job they picked. They assumed the inexhaustible human work would be maintaining and adapting software, and that is precisely the occupation whose exposure is the live question in 2026. frey-osborne-2013 and simon-1965 are where this canon keeps its automation forecasts; this is the entry that shows the same error being made in the opposite direction, by people who could not imagine the machine doing the debugging.

6. Kleinrock, 3 July 1969, and a self-grade worth recording. The press release closes: "As of now, computer networks are still in their infancy. But as they grow up and become more sophisticated, we will probably see the spread of 'computer utilities' which, like present electric and telephone utilities, will service individual homes and offices across the country." Verdict: a hit, including the metered-utility framing, which is a fair description of how compute is now sold.

The canon's rule for prediction entries — where the predictor has graded themselves, record their grade and an independent one, and say they differ — is not binding on a moment, but this case is instructive enough to apply it anyway. In 2005 Kleinrock graded himself in print. He decomposed the 1969 vision into five elements — the technology will be everywhere; always accessible; always on; anyone will be able to plug in from any location with any device at any time; and it will be invisible — and scored it: "The Internet almost got it right. Indeed, the first three elements have already come about. However, the Internet as we know it today has not yet achieved the last two elements of the vision." Three out of five, by his own hand, in 2005.

An independent grade in 2026 differs in both directions, and the direction of the difference is the finding. Element 4 he under-scored: he attributed the gap to TCP/IP's assumption that a user, a device and an address stay coupled, and that gap was substantially closed by mobile networking in the decade after he wrote. Element 5 he over-scored by conceding it: "invisible" is further away now than in 2005, because the intervening business model made the network the most conspicuous surface in ordinary life rather than a utility you forget about. Net: four out of five on the technology, and a miss on the element he thought was merely lagging. Note the asymmetry against the base rate the canon collects this kind of grading for — the house expectation, set by kurzweil-2005, is that self-assessments run generous. Kleinrock's ran conservative. One case is not a base rate, but a base rate assembled only from the generous cases would be wrong too, and this is the counter-example.

7. What none of them predicted: who would own it. Across the 1968 paper, the 1969 press release and the 1981 BBN history, the question of who owns the network is essentially absent — the closest anyone comes is Licklider and Taylor's privilege-or-right sentence and a passing note that "in a commercial system, the remote processor would at this point record cost information for use in billing." They were building a public utility with public money for public institutions and the ownership question did not look like a question. It is now the question. llama-weights-2023 is where this canon argues it; 1969 is where a reading can show it was not asked.

8. BBN's own claim, 1981. "The success of the program far exceeded even the most optimistic views at the time of inception." Verdict: true, and unusually so for a contractor's self-assessment of its own contract.

---

Commonly misused as

Not required for moment, but this is the entry's second job. The 1969 nodes are made to say four things they do not say, and the misuses run in opposite political directions, which is the tell that the event is being used rather than read.

1. "The internet was built to survive a nuclear war." The most durable technology myth there is. What is true: Paul Baran's RAND work of the early 1960s was about survivable communications, and BBN's own history records that RAND "was particularly concerned with the question . . . of survivability of a distributed grid network after heavy node and link failures." Packet switching partly came out of that concern. What is false is the attribution of the motive to the ARPANET. The two men who made the decision have both denied it on the record. Charles Herzfeld, ARPA's director when it was approved: "The ARPANET was not started to create a Command and Control System that would survive a nuclear attack, as many now claim. To build such a system was, clearly, a major military need, but it was not ARPA's mission to do this; in fact, we would have been severely criticized had we tried." Taylor: "The creation of the ARPAnet was not motivated by considerations of war." And BBN's history states the goal as resource sharing and productivity, not survivability.

Why the myth costs something. It reframes the network's robustness as a requirement that was specified and met, when the actual requirement was let researchers use scarce machines they do not own and cannot modify. That second requirement is the one that generalises to 2026 — it is the shape of every argument about access to compute — and the myth hides it behind a more cinematic story. A canon that lets the nuclear version stand loses the only part of the 1969 design that is still live.

2. "The government invented the internet, therefore government can do it again." The overreach on the friendly side. Section 2 states the limit: this is an existence proof, not a mechanism, and it is selected on its outcome. The sharper objection is internal to the entry — the funding regime that produced the ARPANET was restricted by Congress in the same month the third node was installed (Mansfield Amendment, P.L. 91-121, November 1969). Whatever it was that ARPA had between 1962 and 1969, the United States chose not to keep it, and the argument "do that again" therefore has to specify which thing, in a political system that has already declined it once.

3. "Actually the private sector invented it." The overreach on the other side, and it has a canonical instance: **L. Gordon Crovitz, in the Wall Street Journal, July 2012, "Who Really Invented the Internet?", which assigned the credit to Xerox PARC. It was rebutted within days by Vint Cerf, by Xerox itself, and — most awkwardly — by Michael Hiltzik, whose history of PARC was Crovitz's main source and who said the book made the opposite argument. The document trail is not ambiguous: ARPA issued the RFQ, ARPA chose the bidder, BBN built to an ARPA specification, the sites were ARPA contractors, and Cerf and Kahn did TCP/IP under federal funding. The most efficient refutation is a person: Bob Taylor started the ARPANET at ARPA and then ran the Computer Science Laboratory at Xerox PARC.** The claim has to route through the one man who is on both sides of it, and he spent the rest of his life saying the government did it.

Carry both misuses together or neither. An entry that debunks only the convenient one is doing politics with a canon, and rule 7 is easier to keep when the file refuses both directions in the same paragraph.

4. "ARPANET" used as a synonym for "the internet." The commonest small error and the one most likely to corrupt a reading. The ARPANET is one network with a single owner, a single contractor and a single protocol. The internet is the federation of networks that can join without anyone's permission, and that property arrives with the TCP/IP cutover of 1 January 1983 — proposed in proposals.md as tcp-ip-1983 and not yet written, which is a real gap, because almost everything the phrase "the internet did X" is used to mean is a property of 1983 rather than of 1969. If a reading needs uncontainability, that is 1983 and it should say so. 1969 gets resource sharing and heterogeneous hosts, which are the properties that actually map onto 2026's arguments about model access.

5. A smaller one: "packet switching was invented at UCLA." The priority is genuinely disputed among Paul Baran (RAND, On Distributed Communications, 1964), Donald Davies (National Physical Laboratory, 1965–66, who coined the word packet) and Leonard Kleinrock (whose 1961 MIT work was queueing-theoretic). Davies contested Kleinrock's claim in work published after his death — "An Historical Study of the Beginnings of Packet Switching," The Computer Journal 44(3), 2001, pp. 152–162 — and the dispute has never been settled. This canon should not adjudicate it and this entry does not depend on it. What 1969 demonstrates is that a packet-switched network of heterogeneous hosts worked, which is a claim about UCLA, SRI, UCSB and Utah collectively and about none of them individually. A reading that needs the invention should cite the inventors' own papers; a reading that needs the demonstration can cite this.

---

Sources

Read directly, at source. Two primaries are image-encoded PDFs and were read page by page as rendered images; every quotation attributed to them is from the document, not from a secondary account. Fetch failures are recorded at the end rather than papered over.

Secondary, and labelled as such.