War at the Speed of Thought
“And the battle was there scattered over
the face of all the country: and the wood devoured more people that day than
the sword devoured.”
— 2 Samuel 18:8
That verse sits at the head of a 1970 United States government report
titled Forest Fire as a Military Weapon. The men who wrote it — Forest
Service scientists, working under an order from the Advanced Research Projects
Agency — were describing how to burn the jungles of Vietnam to the ground, and
they reached, on the first page, for scripture. The forest of Ephraim, where
the wood killed more men than the sword. They knew exactly what they were
building, and they gave it a liturgy. Hold that gesture; the whole of what
follows is a variation on it.
The Expanding Battlespace — When Reach Exceeds Judgment.
In 1968 Nigel Calder gathered sixteen scientists from six countries and
asked them to forecast the weapons of the wars to come. Unless Peace Comes
is what they produced, and its most durable chapter is not the one on wrecking
the environment, nor the one on poisoning a battalion with hallucinogens in its
morning coffee. It is Harvey Wheeler’s, on computers — and it reads less like a
prediction about machines than a prophecy about time.
Wheeler’s observation was that a philosopher thinks in generations and a
statesman in a decade, but a missile-warning system resolves its entire world
in three millionths of a second. In that window the old buffer disappears — the
days a fleet’s approach once bought for a note of inquiry, a protest, a second
thought. War and politics collapse into what he called a simultaneous
synthesis: the crisis exists the instant the machine perceives it, and the
highest authority is forced to act with a field commander’s speed and a field
commander’s granularity.
Read the rest of the book through that chapter and a single tendency runs
the length of it. Essay by essay, the human being is reclassified — from master
of the battlefield to what the authors, with real coldness, call a grave
complication. We are too slow, too heavy, too fragile. And we hesitate. The
soldier is designed out; the decision is handed to whatever can keep the pace.
The Machinery Arrives
That was the forecast. Here is the arrival, and it comes not as a
manifesto but in the mildest prose imaginable — a Canadian defence journal, ON
TRACK, January 2025, an entire issue titled “National Security in the Age
of AI and Robotics.” No fever in it. Footnotes, acknowledgements to anonymous
reviewers, the measured cadence of people doing their jobs well. Which is
exactly what makes it worth reading closely.
The keystone essay asks how to catch the digital dragon — how a middle
power keeps pace in an arms race run in code rather than steel. Its authors are
serious people; one has spent a career in Canadian intelligence. They reach, as
everyone now does, for the “Oppenheimer moment,” and a footnote frets —
correctly — that the technology has outrun the arms-control instinct, raising
the risk of a war begun by accident or by machine miscalculation. Good. That is
the right thing to fear.
And then, a few pages on, describing the pressure China’s pace puts on
Western decision cycles, they write that the compression of time all but
eliminates the possibility of indecision, inaction, and issue-avoidance. They
mean it as a virtue. The tempo no longer permits a leader to wait, to sit on
his hands, to decline the question. The essay names the closing of that space
and files it under progress.
Hold that sentence too. Set it down beside the one from Second Samuel.
If the digital-dragon essay closes the space for hesitation, the next one
shows you the machine being built to fill it. Its author is a retired
brigadier-general who once ran Canada’s strategic signals intelligence and
cyber operations and served as vice director at United States Cyber Command —
the rare writer here who has actually stood watch. He is careful, and his care
is the point. Twice he insists that humans must remain “in the loop,” that
decisional processes must be guided by people and not corrupted by the systems
serving them. He is not naïve about what he is describing.
And yet, pages earlier, he lays out the technique with evident
admiration: inverse reinforcement learning that reads a satellite’s intent
off its behaviour — that infers, without a human asking, whether an object in
orbit is benign or hostile, whether an anomaly is a malfunction or a mask. The
machine performing the exact inference a watch officer once performed: is this
thing what it appears to be, or is it a threat. He wants the human in the loop
and describes, in the same essay, the apparatus for closing the loop without
one. The contradiction is not sloppiness. It is the whole predicament, sitting
inside a single honest author who can see both halves and cannot reconcile
them.
The third essay moves the logic from the satellite to the soldier. A
retired US Army officer argues that Western war is inherently multinational and
that robotics and AI must therefore be shareable across allies — a sensible
point about interoperability. But the ground under it is the reclassification
Calder’s authors forecast. The US Army, he reports approvingly, “will not trade
blood for first contact”; the front edge of future formations will be machines.
And because a lost robot carries none of the symbolism, cost, or grief of a
lost soldier, the willingness to send it into risk rises. The threshold to act
drops precisely because the thing acting is no longer a person. The grave
complication, designed out — now a line item in a procurement argument. To his
credit, he ends on a caution worth keeping: there are no wonder weapons, and
very few game-changers. Everyone is watching everyone; every edge is temporary.
It is the one moment in the collection where someone says, plainly, that the
machinery does not deliver what its momentum promises.
Then the last essay, and this is the one that should make you sit up,
because it is trying to solve the problem the others are creating. A team at
Defence Research and Development Canada has built a prototype called
Mockingbird, and its purpose is to tell whether a stream of text is human or
automated — to characterize a source, weigh whether a channel can be trusted,
sort genuine signal from machine-generated noise flooding the information
space. This is the receiver problem stated in plain defence language: not what
is being said, but whether the thing saying it is real.
And their own demonstration undoes them. They take one essay — mostly
machine-written, two human sentences buried inside — and hand it to six
commercial detectors. The results are in their Table 2. One tool calls the
whole thing AI. One calls the whole thing human. A third splits it. A fourth
puts it at seventy-two percent. Six receivers, one signal, no agreement — and
the human sentences hidden inside slip past nearly all of them. Mockingbird
exists because the machinery to judge the machinery is itself unreliable. The
institution is trying to automate the very act of discernment, and its opening
exhibit is a row of instruments contradicting one another over a single page of
text.
Four essays, then. The pace that forbids waiting. The inference that
needs no watch officer. The threshold that falls when no person is at risk. And
the discernment we can no longer perform, now handed to tools that cannot
agree. Set them side by side and a room takes shape with no window in it —
every wall a reasonable decision, made by a capable person, for a defensible
reason.
One Institution, Sixty
Years
Here is what the mild prose of that journal does not tell you: the room
was not built by accident, and it was not built recently. It was built by a
single institutional lineage, and the Forest Fire report is one of its
early bricks.
That same agency — ARPA, later DARPA — is the thread that runs through
the whole story. In 1960, its Behavioural Sciences office funded the man who
would become the patron saint of human-machine partnership, J. C. R. Licklider,
whose paper Man-Computer Symbiosis argued that the machine should do
“the routinizable work” while the human “set the goals, formulate the
hypotheses, determine the criteria, and perform the evaluations.” Licklider had
timed his own workday and found that eighty-five per cent of it went to getting
into a position to think — plotting graphs, chasing references.
Symbiosis was his answer: clear away the drudgery so the human is freed for the
one thing that matters, the judgment. He wanted to give the evaluator more
room.
The same agency then spent Vietnam taking the room back. It ran the
forest-fire research. It designed the McNamara Line — the electronic fence of
seeded sensors strung across the infiltration routes, chaired by the
geophysicist Gordon MacDonald, the very man who in Unless Peace Comes
had written the chapter on wrecking the environment. MacDonald’s governing
principle was elegant and terrible: the key to geophysical warfare is finding
the instability to which a small energy release triggers a vastly greater one.
A jungle primed to burn is exactly such an instability. So is a launch-warning
system primed to escalate.
And in 1969, the commander of the war said the quiet part into a
microphone. Speaking to the Association of the United States Army, retired
four-star General William Westmoreland described the fence as the threshold of
something new: “I see battlefields on which we can destroy anything we
locate through instant communications and almost instantaneous application of a
lethal firepower.” Locate, communicate, destroy — the interval squeezed
toward zero, announced not as a danger but as an achievement, from a podium, to
a room of former soldiers who were meant to applaud.
The doctrine that grew from that fence eventually got a name: the system
of systems — the networked lattice of sensors and shooters that Annie
Jacobsen, DARPA’s historian, calls the most revolutionary military technology
of the twentieth century after the hydrogen bomb. And here is the detail that
should stop you cold. Open the ON TRACK issue again. Its keystone essay,
describing what China has built and Canada must match, quotes the People’s
Liberation Army’s own doctrine for exactly this: a “network information
system-of-systems” that identifies vulnerabilities and strikes them at
machine speed. The phrase from Westmoreland’s fence, arriving in a 2025
Canadian defence journal as the future we must keep pace with. Not a metaphor
carried across the decades. The same words.
A system of systems is precisely the architecture that has no room for a
hesitating human. Its entire design intent — Westmoreland said it plainly — is
to compress the interval until the human interval vanishes. It is the
engineered absence of the pause. Which is the one thing that, twice in the age
of these machines, kept the sword sheathed.
The Pause
On the night of 26 September 1983, a lieutenant colonel named Stanislav
Petrov was the duty officer at Serpukhov-15, the bunker south of Moscow that
watched the sky for American missiles. Shortly after midnight the system he was
built to trust reported a launch from the United States. Then a second. Then a
third, a fourth, a fifth. The screen said the highest confidence. Protocol was
a single unbroken line: report up the chain, and the chain would decide in the
minutes it had, which were not many.
Petrov waited.
He reasoned — with no time to reason — that a true American first strike
would come in the hundreds, not in five; that the system was new and the
satellites could be fooled by sunlight on high cloud; that a signal this clean
was, somehow, wrong. He judged the launches to be noise. He picked up the phone
and reported a system malfunction, and then he stood there not knowing, for the
length of the flight time of missiles that were not coming, whether he had just
let his country die.
He had judged correctly. Sunlight on high-altitude cloud, refracted into
the satellites’ sensors. There were no missiles.
And it was not the first time a man had stood in that exact spot.
Twenty-three years earlier, on 5 October 1960 — the day the Ballistic Missile
Early Warning System came online — the radar at Thule, Greenland, reported a
massive Soviet launch and drove the alert to its highest level. The commander
in chief was airborne and unreachable; the call passed to his deputy, a
Canadian air marshal named Charles Roy Slemon. With the authority to begin
escalation toward retaliation, Slemon did not escalate. He asked a single
question: where was Khrushchev? At the United Nations, in New York, came the
answer. Slemon judged that the Soviets would not open a nuclear war with their
own premier standing in the target country, noted that no intelligence
corroborated an attack, and broke protocol to hold. The cause of the alert, it
turned out, was the moon — rising over Norway, its echoes read by a new machine
as a thousand incoming warheads.
Two men, two false signals, two refusals — one Western and one Soviet,
one at the birth of the missile age and one at its height. Neither trusted the
confident screen. Both reached, in the compressed seconds they had, for a piece
of context the machine did not possess: would the enemy really do this, now,
like this? And here is the part the ON TRACK authors would find inconvenient.
At Thule, after Slemon’s night, that same agency — ARPA’s 474L office — went to
work teaching the BMEWS computers to reject echoes from the moon. The
institution that built the reach also built the patch. What it has never built,
and by its nature cannot, is the officer who thinks to ask where Khrushchev is.
Now go back and read the collection’s own words over those two bunkers.
The tempo all but eliminates the possibility of indecision, inaction, and
issue-avoidance — and the entire act, both nights, was indecision, was
inaction, was the refusal to pass the question up. The machine infers intent
without human intervention — and machines inferred intent on both nights,
at the highest confidence, and were wrong. The threshold to act falls when
no person is at risk — and the only thing between the inference and the
launch was a person, at risk, who declined. Everything the 1968 forecast set
out to remove, and everything the 2025 collection files under progress, is the
thing that saved us. The hesitation was not a grave complication. It was the
last working part.
Signal and Noise
Here is the key that turns the whole lock. Claude Shannon taught us to
separate the signal from the noise — and the hardest problem in his theory was
never transmission but reception: the quality of the receiver, its ability to
know which is which under load. Every essay in ON TRACK is a receiver
problem in disguise. Mockingbird’s six detectors, contradicting each other over
one page, are receivers failing in the open. The satellite-intent algorithm is
a receiver we are teaching to judge without us. And Petrov and Slemon, at their
screens, were receivers of the highest quality the system ever had — precisely
because they could hold a signal and doubt it at the same time. The machinery
of the system of systems is machinery to make the receiver faster. The two
officers are the reminder that the thing we actually needed was for the
receiver to be right, and that speed and rightness are not the same
virtue, and under enough compression they become enemies.
None of the ON TRACK authors is a villain. The brigadier wants the human
in the loop. The collection frets, honestly, about accident and accountability
on nearly every page. Westmoreland believed he was describing progress. The men
who put Second Samuel at the head of a manual for burning a country believed
they were being literate and serious. This is the Manhattan register, not the
register of malice: capable people, each step defensible, the aggregate quietly
foreclosing something none of them would choose to lose. That is what makes it
worth writing about. The danger was never that someone decided to remove the
human. The danger is that no one did — that the window closed one reasonable
brick at a time, over sixty years, in the plainest prose, blessed with the
right epigraph.
I keep returning to two instruments pointed at the sky. One is a radio
dish, turned outward, waiting through the night for a signal it prays is real —
the whole discipline of listening built on the hope that the noise will someday
resolve into someone. The other is Petrov’s screen, which showed him a signal
at the highest confidence and asked him to believe it, and the whole of that
night turned on his praying it was not real. Same instrument, in the end. Same
receiver, holding a signal it cannot yet trust, in the dark, while everyone
sleeps. The only question either one ever asks is whether there is a human
awake at the desk, willing to wait, and good enough to know — and whether we
are building a world that still leaves room for one.
The Composite Judgment — reach and reckoning, and the mutual doubt
that binds them.
A note on the
collaboration
This essay and its companion graphics were built by three hands. The
written argument was developed with Claude (Anthropic), across many
turns of drafting, structural critique, and source-hunting — the machine
reaching, surfacing precursors and chasing the ARPA thread faster than one mind
could, while the reckoning and the responsibility stayed with the author. The
two posters — The Expanding Battlespace and its earlier iterations —
were rendered with ChatGPT (OpenAI) and refined across successive rounds
of critique. The Composite Judgment diagram was drawn by Claude. The
direction, the judgment, and the final word throughout were the author’s own.
Which is, more or less, the argument.
Items to verify against primary text before publication.
• The Serpukhov-15 incident (26 September 1983): the five-launch report, the
high-cloud/sunlight false signal, Petrov’s decision to report a malfunction.
Well documented but varies in small details across accounts.
• The BMEWS / Thule incident (5 October 1960): Slemon’s escalation authority, the
Khrushchev question, the moonrise cause. Solidly documented
(nuclear-close-calls literature; Jacobsen, The Pentagon’s Brain); treat the
“let’s sit back and enjoy the show” quotation as folklore unless a sober source
is found.
• Forest Fire as a Military Weapon (June 1970, Project EMOTE,
ARPA Order 818, DTIC AD0509724): the report and its biblical epigraph are confirmed; the
specific attribution to 2 Samuel 18 should be read off the document’s opening
page.
• Westmoreland’s 1969 AUSA address and MacDonald’s 1985 JASON
banquet remarks: verify
wording against The Pentagon’s Brain (electronic-fence chapter, approx. p.
211).
• Licklider, Man-Computer Symbiosis (1960): the “set the goals, formulate the
hypotheses” passage and the “85 per cent” passage against IRE Transactions on
Human Factors in Electronics, HFE-1, March 1960.
• Unless Peace Comes chapter attributions: Wheeler on computers, MacDonald on
the environment, Thring on self-controlling weapons — confirmed against the
1968 Viking edition’s table of contents.
