Big minds are slow minds

Why can't a bigger mind just think faster?

Picture a team making decisions by shouting across a field. Sound takes time to travel, so communication limits how many rounds of discussion the team can complete per minute. Spreading more people over a bigger field increases that delay.

A distributed mind faces a similar constraint. A step that depends on several parts must wait for messages between them. Those messages can travel no faster than light.

If a mind needs every part to communicate with every other within one clock tick, a signal must cross its full width in that time. A faster shared clock therefore requires a smaller mind. At a billion ticks per second, the maximum width is the distance light covers in a billionth of a second, about the length of a ruler.

The people of the Vigil live on both sides of this rule. A chip is a mind-part small enough to keep one clock. By light's measure a human brain is tiny too: about a hand's length across, so light would cross it in 0.50 nanoseconds. Then there is the swarm. Its computing zone, zone C, is a ring around Vesper about as wide as Mercury's orbit round the Sun, and light needs about 7 minutes to cross it. A mind spread across all of that zone could take a step that needs every part of itself no more often than once per crossing: about 9 such steps an hour. Whether one mind can live that spread out is still argued frontier. The light delay is not.

A large mind can still run many small processors quickly and in parallel. It combines their answers less frequently, because each exchange across the whole system takes time.

Guess first A mind the width of zone C runs at a 1 GHz clock. How many ticks of that clock pass while one signal crosses it?

About 4.0 × 10¹¹. Light needs 7 minutes to cross, and a billion ticks happen every second, so that many ticks go by while one message is in flight.

Another option is to match the clock to the communication delay. A mind as wide as zone C, ticking once every 7 minutes, allows one crossing per tick. Many of the Vigil's residents choose a slower pace to save energy and coordinate with friends across the swarm.

How big, how fast, how much?

The relation is one line:

time for one synchronized step = size ÷ c

Here c is the speed of light, 2.997925 × 10⁸ metres per second. A mind of size L can take at most c ÷ L synchronized steps each second.

Run the ruler first. A clock at 1 GHz ticks every billionth of a second, and in that time light travels 30 centimetres. That is the widest a mind can be and still hear from every part within one tick. If the parts must also answer (ask, then reply), the limit halves: a round trip takes twice the time.

Take a brain as about 0.15 metres across. Light crosses it in 0.50 nanoseconds, so a brain-sized mind could keep step at up to 2.0 GHz. Real brains don't come close. Signals in the fastest nerve fibres travel at about 100 metres per second, 3 × 10⁶ times slower than light, and a signal needs about 1.5 milliseconds to cross. A brain operates far below its light-speed limit, using local processing without requiring every part to agree at once.

Now the swarm. Zone C runs from 0.1 to 0.4 astronomical units from Vesper. Light from Vesper needs 50 seconds to reach its inner edge and 200 seconds to reach its outer edge. Two nodes on opposite sides of the outer edge are 0.8 AU apart: a signal between them needs at least 7 minutes, and a question and its answer take 13.

A mind the size of	Light crosses it in	Fastest shared clock	Steps an hour
A chip at 1 GHz (30 cm)	one billionth of a second	1 GHz	3.6 × 10¹²
A brain (0.15 m)	0.50 ns	2.0 GHz	7.2 × 10¹²
Zone C, edge to edge	7 minutes	2.5 millihertz	9

The chip and the swarm differ by a factor of 4.0 × 10¹¹ in how often they can take a shared step. A thought that needs a thousand shared steps takes 1 microsecond on the chip and about 111 hours across zone C, close to 5 days.

What does a big mind give up, and what does it keep?

It gives up depth: the number of steps in a row, each depending on all the last. Anything that needs the whole mind to agree, over and over, gets slower in proportion to its size.

It keeps width. Heat sets a budget for how much a mind can forget each second (the The cost of forgetting floor), and that budget does not care how the mind is laid out. Say the Vigil's whole baseline of 1 × 10¹⁵ watts leaves as heat into 30 K radiators, and every erasure costs no more than the floor at that temperature. Then the swarm can forget at most 3.5 × 10³⁶ bits every second, however it is organized. Run it as a fabric of 1 GHz cells and that comes to 3.5 × 10²⁷ bits per tick of the cells' clock. Run it as one mind stepping every 7 minutes and each step gets 1.4 × 10³⁹. The same work per second, delivered in a few huge steps instead of many small ones. A resident's whole state is about 1 × 10¹⁷ bits illustrative, so one step of the swarm-mind has room to forget that much 1 × 10²² times over.

How fast a thinker of a given size can keep one clock, on log axes. Light takes time to cross the thinker, so the fastest tick everyone can share falls in step with its size.

With JavaScript on, this is a calculator: slide the size of a mind from a millimetre to the width of zone C and it shows how long light takes to cross, the fastest clock that stays in step, and the most a petawatt of heat into 30 K radiators lets it forget in one step. The table above gives the same numbers for three sizes.

Open everything

Why does light set the limit, and what does a mind lose by being big?Light cones, and a simple model of time against size

A step that needs everyone

Two events, one at place $x_1$ at time $t_1$ and one at $x_2$ at time $t_2$ , can influence each other only if light has time to get from one to the other:

$|x_2 - x_1| \le c\,|t_2 - t_1|$

The set of events that can reach a given event, or that it can reach, is its light cone. Events outside each other's cones, called spacelike separated, cannot affect each other at all, and observers moving differently will disagree about which came first. So there is no universal "now" across a big mind. A shared clock requires a synchronization convention and messages between the parts. Leslie Lamport made the same point about distributed computers in 1978, borrowing it from relativity: the order of events that means anything is the one messages can carry, and clocks are kept in step only by exchanging messages.

Now take a mind of largest dimension $L$ whose next step depends on the state of all its parts. The step cannot begin until the farthest part's news arrives, so

$t_{\text{step}} \ge \frac{L}{c}, \qquad f_{\text{sync}} \le \frac{c}{L}$

If the parts must also acknowledge, the round trip doubles it: $t_{\text{step}} \ge 2L/c$ . Seth Lloyd writes the one-way time for a computer of radius $R$ as $t_{\text{com}} = 2R/c$ , which is the first bound with $L = 2R$ .

A model of one thought

Break a thought into total work $W$ (elementary operations) and $s$ rounds where every part must hear from every other. Let $N$ processors each run at rate $f$ . The time the thought takes is the work, spread out, plus the waiting:

$\tau = \frac{W}{N f} + s\,\frac{L}{c}$

The first term is the part more processors can shrink. The second is the serial part in Gene Amdahl's sense, and adding processors does not touch it. It grows with size.

What fixed heat does to it

Suppose each operation erases one bit, at the least possible cost $k_B T \ln 2$ (the The cost of forgetting floor, with $T$ the temperature the heat goes into), and that the mind may dissipate power $P$ . Then

$N f \,k_B T\ln 2 \le P \quad\Longrightarrow\quad N f \le \frac{P}{k_B T\ln 2} \equiv B$

$B$ is the number of operations per second the heat allows. It depends on the power and the temperature, and not on size or clock. Substituting:

$\tau \ge \frac{W\,k_B T\ln 2}{P} + s\,\frac{L}{c}$

At fixed heat, the first term is the same for every layout. Size does not make the work go faster. It only adds the second term. Lloyd found the same thing for the limit set by energy: spreading the energy over more gates makes each slower and leaves the total rate unchanged, so parallelization can make some computations more efficient but does not change the total number of operations per second.

The first term assumes every operation erases a bit. The Vigil's mostly reversible computation reduces this heat cost. The communication delay in the second term still applies whenever a step needs information from across the system.

Why be big at all?

Memory, and heat. Information needs room: if each cubic metre can hold at most $\rho$ bits, a mind that must hold $M$ bits has $L \gtrsim (M/\rho)^{1/3}$ . And heat needs surface. The Vigil's baseline petawatt, leaving at 30 K, needs about 43 Earths' worth of radiating area, and the machinery that makes that heat has to sit near the radiators that shed it. Since $\tau$ grows linearly in $L$ , the best size for a task is the smallest that holds what it must remember and sheds what it must shed. For the swarm, those memory and cooling requirements lead to a large physical extent.

Matching the clock to the size

Measure delay in the mind's own ticks: $\lambda = fL/c$ , the number of ticks that pass while news crosses the mind. Anders Sandberg used the same ratio in 1999 and called it $S$ . For $\lambda \le 1$ the whole mind hears itself within a tick: the regime of the chip, and of a swarm-mind running slow on purpose. For $\lambda \gg 1$ a global step costs many ticks, and the mind is a federation of fast local workers that meet occasionally. A fabric of cells of size $\ell$ ticking at $c/\ell$ inside a mind of size $L$ does $L/\ell$ local ticks per global tick. For $\ell$ = 30 cm and $L$ = zone C, that is 4.0 × 10¹¹. Lloyd's measure of parallelism, $t_{\text{com}}/t_{\text{flip}}$ , is the same kind of ratio.

What have people worked out about it, and what is still argued?Jupiter brains, parallel against serial, and agreeing at a distance

Planet-sized minds, in 1999

Anders Sandberg's 1999 paper on "Jupiter brains" works through the physics of minds the size of planets: how dense, how fast, how big and how hot they can be, and how they talk to themselves. Its section on delays measures communication lag by the ratio above, $S$ = length ÷ (signal speed × clock time): how many clock cycles pass while news arrives from elsewhere in the system. For a human brain he takes a neural cycle of 1 millisecond, signals at about 100 metres per second and a length of 0.10 metres, which gives $S$ of about 1: the brain waits on itself for about one cycle. A chip a few centimetres across with a 1 GHz clock has $S$ well below one and rarely waits on itself. His way out for bigger minds is the brain's own trick, layers: fast low levels do most of the work locally, slower high levels coordinate them, and the mind as a whole can spread far beyond its light-limit while its top levels run slow. This is the layered architecture described above. The paper explores physically permitted designs using assumed materials and engineering capabilities frontier.

Parallel against serial

Lloyd's 2000 paper quantifies the tradeoff between communication and processing speed. He measures how parallel a computer is by the time a message takes to cross it, $2R/c$ , divided by the time a bit takes to flip. For his idealized one-kilogram, one-litre "ultimate laptop" that ratio is about 1 × 10¹⁰, so even at the physical limit such a computer is highly parallel. Compressing a computer makes it more serial, and only at the final limit, a black hole, is it entirely serial: there a bit flips in about the time a signal takes to travel round the horizon.

Gene Amdahl made the older, practical version of the point in 1967: the part of a job that must run one step after another sets a ceiling on what any number of processors can do. Light delay turns that serial part into a physical quantity, and it grows with size.

Agreeing across a distance

Separated parts of a mind must sometimes make one decision: which of two plans to adopt, which copy of a memory is the real one. Computer science has proved how hard that is.

In 1985 Michael Fischer, Nancy Lynch and Michael Paterson proved that in an asynchronous system, where messages can take arbitrarily long, no deterministic protocol can guarantee agreement if even one process may crash. A slow message and a dead sender look the same.
Light gives a swarm known delays instead of arbitrary ones: in a known layout, a message on a working link arrives in a time anyone can compute in advance. That moves it toward the synchronous model, where agreement is possible. Even there, Fischer and Lynch showed in 1982 that tolerating $f$ faulty parts that may say anything needs at least $f+1$ rounds of messages. Every round across zone C costs 7 minutes.
Lamport's 1978 logical clocks show that a consistent order of events needs no global time, only messages. The cost is that agreeing on an order takes at least one crossing.

So a mind the width of zone C can decide, but each decision costs several crossings, tens of minutes rather than nanoseconds. It does better to decide rarely, locally, and in advance.

Still argued

Does a mind need a single clock at all? Brains don't have one. They work with local, asynchronous activity and slow, loose coordination, which is how a mind can be larger than its light-limit. Whether a swarm-mind should count as one mind, or as a society of fast small ones in constant correspondence, depends on how minds are defined frontier.
How fast do large minds really run? It depends on how much of a thought must be global. As of 2026 no one has measured that for a mind, because no one has built one.
Can an architecture reduce the delay? Hierarchies, prediction and caching hide latency by guessing what other parts will say. They can't tell a part something it has not said. No known scheme beats the light-cone bound on genuinely new information, and no quantum effect carries a message: entanglement can't be used to signal (the no-signalling theorem).