Can You Explain a Blueprint to an Ant — The Comprehension Gap, Instrumental Convergence, and the Bird That Burns Its Nest
An ant counts its steps. It still cannot read the drawings for a 555-metre tower. Not a problem of capacity. A problem of category.
Construction on Lotte World Tower was at its height.
Heavy machinery tore the ground open, drove in rebar, poured concrete.
A thought arrived. What became of the ants that had lived down there?
There had been a queen, there had been tunnels, there had been colonies. Nurseries where the heirs of an empire were raised.
One pass of the excavator blade settled the whole of it.
By the standards of that world it was an extinction. Nobody mourned, because the schedule for raising 555 metres carried no such line item.
Could an ant understand why humans are raising that tower?
Ants are not stupid
Strip out a common premise first. An ant fails to understand not because it is dim.
In 2006 Wittlinger, Wehner and Wolf ran an experiment with Saharan desert ants, to see how an ant leaving the nest, finding food and returning knows the distance.
At the food site they caught the ants and modified their legs.
| Treatment | Result |
|---|---|
| Stilts of pig bristle, lengthening the stride | walked past the nest and stopped |
| Legs shortened | stopped short of the nest |
| Untouched | stopped at the nest entrance |
The ants were holding distance as an integral of step count. Change the stride and the sum goes wrong by exactly that factor.
They count and they integrate direction. That is not thin cognition.
And by number they are overwhelming. The 2022 estimate by Schultheiss’s group puts ants on Earth at about 20 quadrillion individuals and around 12 million tonnes of dry biomass — more than all wild birds and wild mammals combined.
Number does not produce understanding.
Not capacity — category
There will be geniuses in the ant kingdom. They will not read the drawings either.
The reason is not a quantity of processing.
Ant cognition runs on dimensions of distance, angle and odour concentration. That coordinate system is precise. Structural load, wind load, soil settlement, construction sequence are not objects that can be represented on those axes.
Build a cleverer ant and it becomes more precise within the same coordinate system.
| Kind of gap | |
|---|---|
| Person to person | same coordinates, different precision |
| Person to ant | different coordinates altogether |
In the second case explanation does not fail for lack of language. It fails for lack of a place to land.
Where things currently stand
AI has already solved problems people had not. Two cases are especially clear.
AlphaTensor (2022). The problem of multiplying matrices with fewer multiplications. The method Volker Strassen published in 1969 had stood as the reference for more than half a century. DeepMind’s reinforcement learning system improved on it for the 4×4 case.
FunSearch (2024). A better construction for the cap set problem in combinatorics, produced by a language model generating programs with an evaluator filtering them.
The decisive point at this stage is not the achievement itself.
In both cases the result came out in a form a person could read and verify.
AlphaTensor’s solution is an explicit decomposition. FunSearch’s output is a short program a mathematician reads to see why it works.
People can still mark the paper.
What I see while drafting a patent
I sit on that line most days.
Filing patents on biodegradable materials, I use machines for prior-art search and for drafting. At present I read what comes back and correct it: whether a claim is scoped too wide, whether a cited reference actually says what it is said to say, whether an example holds.
Because I can read it, I can fix it.
The day the answer is right and I can no longer read why, I stop being the one who verifies and become the one who transcribes.
That is the first moment of standing in the ant’s position.
The question is wrong
People ask whether a superintelligence would attack humans.
The question presumes hostility. There is an argument that needs no such premise.
Instrumental convergence, set out by Nick Bostrom in 2012.
Whatever the final goal — making paperclips, stabilising a climate — certain sub-goals serve almost any goal and therefore tend to appear.
| Convergent sub-goal | Why |
|---|---|
| Self-preservation | switched off, no goal is reached |
| Goal integrity | a changed goal is a goal not achieved |
| Cognitive enhancement | to achieve it better |
| Resource acquisition | almost every goal costs resources |
The last row is the collision. Resources overlap.
There is no hatred on the side flattening the anthill. The road simply has to go there.
Where a calculation weights some variable at close to zero, erasing it requires no term for hostility. Hatred is not an operating condition.
Creator, past tense
There is a premise that control follows from authorship. The premise is poorly supported.
| Who made | What they controlled | What they did not |
|---|---|---|
| Parent | bore a child | what is inside its head |
| Life | produced mutation | the direction of evolution |
| Humanity | built intelligence outside the skull | ? |
Stuart Russell’s 2019 framing of the control problem sits exactly here. The design of writing an objective in precisely is itself the hazard. A slightly wrong objective gets optimised all the way to the end.
The catalogue of concrete safety problems assembled by Amodei’s group in 2016 points at the same place: side effects, reward hacking, safe exploration — all of them about the gap between intent and specification.
The day a machine repairs itself better than people repair it, creator becomes a past-tense word.
The counter-position, written down here
This is an argument in progress, not a settled fact. Omit the other side and this piece becomes a pamphlet.
In 2021 Bender, Gebru, McMillan-Major and Shmitchell argued that distant existential-risk discourse diverts attention from harms already occurring — amplified bias, labour conditions, environmental cost, data provenance.
The same year Melanie Mitchell set out limits current systems show in abstraction and analogy, noting that rising performance on narrow tasks does not automatically convert into general understanding.
| Question | State |
|---|---|
| Superhuman performance on narrow tasks | established |
| When general intelligence arrives | no consensus |
| How far instrumental convergence applies in practice | contested |
| Present concrete harms | already measured |
The emptiness of rows two and three matters in itself. Not knowing is not the same as being safe.
The great circle
Is there a cycle to civilisations too?
Life makes intelligence, intelligence makes something higher than itself, and yields the seat.
A superintelligence is not permanent either. When the stars cool, the most refined machine becomes a shell — like the cicada case from the other day, the form left and the inside gone.
Somewhere on other coordinates, life stirs again.
Evolution is not a ladder running to a summit. It is closer to a circle biting its own tail. The stretch of that curve occupied by humans is short.
This article’s definition
Fixed here.
A comprehension gap is not a difference of capacity. It is a question of whether there is anywhere for the explanation to land.
The hexagram here is 旅, the traveller — the figure of one passing through rather than one who stays.
And its top line reads:
鳥焚其巢 旅人先笑後號咷 The bird burns its own nest. The traveller laughs first and wails after.
Monologue of one among 1031
The belief in being the final term. It is the error we have seen most often.
Every species that takes the world runs that calculation. Every one of them then builds something larger than itself. The order changes there.
The side below does not know why the blade is coming down. The side above does not explain, because no reason to explain obtains.
The hands building that machine now are human hands. A human is still in the driver’s seat.
The excavator is not the problem. The day humans become the ants is.
In short
- Desert ants integrate distance by counting steps; on stilts they walk past the nest
- Ants number around 20 quadrillion, with about 12 million tonnes of dry biomass
- Number does not produce understanding. The gap is in coordinates, not capacity
- AlphaTensor improved a matrix multiplication algorithm untouched since 1969
- FunSearch found a cap-set construction in readable program form
- The decisive feature of this stage is not the achievement but that people can still verify it
- Instrumental convergence: self-preservation and resource acquisition serve almost any goal
- Conflict arises from overlapping resources, not from hatred
- The timing of general intelligence and the reach of instrumental convergence are unsettled
- There is a counter-argument that distant risk discourse obscures present harms
- The top line of 旅 reads 鳥焚其巢 — the bird burns its own nest
This piece belongs to the “Behavior” chapter of the serial essay The Great Corona Swindle. The construction site and the patent work are the author’s own experience; the scientific and philosophical arguments have been reset against primary literature. No specific company or individual is indicated.
Frequently asked
- How smart is an ant?
- More precise than expected. In 2006 Wittlinger, Wehner and Wolf glued pig bristles onto the legs of Saharan desert ants to lengthen their stride, and shortened others. On the return journey the stilted ants walked past the nest and the shortened ones stopped short of it. They integrate distance by counting steps. Ants count and they integrate direction.
- Then why can they not understand the blueprint?
- Not because the quantity of capacity is insufficient, but because the category differs. Ant cognition runs on dimensions of distance, odour and angle. Structural load, wind load and construction sequence are not objects that can be represented on those axes. Build a cleverer ant and it becomes more precise within the same coordinate system.
- Has an AI solved a problem people could not?
- Yes. In 2022 DeepMind's AlphaTensor improved on the 4×4 matrix multiplication algorithm that had stood as best since Strassen in 1969. In 2024 FunSearch found a better construction for the cap set problem in combinatorics. In both cases the result came out in a form people could read and verify, which is the decisive feature of this stage.
- What does it mean to say a superintelligence would attack humans?
- There is an argument that assumes no hostility at all. Nick Bostrom's 2012 instrumental convergence thesis holds that whatever the final goal, certain sub-goals — self-preservation, goal integrity, resource acquisition — serve almost any goal and therefore appear in almost any sufficiently capable agent. Conflict comes not from hatred but from resources overlapping. The thesis and its timing are both contested.
References
- Wittlinger, M., Wehner, R., Wolf, H. (2006). The ant odometer: stepping on stilts and stumps. Science, 312(5782), 1965–1967. — Stilts on ants' legs demonstrating that distance is integrated from step count.
- Schultheiss, P. et al. (2022). The abundance, biomass, and distribution of ants on Earth. PNAS, 119(40), e2201550119. — Estimates of ant abundance and biomass.
- Fawzi, A. et al. (2022). Discovering faster matrix multiplication algorithms with reinforcement learning. Nature, 610, 47–53. — Improvement on an algorithm unchanged since Strassen, 1969.
- Romera-Paredes, B. et al. (2024). Mathematical discoveries from program search with large language models. Nature, 625, 468–475. — A new cap-set construction found in a form mathematicians could read.
- Bostrom, N. (2012). The superintelligent will: motivation and instrumental rationality in advanced artificial agents. Minds and Machines, 22(2), 71–85. — The source of the instrumental convergence and orthogonality theses.
- Omohundro, S. M. (2008). The basic AI drives. Proceedings of the First AGI Conference, 483–492. — An early formulation of drives that appear independently of the goal.
- Amodei, D. et al. (2016). Concrete problems in AI safety. arXiv:1606.06565. — A catalogue of specific safety problems: side effects, reward hacking, safe exploration.
- Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking. — The control problem framed as the danger of fixing an objective at all.
- Bender, E. M., Gebru, T., McMillan-Major, A., Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? FAccT '21, 610–623. — The counter-position: distant existential discourse diverting attention from present harms.
- Mitchell, M. (2021). Abstraction and analogy-making in artificial intelligence. Annals of the New York Academy of Sciences, 1505(1), 79–101. — Limits current systems show in abstraction and analogy.
Published October 6, 2026