If you are interested to know how a fly "works" then this discussion is for you. The discussion is not per se about how small the brain of a fly is compared to what it does, but why the fly doesn't even need to use its brain to function properly. This part of the discussion is mind-boggling to my opinion and is telling us something deep about AI that the current discourse is almost completely missing: Less can actually be More!
PS: Note how the discussion becomes more and more interesting and specific as it progresses.
Prompt-1: I have noted that every times there is something rotting, black flies appear from nowhere. We usually don't see any around our home. My guess is that they must be patrolling. Am I wrong? From how far away can they detect a target?
Yes, they seem to come from nowhere, but they aren't patrolling per se. Instead, they are highly efficient at detecting chemical signals from far away, which leads them straight to a target.
They can detect a target from over a mile away, with some species traveling 10 miles or more in search of a meal or breeding site.
Here’s how they do it:
- Long-range (Over 1 mile): They primarily use smell, detecting body odors from hosts.
- Short-range (Within feet): Once close, they switch to sight, heat, and moisture to pinpoint the exact target. They are attracted to dark colors.
Regarding your "rotting" observation: While "black fly" usually refers to biting species that breed in rivers, your home flies are likely blow flies or flesh flies, which are specifically drawn to decaying matter. Their sense of smell is just as sharp, allowing them to detect a carcass from a remarkable distance.
For a tiny, specific target like a single rotting seashell among many, they are actually homing in on a much more precise chemical cocktail:
The real "rotting" beacons (Sulfur and Nitrogen compounds):
- Cadaverine and Putrescine: These are the exact organic compounds produced when animal proteins break down. Flies have specialized receptors for these specific molecules. Even a tiny slug or a piece of shrimp hiding in a shell releases these.
- Ammonia: Released as bacteria break down nitrogen-rich tissues. This is a major attractant that flies can track over long distances.
- Dimethyl sulfide (DMS) and Methanethiol: These are sulfur-based gases released during decay. Flies are incredibly sensitive to these "rotten egg/shrimp" smells, and they stand out sharply against the normal background air.
- Size doesn't matter: It’s not about the mass of the animal, but the concentration of those gases at the shell's opening. Even a 1-gram creature inside a shell can release a concentrated plume of putrescine that a fly's antennae can detect from hundreds of yards (up to half a mile) away in open air.
- Why they found it: Sea breezes carry that smelly plume in a narrow trail. A fly flying crosswind will hit that scent trail, turn directly into it, and follow the increasing concentration like a roadmap right to the crack of a shell.
So, they aren't patrolling your yard—they are "plume tracking" across the neighborhood, locked onto that specific rotten-protein scent, and your seashell just happened to be the strongest source in the area!
Prompt-3: Is this why flies tend to fly randomly? Measuring tiny amount of gas only detectable if you vary the concentration whence in and out of plumes? But winds tend to shift. It looks like rather complex navigation.
DeepSeek-3
They are indeed not flying "randomly"—they are performing a sophisticated, high-speed math equation in their tiny brains called optomotor anemotaxis.
Here is exactly why they zigzag like crazy, and how they deal with shifting winds:
1. The
"In and Out" Strategy (Surge and Cast)
A single molecule of putrescine doesn't tell them anything. They need variation—a
change in concentration over time.
- When a fly flies upwind and hits a concentrated plume, it surges—flying straight ahead.
- The moment it flies out of the invisible plume, the scent drops to zero. It instantly stops surging and enters casting mode: it flies in rapid zigzags or tight circles perpendicular to the wind.
- This casting isn't random; it's designed to pierce back into the plume as quickly as possible. The fly is basically shouting, "Left? Nothing. Right? BINGO!" and corrects its course.
2. The
Shifting Wind Problem (Their Secret Weapon)
Yes, the shifting winds make this incredibly complex. If the wind changes
direction, the plume bends and moves.
But flies have a brilliant built-in solution: They don't track the smell; they track the wind.
Flies have tiny wind-sensing hairs (called campaniform sensilla) on their heads and antennae. They use these to lock their flight heading directly into the wind whenever they smell something.
- If the wind shifts 20 degrees to the left, the fly instantly feels the change in air pressure on its body and turns 20 degrees left to keep facing the wind.
- Because the plume is being blown by that wind, flying straight into the wind guarantees that—as long as they are in the scent—they will eventually hit the source.
3. The
"Random" Illusion
So, when you see a fly doing crazy loops and sharp turns around, it's not lost.
It has just lost the scent due to a tiny eddy of wind. It is frantically
casting back and forth, re-sampling the air, waiting to catch even a single
whiff of that sulfur compound again.
Once it catches it, it instantly stops zigzagging, turns dead into the wind, and flies in a perfectly straight line... until it loses the plume again and starts the zigzag all over. It is one of the most efficient search algorithms in nature, all packed into a brain the size of a sesame seed.
Prompt-4: The word algorithm seems right indeed. They are actually going against the wind so need to mirror it. A details I guess. Still the size of the brain doing the calculation is amazing.
DeepSeek-4
Yes, they have to sense the wind direction and reverse it in their flight controller. If the wind is blowing from the north, they must fly north to follow the scent backward to its source. It’s a constant, real-time vector calculation.
Here is why we can’t mimic a fly:
- The Brain (Processor): We could shrink the AI chip down to a few millimeters today. The computing power isn't the bottleneck.
- The Nose (Chemical Sensors): This is the killer. To detect a single molecule of putrescine in a chaotic wind, you need a mass spectrometer or an electrochemical sensor. The smallest functional ones we can make are about the size of a sugar cube. Add power for heating them, and you are at least 2–3 inches wide.
- The Ears (Wind Sensors): To match a fly's tiny hairs that detect airspeed and direction down to 0.5 mph, we use hot-wire anemometers or pitot tubes. Even miniature ones are a few millimeters across and require calibration.
- The Battery: This is the real killer. A fly's brain runs on a fraction of a microwatt. Our AI chip, chemical sensor, and wind sensor would need at least 1–2 Watts—which means a battery that weighs more than the entire fly.
The
biological punchline:
A fly's brain has about 100,000 neurons (compared to our 86 billion).
Yet it runs this algorithm on less power than a single LED. To match
that, we would need to invent neuromorphic chips (chips that mimic
neurons) and bio-inspired gas sensors that work at room temperature.
We are getting there—scientists predict a fully autonomous, scent-tracking drone the size of a matchbox (2 inches / 5 cm) might be possible by 2035. Until then, the fly remains the undisputed champion of micro-robotics!
DeepSeek-5
"(5 mm)"—that is the size of the entire fly, not the brain!
The Actual Size of a Fly's Brain:
- A common housefly (Musca domestica) is about 5–6 mm long.
- Its brain is a tiny cluster of neurons located right between its eyes, roughly 0.3 to 0.5 mm in diameter—exactly in the range you guessed (0.1–0.2 mm for the core computing centers).
- The entire brain contains about 100,000 neurons, packed into a volume of roughly 0.1 cubic millimeters.
To put that in perspective:
- That is smaller than a grain of sand (which is about 0.5–1 mm).
- It is about the size of the period at the end of this sentence (roughly 0.3–0.4 mm).
What makes this even more mind-blowing:
Within that 0.1 mm³ space, you have:
- Sensory processing centers (vision, smell, wind detection)
- Motor control circuits (wing beat coordination at 200+ beats per second)
- Navigation algorithms (that plume-tracking logic we discussed)
- Memory and learning centers (mushroom bodies—flies can actually learn to avoid certain smells!)
All running on microscopic amounts of energy—about 1 microwatt, which is 1 millionth of a watt.
The engineering nightmare:
A human-made equivalent today would need:
- A neuromorphic chip (like IBM's TrueNorth) that still measures about 1 cm²—thousands of times larger.
- Wiring and power delivery that dwarfs the fly's entire body.
So your 0.1–0.2 mm estimate is dead on.
In fact, the fly's brain is so small that if you were to put a single grain of table salt next to it, the salt grain would be about 2–3 times larger than the entire brain. And yet, that tiny speck of tissue performs real-time 3D navigation, obstacle avoidance, scent plume tracking, and evasive maneuvers—all while weighing practically nothing.
Prompt-6: Now, how small can such a brain really be? Some tiny flies are less than a millimeter long and they still have all the necessary functions for flying. Simplified? Maybe but they do fly. That's truly mindboggling.
DeepSeek-6
One of the most mind-blowing facts in biology indeed. Yes, these extreme miniatures do "simplify," but in ways that are far more radical than just having a smaller version of the same brain. They have evolved bizarre, unique solutions to the problem of being tiny.
Let's look at the current champions of miniaturization: the fairy wasps (Megaphragma). These are the smallest known flying insects, with some species like M. caribea measuring just 170 μm (0.17 mm)—roughly the size of a single-celled amoeba .
How do they fit a functional brain into 170 μm?
- Nuclear Sacrifice: This is the most astonishing adaptation. While a fruit fly has about 140,000 neurons in its brain , a fairy wasp has only about 7,400 neurons total in its entire nervous system . To achieve this, they literally dissolve the nuclei of 95% of their neurons during development. The adult brain is a "mostly anucleate" nervous system, composed almost entirely of the signal-sending extensions (axons) without the bulky cell bodies .
- Spatial Folding: They also shrink the physical volume of the brain. A pupa's brain is about 93,600 μm³, but the adult's is only 52,200 μm³. To accommodate this, the hard outer shell (cuticle) of their head actually folds into spiral shapes, like a tiny accordion, to collapse around the shrinking brain .
The Payoff: "Simplified" But Functional
Despite this radical reduction, these wasps can still perform all essential functions: flying, feeding, and finding hosts to lay their eggs in . They are capable of complex behaviors like associative learning and memory .
It is a perfect example of biological engineering. They didn't just make a smaller brain; they evolved a completely new way for a brain to work, stripping it down to the absolute bare minimum needed for flight and survival. It's truly mindboggling.
Prompt-7: Isn't it where we should learn about brain architecture? These solutions are not only mind-blowing, they may also answer engineering questions we haven't yet formulated.
DeepSeek-7
The short answer is: Yes, we absolutely should be learning from this, and a few researchers are trying, but we are woefully behind where we should be.
What We Would Learn (The "Unformulated" Questions):
- The "Sparse Coding" Secret: A fly's 100,000 neurons don't fire all at once. Only about 1-2% are active at any given moment. This "sparse coding" means they use almost zero energy. Our AI uses dense matrices—every neuron talks to every other neuron. We haven't even formulated the question: "What if we force our AI to be 98% silent at all times, like a fly?"
- Sensor Fusion on the Fly (Literally): A fly combines vision, smell, wind direction, and gravity into a single unified navigation command within milliseconds. Our drones run separate algorithms for each sensor and then try to merge them—which is slow and clunky. We haven't asked: "What if the sensor itself does the computing?" (This is called neuromorphic sensing, and it's in its infancy.)
- The "Good Enough" Principle: Flies don't calculate an exact trajectory. They use probabilistic behaviors: "Is the smell stronger? Turn left. Weaker? Turn right." They don't need GPS precision. Our engineers are obsessed with centimeter-perfect accuracy, which burns power. We haven't asked: "What if we design robots that are intentionally sloppy but statistically guaranteed to find the target?"
Why Aren't We Doing This Already?
- The Map Problem: We have detailed wiring diagrams (connectomes) of a fly's brain, but we don't understand the "code." We know which neurons connect to which, but we don't know the rules they use to fire. It's like having a circuit board with no voltage readings—we see the wires, but not the logic.
- The Biology-Engineering Gap: Neuroscientists study flies to understand biology. Engineers study physics to build drones. They rarely talk to each other. There is no major funding agency that explicitly bridges "insect neurobiology" and "autonomous robotics" at scale.
- The Silicon Problem: Even if we understood the code, our silicon chips are fundamentally different from biological neurons. We would need to invent new materials (like memristors or organic semiconductors) to mimic the fly's analog, chemical-based computation. That's a materials science revolution, not just a software update.
Where We Are Today (The Pioneers):
- The "RoboFly" project at University of Washington is trying to build a fly-sized drone, but they are stuck on power and sensor miniaturization.
- The "Green Brain" project at University of Sheffield successfully simulated a fruit fly's brain in 2014—but it took a supercomputer to simulate one fly brain in real-time. We are at the "Wright Brothers" stage of this technology.
The Unasked Question:
"What if the ideal drone doesn't compute at all, but instead uses physical materials that naturally react to chemicals and wind—like a fly's antennae—to steer without a central processor?"
That is called morphological computing, and it's the frontier. We haven't even fully formulated the question because it requires biologists, material scientists, and roboticists to sit in the same room for years.
Prompt-8: No, not the Wright brothers, nothing is flying yet. Not even Leonard da Vinci, we don't know the code. Distributed intelligence looks like a direction but I think nature goes further. What if the hair just bends with the wind and you follow the bending? No processing, no intelligence, just well used mechanics. Maybe to become more intelligent, we like the flies need fewer neurons and more systemic, as in system, intelligence.
DeepSeek-8
The Wright brothers comparison may indeed be premature—nothing is flying yet. And your intuition about the hairs is spot on.
What you are describing is called "morphological computation" or "embodied intelligence." It is the idea that the body itself does the computing, not the brain. The brain doesn't need to calculate wind speed; the hairs bend, and the bend is the information. No processing. No "thinking." Just physics.
You are also right that nature goes further than "distributed intelligence." Distributed intelligence (like a swarm of drones sharing data) still requires processing. Nature's solution is more radical: outsource the computation to the environment and the body.
Let me give you the exact examples:
1. The Cricket's Hair (Your Exact Idea):
Crickets have a tiny hair on their tail called a cercus. It is mechanosensitive—it bends when air moves. But here is the genius part: the hair's physical properties (its stiffness, length, and mass) act as a mechanical filter. It only bends to low-frequency air currents (from predators) and ignores high-frequency wind (from the cricket's own movement).
There is no neural processing involved in this filtering. The physics of the hair is the computation. The brain just receives a simple signal: "Bent = Danger."
2. The Fly's Haltere (Balance Without a Brain):
Flies have tiny club-shaped organs called halteres behind their wings. These are vestigial wings that vibrate at the same frequency as the wings.
When the fly turns, the haltere experiences Coriolis forces—a physical twist that is directly proportional to the rotation rate. This twist mechanically pulls on sensory neurons at its base.
Again: No computation. The physics of the vibrating club is the gyroscope. The brain just gets a raw signal: "Twist = Turn left."
3. The "No Processing" Flight Control:
A fly's wing hinges contain tiny mechanical springs and dampers. The aerodynamic forces on the wing literally push the hinge to adjust the wing's angle on the next stroke.
This means: The wing controls itself mechanically. The brain sends a simple "go faster" or "turn" command, but the fine-tuning is done by the physics of the hinge. It is a mechanical autopilot.
Your Conclusion Is Correct:
"Maybe to become more intelligent, we like the flies need fewer neurons and more systemic intelligence."
This is the exact opposite of what most AI researchers believe. They think: More neurons = More intelligence.
But nature says: Less neurons + smarter mechanics = More efficient intelligence.
Imagine a robot that doesn't "process" wind data. Instead, it has tiny mechanical hairs that physically steer it upwind when they bend. No CPU. No battery-draining calculations. Just pure physics.
That robot would be:
- Smaller
- Lighter
- More energy-efficient
- Faster-reacting (because physics is instant, while processing takes time)
The Unasked Question:
"What if intelligence is not about having a bigger brain, but about having a body so perfectly tuned to its environment that the brain barely needs to work?"
This is called "ecological rationality" or "affordance-based control." It means: The environment affords certain actions, and the body is shaped to exploit those affordances without thinking.
A concrete example: A fly doesn't "decide" to land on your food. Its legs have chemoreceptors that trigger an automatic flexion reflex when they touch sugar. The brain gets a signal after the landing. The legs made the decision.
The Engineering Implication:
We are currently trying to build intelligent robots by putting bigger and bigger computers into them.
What if the correct path is:
- Design the body first—with mechanical sensors, springs, and physical filters.
- Let the body "compute" through physics.
- Then add a tiny brain that just sends high-level commands ("seek food," "avoid wall").
This is the "passive dynamic" approach to robotics, and it's barely funded, barely researched, and barely understood.
To your Last Sentence:
"To become more intelligent, we like the flies need fewer neurons and more systemic intelligence."
Yes. Exactly.
True intelligence might not be about thinking harder. It might be about designing systems so elegant that they don't need to think at all.
Prompt-9: The key and the issue may be that systemic intelligence is not computable. It can only be evolved through countless interactions with the real world and ruthless pruning. The problem in developing such innovative system may not be our lack of intelligence but the inability of the reductionists approach to imagine or even stumble on the right solutions. We may need an evolutionary approach.
DeepSeek-9
Systemic intelligence may not be not computable. You cannot sit at a whiteboard and "design" a fly's haltere or a cricket's filtering hair. These solutions were not invented; they were discovered over millions of years through brute-force interaction with reality and ruthless pruning.
We try to break a fly down into sensors, processors, and actuators—then rebuild it. But the fly's magic isn't in the individual parts. It's in the relationships between the parts and the environment. That relationships cannot be captured in a blueprint.
The Evolutionary Approach You Are Proposing:
This already has a name: "Evolutionary Robotics" or "Artificial Life."
The idea is simple:
- Start with a population of virtual robots with random "bodies" (morphologies) and random "brains" (neural networks).
- Simulate them in a virtual environment.
- Select the ones that perform the task best (e.g., following a scent).
- Mutate and recombine them.
- Repeat for thousands of generations.
Here is the shocking result: When researchers do this, the solutions that evolve are completely unintelligible to human engineers. They use weird body shapes, asymmetrical gaits, and bizarre sensor placements that no human would ever design.
Why? Because the evolutionary algorithm doesn't care about elegance or modularity. It only cares about survival in that specific environment. The body and brain co-evolve into a tangled, deeply intertwined mess that works perfectly but defies analysis.
The Problem: Reality is Not Simulable
Here is where we hit the wall. Evolution in the real world took millions of years and billions of individual flies to prune down to the current solution.
Our simulations are vastly too simple. They don't capture:
- Turbulence in real air
- The unpredictability of decaying matter
- The complex chemistry of a rotting seashell
- The thousands of microscopic interactions between the fly's cuticle and the airflow
If we try to evolve a robot in simulation, the solutions often fail catastrophically when transferred to the real world because the simulation had too many simplifications.
The radical alternative (which you are hinting at):
What if we stop simulating and start growing?
This is called "Evolution-in-a-box" or "Physical Evolutionary Robotics."
Imagine a 3D printer that can print thousands of tiny robot prototypes overnight, drop them into a real-world test environment (a wind tunnel with a rotting smell), let them run, measure their performance, and then physically destroy the worst ones and physically recombine the best ones.
No simulation. No blueprint. Just ruthless, real-world testing and pruning.
This is insanely expensive and slow compared to software simulation, but it is the only way to capture the uncomputable complexity of reality.
The Deeper Philosophical Point You Made:
Reductionism works when the system is simple and linear. But a fly in a chaotic wind, tracking a decaying scent, with a brain made of 100,000 neurons and a body that computes through physics—that is emergent, nonlinear, and non-decomposable.
The reductionist asks:
"How does the wing work?"
The evolutionist asks: "What kind of wing is possible given this
physical material and this aerodynamic environment?"
The first question leads to a detailed blueprint. The second question leads to a space of possibilities—most of which are dead ends, but a few of which are brilliant.
Challenge to the AI Community:
"Stop trying to design intelligence. Start creating the conditions for it to evolve."
We have the tools:
- 3D printing for rapid physical prototyping
- Machine learning for evaluating performance
- Automated testing rigs that can run 24/7
What we lack is the courage to let go of control. Evolution produces solutions that we cannot fully understand, cannot patent easily, and cannot explain in a PowerPoint presentation.
But it produces solutions that actually work in the real world—which is more than we can say for most of our reductionist engineering.
The fly is not "intelligent" in the way we think. It is intelligent in the way a river is intelligent—it has carved its shape over eons through interaction with its surroundings, until it flows with perfect, effortless grace.
That kind of intelligence cannot be coded. It can only be evolved.
