What Does a Fruit Fly’s Brain Have to Do with Us?
On this Friday evening, September 11, 2026, I want to discuss Google’s announcement about a map of the fruit fly brain’s neural network. Researchers have been studying the fruit fly brain and attempting to uncover its complete network structure for a long time. Now comes the news that the structure has been implemented as a digital map that other researchers can use. This struck me as a very significant development. What matters to me is not simply that we have mapped a small insect’s brain, but the technology that made it possible and the influence this research will have in return.
“What does a fruit fly’s neural network have to do with us?” You might well ask that. In the past, we might have treated it as an interesting achievement from some laboratory and moved on. But AI is now taking on complex work, processing patterns and conducting analysis, changing the speed of research itself. Venerable Dao commented, “There is astonishing new news every day.” Yes. AI is one reason these announcements keep coming. Work that people would previously have struggled to handle has become manageable.
The interesting point is that the influence does not run in just one direction. AI accelerated research into the fruit fly’s neural network, but research into the neural networks of humans and other living things helped bring AI into existence. Attempts to understand biological intelligence contributed to artificial intelligence, which is now used to study biological intelligence. Won’t the structures and mechanisms uncovered through this research help advance AI in turn? That reciprocal relationship is what I want to discuss.
The Shock of Image Recognition Before AlphaGo
Ask people when AI first made a powerful impression on them, and many will recall the 2016 match between Lee Sedol, a ninth-dan Go player, and AlphaGo. Yet several years earlier, there was an image-recognition competition that the computing industry regarded as important. A computer was shown an image and had to identify what it depicted. Was it a cat or a dog? That may seem trivial to a person, but it was difficult for a computer.
The puzzles you solve when entering a website and declaring “I am not a robot” concern the same issue. You identify blurred or distorted numbers, or choose cats from a collection of pictures. We can recognize letters A and B even when other marks are mixed in beside them. Computers used to struggle with this. That is why, two days ago, I discussed the news that AI had solved all 48 CAPTCHA stages. We need to consider what it means when existing methods can no longer easily distinguish people from computers.
This is also why the results Geoffrey Hinton and his students achieved in an image-recognition competition mattered. More than one person is called a godfather of AI, but Hinton is a prominent example. He worked as a professor and at Google, and his students have participated in core research and development at OpenAI and other AI companies. Using graphics cards and artificial neural networks, they achieved results far beyond those of earlier competitions. The ability to tell whether a picture showed a cat or a dog had substantially improved.
The industry began to gain confidence: “Artificial intelligence really is possible.” This was no longer merely an idea or a claim; a competition had produced a result. The computing industry began investing seriously, and translation changed afterward as well. Previously, people wrote algorithms and translations followed those rules, often with terrible results. The turning point I am referring to is Google Translate’s switch to AI translation in 2016. In the same year, Google DeepMind’s AlphaGo played Lee Sedol, giving us another shock.
To Build Intelligence, We Have to Study the Intelligence That Exists
At the center of this process is the ability to distinguish things and ask, “What is this?” We recognize numbers and images, finding patterns even in indistinct marks. To implement that ability in a machine, researchers turn to how people perceive and how living things receive and distinguish sensory information. AI research is therefore deeply connected with the study of human cognition.
The researchers at the forefront of AI development do not come only from mathematics or computer science. Some are connected with biology, cognitive research, and psychology. Psychology and computing may appear far apart. Yet building artificial intelligence means trying to understand existing intelligence and considering how to implement it. It is natural that people who study what we see, what we distinguish, and how we judge should contribute. Hinton is one person with connections to those fields.
AI did not emerge from mathematics alone. Biology, cognitive research, and advances in computing influenced one another. AI began performing some of the distinctions, perceptions, and judgments made by humans and other living things, and now it handles language too. You have probably heard the term LLM. Could questions about how humans learn language, what language is, and how we use our brains have had no influence on this development? Numerous disciplines are involved.
What happens, then, when the AI developed through that process returns to studying biological brains? We see achievements like this map of the fruit fly’s neural network.
Having Male and Female Maps Is Different from Knowing Everything
In translation, the Google Research announcement describes a milestone in connectomics, the study of connectivity. Pause for a moment: AI is also providing the translation we are reading. Google Translate may not offer the same translation ability as Gemini or ChatGPT, but it still makes the material accessible. AI is already involved in the very process of reading an article about AI research.
The announcement specifies a male fruit fly because a female map had appeared earlier. Comparing the male and female network maps displayed above and below one another helps us understand the significance of identifying structural differences. A living body is not laid out on a flat plane. The brain and the neural networks extending through the body exist in three-dimensional space, so the map concerns a three-dimensional arrangement of connections.
However, hearing that the complete structural map is finished should not lead us to think everything has been explained. The neural network of a particular male fruit fly has been mapped. Consider human DNA. Revealing the structure of human DNA does not mean that every individual human’s entire DNA has been separately mapped, does it? People are not all identical. We should understand this achievement as mapping the structure of a particular individual as well.
This distinction does not diminish the achievement. Completing even one individual’s structural map is extraordinary. But a map that copies a structure is different from a complete understanding of how a living brain learns and changes. If we overlook that difference, we will misunderstand both the game experiments discussed below and the direction of future research.
166,000 Neurons and 125 Million Connections
The first thing that caught my attention was the numbers: more than 166,000 neurons and 125 million synaptic connections. The announcement describes this as the largest scale to date in terms of neuron count, providing essential material for scientists who use fruit flies as a model organism to investigate brain function. But the number of neurons and the number of connections differ enormously. Why?
Think of a synapse as a connection between neurons. If each neuron paired with only one other neuron and that was all, pairing 166,000 neurons would require about 83,000 connections. Yet the reported number is 125 million. A neuron does not connect only to one other neuron; it has numerous connections with many different neurons.
Of course, not every neuron will have the same number of connections. Some may have far more, while others have relatively few. Network research also examines such differences. Let us use cities as an analogy. Korea has many cities; imagine Seoul with ten million people and Busan with 3.5 million. I am not discussing precise population statistics here. I am asking you to substitute connections for population: suppose a neuron called Seoul has ten million connections and one called Busan has 3.5 million.
Seen that way, two cities can occupy very different roles within a network. Some places concentrate many connections and play a central role. I briefly forgot the term during the broadcast: a hub. It is easier to understand the arrangement if we treat neurons as nodes and synapses as connections between nodes. We are not looking merely at small dots but at an enormous number of relationships between them.
Six Degrees of Connection and Pictures of Artificial Neural Networks
You may have heard that you can reach anyone in the world through six degrees of connection. Suppose I want to send a message to President Lee Jae-myung. The idea is that I could follow someone I know, then someone that person knows, and reach him after several steps. That may sound implausible at first, but it becomes more likely when a network includes hubs with enormous numbers of connections. People do not all have the same number of relationships.
I bring this up because neural networks can also be considered in terms of connections. Neurons are nodes, and synapses are connections. There are around 166,000 neurons but more than a hundred million connections. In illustrations of AI, you have probably seen rows of dots with dense lines running between them. One node connects to several others. These are familiar pictures in explanations of deep learning.
The name artificial neural network makes sense when we see that it follows aspects of the relationship between neurons and synapses. It does not reproduce a biological network in its entirety; it imitates only a small part. A simple diagram may show nodes in one layer connected to those in the next, without a direct connection from an earlier node to a distant layer. Other approaches then introduce connections that skip layers and reach nodes farther away.
This also leads me to consider the possibility of three-dimensional connections. An artificial network drawn on a screen looks planar, while biological neurons and synapses connect in actual three-dimensional space. Considering those spatial relationships greatly multiplies the possibilities. They are difficult for a person to design and understand mentally. Might artificial neural network research move toward dealing more actively with such three-dimensional connections? That is my own thought. Others may be thinking along similar lines, but I see biological network research as potentially contributing to that direction.
Learning Means That Connections Change
Why have I spent so much time on neurons and synapses while discussing fruit flies? Because learning is connected with changes in synapses. Neurons may also appear or disappear, but connections between neurons become stronger or weaker. In AI, these relationships are expressed numerically as weights. Researchers construct a virtual network and alter those values during learning.
When we remember something, strengthen a memory, forget, learn something new, or experience emotional changes, changes also occur in the brain. These relate to the strengthening or weakening of synaptic connections and the formation of new connections. In an artificial network, we can think of this in terms of changes in connection weights. The terminology may sound complicated, but the point is that remembering and learning do not happen without any change.
You may have heard of neural plasticity. Our brains are not fixed forever in one state. Their structures and roles change according to what we use and experience. This does not mean the entire brain is replaced with something completely different every time. It means that functions and connections can change, allowing a part to take on different work.
From Vision and Hearing to the Problem of an Unchanging Soul
Suppose an accident damages a part of the brain that processes visual information. The eyes themselves remain intact and send signals, but the person cannot see because the processing area is not functioning properly. Must that person necessarily remain unable to see? Could another part take on visual processing and allow function to recover even if the damaged part remains damaged? Questions like these help us understand what neural plasticity means.
Now consider the reverse situation: visual signals stop arriving because a person has lost their sight. The brain area that processed visual information is still there, but the input is absent. Will it simply sit idle? When we discuss cases in which people who lose their sight develop hearing or other senses, we can consider changes in which areas formerly involved in vision take on other sensory processing. Their responsibilities change. This is why we cannot regard every brain function as permanently assigned from beginning to end.
Even without the special term neural plasticity, we should recognize that without changes in the brain, we could not remember or experience anything new. Memory is connected with change. This is exactly where we must question the claim that an unchanging soul is the self and remembers past lives. Let us suppose a soul exists. If it really is something that never changes, how does it remember? Doesn’t remembering require change?
I criticize the notion that an unchanging soul stores memories and reincarnates as bullshit. It is not simply that I dislike the expression. People attach words without asking what memory is or how it relates to change. They must first consider the problem that something unchanging cannot remember. I have moved from brain plasticity to the immortality of the soul and past lives, but this is not an irrelevant digression. It concerns what we call memory and what we call the self.
What If a Fruit Fly Brain Model Played Doom and Mario?
Let us look at an attempt to take a published neural structure and use it for something else. On screen, a fly-like character holds a gun and plays Doom. The old game has been modified slightly to make a fly appear as the protagonist. The material I introduced concerned software engineer Alex Wors experimenting with Doom and Super Mario 64 using a public fruit fly brain model. I briefly hesitated over how to pronounce his name, but the substance of the experiment is what matters here.
This does not mean that an actual living organism has been connected to a game machine. It means taking a published brain connectivity structure as a digital model and testing whether it can do something else. The explanation describes feeding game images, frame by frame, into a model of approximately 166,700 neurons and converting neural activity into game-control signals.
Consider what a frame is. A moving image consists of successive pictures. Older formats sometimes used 16 or 24 frames per second, and around 24 frames gives a sense of continuity. Thirty frames is also used, while 60 frames is frequently discussed in gaming today, with monitors operating at 60Hz or higher. The model receives these individual images. Visual stimuli from the game activate sensory neurons, and the resulting neural activity is linked to character movement or attacks.
The account also describes stimulating particular dopamine neurons when damage occurs in the game to imitate a reward system. To encourage learning, wouldn’t we need to provide a negative stimulus when damage is received and a positive one when an opponent is defeated? That is how the experiment tests whether the model can learn behavior and survive in the game environment. Mario sometimes jumps in irrelevant places on screen. A central display appears to show which parts of the neural network are being stimulated. This does not mean we already know exactly how to use that brain. It means various attempts are beginning.
Copying a Structure Does Not Mean Copying the Principles of Learning
With a virtual fruit fly brain available, we can stimulate it and test what it might learn. But doing this properly requires a better understanding of brain plasticity. How does the network change as a young fruit fly experiences and learns things? Which synapses connect, and how do particular neurons act? What we have now copies the network structure of a particular fly; it does not copy that fly’s entire learning process.
To train the copied structure, the strength of connections must change. We need changes such as weights becoming stronger or weaker. However, changing values in a virtual three-dimensional network is different from checking, one by one, which synapses form and which connections change when an actual fly receives a stimulus. That entire process has not yet been established, has it?
This is why I call the present stage extremely early. Researchers now have to investigate how adult male and female network structures differ and how the various parts function. Nevertheless, elsewhere people are already testing whether these public structures can learn completely different tasks. We do not have to know everything before another line of research can begin. The material already available is enough to generate further questions and experiments.
I am not a specialist who knows everything about connectome research. Nor am I claiming complete knowledge of how far the research has progressed. But there are several things we should understand when encountering this news: the scope of the claim that a structure has been revealed, the problems that remain, and the changing tools used to investigate them.
The Difference in Scale Between Fruit Flies and Humans
The announcement gives 86 billion as the number of neurons in the human brain. Numbers I have heard over the years have varied somewhat, and we should also consider whether every individual has the same number. In any case, this is a different scale from the fruit fly’s roughly 166,000. If there are 86 billion neurons, how many synapses might connect them? It brings to mind immense numbers, in the trillions or beyond. This is why mapping the entire human brain remains difficult.
Even for the fruit fly, changes in every synapse at every moment have not been tracked. The structures of particular male and female specimens at particular moments have been analyzed and mapped. That alone is remarkable. Fruit flies are widely used as model organisms because they are relatively simple, short-lived, and display behaviors such as courtship. There may also be advantages for observing successive generations. But simpler than a human does not mean easy.
Here we can ask again: why is Google announcing a result about the structure of a biological neural network? The announcement describes a decade of collaborative research advancing the use of AI to construct cellular-level maps of entire brains. The human brain is still too complicated, but perhaps a fruit fly with roughly 166,000 neurons is manageable. Even that is an enormously difficult task requiring AI’s assistance.
Connecting Images of Thin Slices into a Three-Dimensional Network
What does making a brain map involve? Following the explanation, we encounter a process of dividing a brain into numerous thin slices, photographing each, then using computers and AI to bring the images together. Think how thin the sections must be when discussing millions of slices from a fruit fly’s tiny brain. This is not a matter of roughly drawing lines and joining them by eye.
You might recall hospital examinations that look inside the body and construct a three-dimensional representation. During the broadcast I thought of entering and leaving a scanning machine, struggled to recall the name, and mentioned MRI. The point here is the construction of a spatial structure from sections. The fruit fly work uses electron-microscope images because examining neural structures requires very precise information at micro- or nanoscale levels.
AI is used in a system that turns flat electron-microscope images into three-dimensional reconstructions and produces accurate nerve-cell shapes. The technology continues to advance. Is the present precision the final limit? Probably not. We can imagine going beyond establishing that neurons and synapses are connected to determining how strong or weak those connections are. The existence of a connection and its strength are different questions.
I regard this as a very early stage. Making one map does not close the subject; it lets us ask what we can examine more precisely with that map. And the more information and finer distinctions we handle, the harder it becomes for people to manage the work alone.
Fish Brains, Sensation and Behavior, and Questions About Disease
Mapping an entire fish brain is also discussed as a future target. Having studied a fruit fly, researchers want to move beyond insects to vertebrates. A zebrafish larva is a young fish. It is introduced as one of the few vertebrates small enough for whole-brain mapping with current technology. Its transparency at the larval stage also makes it possible to measure neural activity during experiments. The questions extend from looking at structure to observing actual activity.
This connects with attempts to understand how the human brain works. Why does Alzheimer’s disease arise, and through what mechanisms does it progress? What processes underlie conditions such as depression or schizophrenia? How do bodily and neural processes affect the mind? What interests me here is the effort to study those mechanisms. There is hope that this work may eventually contribute to treatments for cognitive disorders, better brain health, and the recovery of damaged functions.
The introduction also mentions three papers concerning the visual system, taste, and social behavior. These investigate what is seen and how it is processed, how something is identified as edible and experienced as taste, and what neural activity relates to courtship or interactions with other individuals. I have not opened and examined those papers in detail. I am looking at their titles and descriptions and recognizing the directions in which research is proceeding.
Once a structure of roughly 166,000 neurons can be handled as data, researchers can further investigate and experiment with which areas are stimulated, how actively they respond, and how synapses change and influence one another. The achievement is not merely a beautiful picture of a brain.
How AI Learns Signals That People Struggle to Recognize
Let us return to the starting point. People attempting to create intelligence had to study living things. Whether a fruit fly, earthworm, insect, dog, monkey, or human, they examined what it felt and distinguished and how it judged and learned. Those studies helped AI development, and their results joined with investment and industry to produce today’s vast AI industry. The tools developed through that process are now being used to analyze biological neural networks in return.
The reason is that some patterns are too complex and subtle for people to distinguish individually. Consider, for example, identifying some early indication of dementia in a brain image. A difference may be too subtle for a person to judge readily, while an AI trained specifically to distinguish it can learn the pattern. The finer the difference, the harder it becomes to rely only on human vision and experience.
A robotic arm is another useful example. Suppose someone who has lost an arm is fitted with a robotic one. Signals from the brain travel through nerves by electrical and chemical processes. Could a person simply inspect those complicated signals and immediately say, “This means bend one finger,” or “This means pull the arm back”? They are too subtle and complex.
The situation changes if AI learns the patterns linking that person’s neural signals to movements. It interprets the intention to bend an arm, lift a finger, or make a fist and moves the robotic arm accordingly. That interpretation is needed if the person is to use it like their own hand. This is what I mean when I emphasize that AI’s pattern-recognition ability is used in many places already.
From Autonomous Driving to Dolphins and Cats
Consider what an autonomous vehicle does. Is that a person, a dog, or an obstacle? Is the light red or green? Is that a crossing or a lane marking? Is there fog? Should it stop or proceed, turn left or right? These are all matters of perception, distinction, and judgment. The system learns patterns thousands and tens of thousands of times, extending to hundreds of millions and billions of instances, to drive.
We can ask similar questions about animal signals that humans struggle to understand. In what situations does a dolphin make a particular movement or sound? What is a bird call conveying? What feeling or condition does a cat’s voice and behavior reveal? AI can be trained to distinguish subtle differences that people may fail to recognize despite careful attention. I think a day is approaching when AI will help us understand the emotions and actions of birds and cats and communicate with them.
This pattern-recognition ability returns to understanding the brain’s complex structure and operation. Learning language, feeling an emotion, being hurt, becoming depressed, and undergoing emotional changes involve brain activity. AI is used to uncover which neurons connect through which synapses, and it can also be used to understand activity within that network. Recognizing structure and recognizing patterns of operation advance together.
The Cycle in Which Better Biological Research Produces Better AI
As AI advances, it will become easier to study more complex neural structures and subtler mechanisms. Where will that research lead? Back to better design of AI structures and learning. The three-dimensional networks I mentioned are one possibility. As I see it, present artificial networks are generally understood and designed in planar terms. There may be some spatial approaches, but dealing with the actual three-dimensional connections of biological systems greatly increases complexity. AI helps us understand precisely that complexity.
How does a spider identify prey and weave a web? How do dogs, dolphins, and cats make judgments? Studying neural organization in fruit flies, fish, or humans can directly or indirectly influence AI and robotics. Biological neural network research and artificial neural network research help each other and accelerate each other.
Frankly, how would we individually analyze biological networks of this complexity without AI? It is almost impossible for people to identify and classify every pattern themselves. Training AI for a particular purpose makes previously unmanageable work possible. After a structural map is produced, that analytical ability can also help explain vision, taste, and social behavior in greater detail.
The fact that I have been discussing AGI for several days does not mean that current AI is complete. Compared with what may lie ahead, it is still at a very early stage. There remains the possibility of achieving greater intelligence with the same energy and better capabilities through more efficient neural structures. Biological network research will help us understand that efficiency. Learning what is efficient and what is inefficient then broadens our understanding of living things again. It will influence many areas of medicine and biological research, including questions such as why humans age and die.
Acceleration Across Science and Industry, Not Only Brain Research
This relationship is not limited to AI and biological neural networks. AI can also be used to analyze materials and protein structures and select new substances. Choosing which drug might help against a virus such as COVID-19, or which protein structure might have an effect, also involves distinction and judgment. As these abilities improve, the development of drugs and vaccines should become faster and more efficient.
The same applies to mathematical and scientific problems, industrial technology, materials, and new substances. AI assists that research, and the resulting mathematics, physical understanding, industrial techniques, and materials improve the conditions for AI development, including semiconductors and energy efficiency. The tool helps research; research changes the tool; and the changed tool speeds up the next round of research.
This is why I see us as entering the early stage of acceleration. An achievement in one field does not simply end there. It travels through other fields and hastens further achievements. That is the significance I see in the completion of the fruit fly brain map. It seems that the time when we can actually feel rapid acceleration is approaching.
Why I Cannot Simply Accept an Optimistic Economic Outlook for 2030
I had originally intended to discuss, alongside fruit flies, an outlook for labor and the economy through 2030 published by the organization behind Claude. It separates cases in which AI develops less than expected, at a moderate pace, more successfully, or suddenly reaches much greater intelligence than anticipated, examining the consequences for the economy, labor, and GDP. I plan to leave a detailed review until next Monday.
What is interesting is that it concerns only the period through 2030, not a distant future. It is September 2026, so roughly four years remain. Even looking that far ahead requires several cases and complicated forecasts. That shows how wide the range of change is. I think we are in a situation where even developers cannot precisely predict the capabilities of the next model they are building.
It is in that context that I refer to the earlier broadcast’s discussion of OpenAI’s Astra, the subsequent internal model, and the account of work on the Navier–Stokes equations. Developing and using the next model does not mean knowing all its eventual capabilities in advance. That makes it difficult to express the level AI will reach in 2030 as one definite figure.
However, I do not simply accept that economic report. Parts of it seem far too naive to me, maintaining a position so lacking in candor that it approaches self-justification. It appeared to reserve scenarios involving labor-market collapse, rapidly rising GDP, and the collapse of employment for people in their twenties and thirties for the case of rapid AI progress. Would those problems really arise only in the most rapid case? I think we need to read its assumptions and explanations critically together. But this broadcast has not examined the report in detail, so I leave that as a question for the next discussion.
People Who Speak of Past-Life Memories Without Understanding Memory
Discussing fruit fly neural networks also makes us reconsider how people talk about souls, selves, and a true self. We should not remain distant from the observable world and modern science while presenting our wishes as facts. We need to try to see what is there independently of what we want.
People make claims such as “There are past lives and souls, and memories exist inside them,” “I remembered a previous life from childhood,” or “Something enters the womb.” But what exactly is memory? Without understanding memory, they immediately attach words about it being stored in a soul or about karma. I call that bullshit. Many repeat those claims without even understanding why they are bullshit.
They do not know how our brains develop, how we perceive, distinguish, and judge, or what has so far been observed. I have criticized Buddhist superstitions and misconceptions, and the myths, lies, novels, fairy tales, and wishes of people in the past contained in scriptures. But even before that, we need to learn what modern science means. Something being written in a scripture does not make it an explanation of how memory and life work.
If people had properly learned about evolution, could they speak so casually of creation or reincarnation? They claim to have studied earth science, physics, and biology, yet repeat these claims because they have not considered the implications. The same applies to what AI demonstrates. They memorize words to pass examinations without thinking deeply about what those words mean. I criticize the attitude of people who appear with the titles of scientist or professor and casually make claims about an afterlife. A title does not do our critical thinking for us.
Can Buddhist Practice Be Unrelated to Changes in the World?
During the broadcast I also reflected on how quiet the chat has been lately. Some people may have arrived with an interest in Buddhism and found the discussion of AI unexpected. But we cannot practice Buddhism in isolation from worldly affairs. We need to examine what the changes affecting not only me but Korea and the entire world will produce, how we should live, and what and how we should teach.
Of course, discussing the immediate problems in one’s own life is important too. I do not consider those trivial. You can ask about personal concerns, Buddhism, or questions prompted by another book at any time through comments or live chat. But someone also needs to take an interest in this period of upheaval, interpret it, consider what may follow, and try to understand it. Persistent thought and investigation are necessary.
We cannot demand that everyone do this. If someone is busy making a living, when can they examine all these matters? Those with the capacity should do it. That does not mean I do it because I am wealthy. I am not an affluent “Gangnam leftist”; I am poor. I have put aside the idea of doing other things and decided to do this work, so I ought to do it. Shouldn’t politicians, people called intellectuals, professors, and influential figures also do their part?
I sometimes reflect on whether I am repeating at length what I could explain briefly. I have talked for a long time about one fruit fly, but what we are trying to understand is not confined to that small insect. It concerns what we can claim to know, how we understand intelligence, and how we participate in a changing world. I am glad Venerable Dao found it interesting. Please leave any questions you have, enjoy your weekend, and let us talk again next time.
This article was edited by AI solely from the supplied video captions and description. No external sources were used.