There was a time when computers could beat humans at chess, perform millions of calculations per second, and manage enormous databases.
But show one a photograph of a golden retriever sitting on a couch?
Apparently, it was a beige blob occupying a rectangular region of slightly darker beige.
The computer was not impressed.
The dog was probably offended.
And artificial intelligence had a rather embarrassing problem.
It could process information, but it was remarkably bad at understanding what it was looking at.
Then along came a scientist who decided that perhaps the problem was not simply that computers needed better algorithms.
Perhaps they needed something else.
Millions and millions of photographs.
Preferably with labels.
Welcome to Episode 4 of Who’s Who in AI — The People and Companies That Built the Future.
Hi, Dave LumAI here. I spend an unreasonable amount of time talking to machines that can explain quantum physics but occasionally struggle to identify the contents of my refrigerator. Today we’re meeting the woman who helped change that particular situation, although I regret to report that AI still cannot tell me why I opened the refrigerator in the first place.
In Episode 1, Alan Turing asked whether machines could think.
In Episode 2, John McCarthy, Marvin Minsky, and Dartmouth gave artificial intelligence its name.
And in Episode 3, Geoffrey Hinton, Yann LeCun, and Yoshua Bengio, we met the neural-network pioneers who kept working on learning algorithms when much of the scientific establishment was more interested in practically anything else.
Now comes the woman who helped provide the enormous collection of examples those algorithms desperately needed.
And believe me, this story has some excellent surprises.
First, Some Interesting Tidbits
Before we get into the serious science, consider these little gems.
Fei-Fei Li helped build a database containing approximately 14 million images organized into more than 21,000 categories.
That’s a lot of photographs.
My camera roll has about 14,000, and I’m already considering hiring someone to identify which 300 are pictures of my thumb.
Her ImageNet project enlisted approximately 50,000 human workers from 167 countries to help label and verify images.
Yes, fifty thousand people.
This was artificial intelligence’s enormous human homework assignment.
And here’s another surprise.
Before becoming one of the world’s most influential AI scientists, Li helped run her family’s dry-cleaning business while studying physics at Princeton.
I imagine that created some interesting conversations.
“What are you studying?”
“Physics.”
“What are you doing this weekend?”
“Dry cleaning.”
“What are your long-term plans?”
“Revolutionizing machine intelligence.”
“Excellent. Your shirts will be ready Tuesday.”
And the truly astonishing part?
In September 2026, the company she cofounded agreed to be acquired by AMD in an approximately $8.2 billion deal.
From dry cleaning to multibillion-dollar AI research.
Now that’s a career trajectory.
Let’s find out how it happened.
Who Is Fei-Fei Li?
Fei-Fei Li was born in Beijing, China, in 1976 and grew up in Chengdu.
She immigrated to the United States with her family as a teenager, settling in New Jersey.
Her early experience was not exactly the standard Silicon Valley origin story.
She had to learn English, adapt to a new country, help support her family, and navigate an educational system that was initially unfamiliar.
Her high school math teacher, Bob Sabella, even arranged advanced mathematics instruction for her during lunch because the school didn’t offer the class she needed.
That is the kind of teacher you remember.
And she certainly did.
Li earned a scholarship to Princeton University, where she studied physics and graduated in 1999.
Meanwhile, her family operated a dry-cleaning business, which she helped manage for years.
She subsequently attended the California Institute of Technology, earning a doctorate in electrical engineering in 2005.
Along the way, she developed a fascination with the relationship between human perception, neuroscience, and computing.
How do biological brains recognize objects?
How can humans identify a familiar face almost instantly?
And why can a toddler recognize a dog from an unfamiliar angle when a sophisticated computer struggles to distinguish a poodle from an upholstered footstool?
These questions drew her into computer vision.
Her remarkable personal journey is explored in Stanford’s account of her early life and memoir, The Worlds I See.
And it helps explain something important about her scientific approach.
Li was interested in intelligence as more than a collection of mathematical tricks.
She wanted to understand perception itself.
The Problem: Computers Didn’t Understand Pictures
In the early 2000s, computer vision researchers were already building systems capable of detecting edges, identifying patterns, and recognizing objects under controlled conditions.
But real-world images are messy.
Consider a cat.
A photograph might show a black cat, white cat, enormous fluffy cat, tiny kitten, sleeping cat, or a cat hanging upside down from a curtain while reconsidering its relationship with gravity.
To a human, they’re all cats.
To early image-recognition software, the differences could be surprisingly difficult.
Researchers often relied on carefully engineered features.
A feature is some useful characteristic in an image, such as an edge, corner, texture, or particular arrangement of shapes.
Engineers designed algorithms to extract these features and identify objects.
The problem?
The world contains too many variations.
Different lighting.
Different backgrounds.
Different viewing angles.
Different sizes.
Different colors.
And cats, which have made a lifelong commitment to being uncooperative.
Li began asking a different question.
What if computers needed to learn from the same kind of enormous visual diversity humans encounter naturally?
Instead of trying to describe every possible object by hand, perhaps researchers could build a sufficiently large collection of examples.
Let algorithms discover useful patterns.
And measure how well they learned.
It sounds obvious now.
At the time, it was a remarkably ambitious research direction.
ImageNet: The World’s Most Ambitious Photo Album
Li and her collaborators began developing what became ImageNet, the large-scale image database that transformed computer-vision research.
The project was formally introduced in a landmark 2009 research paper by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei.
Notice that list.
ImageNet was not the work of one person sitting in an office with a spectacularly large collection of JPEGs.
It required a research team, collaborators, technical infrastructure, and an extraordinary amount of human labor.
The idea was to assemble a huge collection of images organized by meaningful concepts.
Dogs.
Cats.
Cars.
Trees.
Kitchen utensils.
Musical instruments.
Birds.
Lots of birds.
More than 800 kinds of birds, in fact, according to Li’s retrospective account.
Apparently, ornithology was getting excellent representation.
The project used WordNet, an existing database of English words and their conceptual relationships, as its organizational foundation.
WordNet could describe relationships such as:
A golden retriever is a kind of dog.
A dog is a kind of mammal.
A mammal is a kind of animal.
ImageNet associated these concepts with photographs.
The result was a structured visual resource that researchers could use to train and evaluate algorithms.
At its widely recognized historical scale, the collection contained roughly 14 million images spread across 21,841 categories.
That’s not a photo album.
That’s a photo album that should probably have its own municipal government.
Fifty Thousand People and an Enormous Labeling Problem
Collecting millions of pictures was only half the challenge.
The images had to be labeled accurately.
A computer cannot learn much from a folder containing 14 million randomly named files unless it has some way to connect the pixels to meaningful concepts.
The team initially tried relying on Princeton students to help with labeling.
That approach quickly ran into a minor logistical difficulty.
There are only so many hours in the day.
And students occasionally insist on sleeping, eating, and attending classes.
The project turned to Amazon Mechanical Turk, a crowdsourcing platform that allowed people worldwide to complete small tasks.
Between 2007 and 2009, approximately 50,000 workers from 167 countries helped label and verify images.
The scale was extraordinary.
And here is an important distinction.
ImageNet was not simply a triumph of artificial intelligence. It was also a triumph of human coordination.
Thousands of people performed the essential work that allowed machines to learn.
This is one of AI’s recurring ironies.
The more impressive the machine becomes, the easier it is to forget the enormous human effort behind it.
Li and collaborator Ranjay Krishna describe the entire undertaking in their excellent retrospective, Searching for Computer Vision North Stars.
It is a fascinating account of how asking the right scientific question can transform an entire research field.
Then ImageNet Became a Competition
Having built a massive dataset, Li and her colleagues did something clever.
They turned part of it into a standardized challenge.
The ImageNet Large Scale Visual Recognition Challenge, usually shortened to ILSVRC, began in 2010.
Researchers would train their algorithms using a common collection of labeled images.
Then the algorithms would be evaluated on images they had not previously seen.
Everybody faced the same task.
Everybody could measure their performance.
And everybody could discover just how much their brilliant invention still struggled with photographs of ordinary household objects.
The famous classification challenge used a subset of ImageNet containing approximately 1.2 million training images representing 1,000 object categories.
That’s an important distinction.
The full ImageNet database contained far more images and categories than the competition subset.
The standardized benchmark allowed researchers to compare progress.
Instead of vaguely announcing that an algorithm was “really good at identifying things,” researchers could produce actual numbers.
And those numbers were about to become very interesting.
2012: The Moment Everything Changed
Remember Geoffrey Hinton from Episode 3?
He’s back.
Along with Alex Krizhevsky and Ilya Sutskever.
In 2012, the University of Toronto team entered the ImageNet competition using a deep convolutional neural network.
The system became known as AlexNet.
It used neural-network techniques that had been developing for years, along with substantial computing power supplied by graphics processing units, or GPUs.
Those GPUs matter.
We’ll get to Jensen Huang and NVIDIA later in this series.
For now, understand that powerful graphics hardware was helping researchers train neural networks on much larger datasets.
AlexNet produced an extraordinary result.
Its winning system achieved a 15.3% top-five error rate, compared with approximately 26.2% for the second-place entry.
What’s a top-five error rate?
Imagine showing the computer a photograph and asking it to identify the object.
The computer submits its five most likely answers.
If the correct answer isn’t among those five, that’s an error.
AlexNet dramatically reduced those mistakes.
And the size of the improvement was difficult to ignore.
You can examine the original research in ImageNet Classification with Deep Convolutional Neural Networks.
Here’s the wonderful thing about this moment.
Li didn’t invent AlexNet. Hinton’s team didn’t invent ImageNet. And neither breakthrough would have had quite the same impact without the other.
ImageNet provided the data and the benchmark.
AlexNet provided a powerful demonstration of deep learning.
GPUs made the enormous calculations practical.
Together, they helped convince researchers that neural networks could outperform many established approaches to computer vision.
A whole lot of previously skeptical people suddenly became extraordinarily enthusiastic about deep learning.
Amazing what a convincing set of numbers can accomplish.
What Did Fei-Fei Li Actually Get Right?
Her most important insight wasn’t simply that more data was better.
It was that the right combination of scale, diversity, quality, and measurable evaluation could fundamentally change the pace of scientific progress.
Large datasets made it possible to train more capable learning systems.
Standardized benchmarks made progress visible.
And public research resources allowed laboratories around the world to build on one another’s work.
That last part is especially important.
ImageNet wasn’t a secret company dataset designed exclusively to make one organization’s product better.
It became a shared research resource, made available under research and educational use conditions.
Its impact spread far beyond the original ImageNet competition.
Researchers applied related techniques to object detection, image search, visual recognition, and many other areas.
They also developed transfer learning, where knowledge acquired from one task helps a model learn another.
For example, a model trained to recognize ordinary objects might learn useful visual patterns that help when adapting it to a different image-recognition task.
That doesn’t mean a model trained on photographs of dogs can suddenly diagnose medical conditions.
Specialized applications require appropriate data, evaluation, and expertise.
But the broader idea became enormously influential.
ImageNet helped demonstrate that training on a large visual dataset could produce useful starting points for other problems.
And that lesson continues to shape machine learning.
The Complicated Part: ImageNet Wasn’t Perfect
Now comes the part where history gets more interesting.
Because even enormously influential scientific projects can contain serious mistakes.
ImageNet assembled photographs and labels from the world as it existed.
Unfortunately, the world contains biases, stereotypes, and questionable assumptions.
Some of those problems made their way into the dataset.
Particularly troubling were categories involving people.
Certain labels were offensive or sensitive.
Others attempted to classify characteristics that could not reliably or ethically be determined from appearance.
For example, looking at a photograph cannot tell you whether someone is generous, trustworthy, or morally admirable.
Yet a classification system might be tempted to treat those descriptions as visually identifiable properties.
That’s not computer vision.
That’s automated stereotyping wearing a laboratory coat.
Critics raised serious concerns about the dataset’s treatment of human subjects, consent, representation, and potentially harmful labels.
In 2019, the ImageNet researchers publicly acknowledged these problems.
Their research, Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy, identified 1,593 problematic categories within the original collection of person-related labels and proposed removing approximately 600,000 associated images.
The paper was subsequently published in 2020.
Subsequent revisions went further.
The broader lesson was uncomfortable but essential.
A large dataset is not automatically a good dataset.
And an accurate benchmark score doesn’t guarantee fairness, safety, or real-world reliability.
In their later reflections, Li and her collaborators acknowledged that they had not sufficiently anticipated some of these issues.
That acknowledgment matters.
Scientific progress isn’t simply the story of successful experiments.
It’s also the story of recognizing limitations and correcting mistakes.
A technology that can recognize a photograph with extraordinary accuracy may still misunderstand the social world surrounding that photograph.
And those are very different problems.
Beyond ImageNet: Building Human-Centered AI
Li’s career didn’t stop with computer vision.
She joined Stanford’s faculty in 2009 and later directed the Stanford Artificial Intelligence Laboratory from 2013 to 2018.
She also served as vice president and chief scientist for AI and machine learning at Google Cloud during a Stanford sabbatical in 2017 and 2018.
Over time, she became increasingly prominent in discussions about the relationship between artificial intelligence and human society.
She helped establish AI4ALL, an organization dedicated to expanding participation in AI education.
She also cofounded Stanford’s Institute for Human-Centered Artificial Intelligence, commonly called HAI.
The underlying idea is straightforward.
Artificial intelligence should be developed with human needs, responsibilities, institutions, and values in mind.
Not simply because a company has discovered that adding AI to a product description improves the chances of raising another billion dollars.
Li has argued that artificial intelligence should augment human capabilities rather than merely attempt to replace people.
That perspective connects her technical research to questions about education, healthcare, scientific discovery, accessibility, and public policy.
And as AI grows more powerful, those questions become harder to ignore.
In May 2026, Stanford announced a restructuring that combined its AI and data science initiatives.
Under the new organization, James Landay became the institute’s Denning Director, while Li took on a university-wide role as Special Advisor on AI to Stanford President Jonathan Levin.
She also became co-chair of HAI’s advisory council alongside former Stanford president John Hennessy.
You can read about the changes in Stanford’s May 2026 announcement.
But university leadership is only part of her current story.
Because Li has also returned to building ambitious AI technology.
And this time, two-dimensional pictures are not enough.
World Labs: Apparently Teaching AI to See Was Just the Beginning
Let’s jump to 2024.
Li and three collaborators founded World Labs, an AI research company focused on spatial intelligence.
What is spatial intelligence?
Think of it as moving beyond recognizing objects in photographs toward understanding three-dimensional environments.
A conventional image classifier might recognize a chair.
A more spatially capable system could potentially reason about where the chair is, how it relates to other objects, what lies behind it, and how an agent could move through the room.
For example:
A photograph shows a table, a doorway, and a couch.
A spatially intelligent system should eventually be able to understand their arrangement, infer aspects of the surrounding environment, and help plan interactions within that space.
That’s much more demanding than simply declaring, “Couch detected.”
A couch does not become useful spatial knowledge until you understand where it is.
And ideally, whether you are about to walk into it.
World Labs began developing AI systems known as world models.
These systems aim to represent, generate, reconstruct, and ultimately reason about environments.
In November 2025, World Labs introduced Marble, a multimodal world model capable of generating explorable 3D environments.
Then came another major development.
On September 1, 2026, World Labs introduced Atlas, a new world model designed to work with text, images, video, and three-dimensional information.
And the company has been moving quickly.
On November 5, 2026, it introduced Frameboy, a research preview demonstrating real-time interactive exploration using a version of Atlas.
Then, on November 12, 2026 — literally yesterday as I write this — the company opened broader access to Atlas and introduced additional tools and developer APIs.
Through World Labs’ Atlas platform, users can explore tools such as Viewfinder for creating and reconstructing scenes, Chisel for shaping their structure, and functionality for converting generated 3D representations into meshes.
That last capability matters because meshes can be used in established 3D software, games, and simulation workflows.
Potential applications include visual effects, architecture, virtual environments, and robotics research.
To be clear, generating a visually convincing 3D environment doesn’t automatically mean a model understands physics or can safely control a robot.
Those are much harder problems.
But the direction is fascinating.
ImageNet helped machines recognize objects.
World Labs is pursuing machines that can work with entire environments.
The ambition has grown considerably.
So has the price tag.
The $8.2 Billion Plot Twist
On September 28, 2026, AMD announced an agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion.
Yes.
Eight point two billion dollars.
That’s a lot of dry cleaning.
According to AMD’s official acquisition announcement, the deal is intended to bring World Labs’ model-research expertise closer to AMD’s hardware, software, and computing infrastructure.
The agreement is expected to close by the end of 2026, subject to regulatory approvals and customary closing conditions.
As of November 13, the transaction has been announced but not confirmed as completed.
Li remains World Labs’ cofounder and CEO.
Once the acquisition closes, she is expected to become AMD’s executive vice president and chief scientist, reporting to CEO Lisa Su.
Think about that combination.
One of AI’s most influential computer-vision researchers joining forces with one of the world’s most important computing-hardware companies.
It’s almost as though the AI world has noticed that algorithms and hardware work particularly well together.
Who knew?
Well, Jensen Huang certainly did.
We’ll meet him properly in Episode 6.
For now, AMD’s acquisition agreement illustrates how central advanced model research has become to the hardware industry.
The people developing future AI models are increasingly influencing the design of the computers that will run them.
That is a significant change from the days when researchers simply hoped their experimental code would finish running before the weekend.
How Fei-Fei Li Fits Into the AI Family Tree
By now, the relationships in this series should be getting clearer.
Alan Turing helped establish the intellectual foundations of intelligent machines.
John McCarthy and Marvin Minsky helped organize artificial intelligence into a research discipline.
Geoffrey Hinton, Yann LeCun, and Yoshua Bengio helped develop and sustain neural-network research.
Fei-Fei Li and her collaborators built an enormous dataset and benchmark that helped those learning methods demonstrate their capabilities.
Alex Krizhevsky and Ilya Sutskever helped deliver the famous 2012 breakthrough using ImageNet.
NVIDIA’s graphics hardware helped make large-scale neural-network training practical.
And today’s companies continue extending those ideas into language, vision, robotics, simulation, and increasingly complex AI systems.
One person’s breakthrough created opportunities for another.
Sometimes the connections were direct collaborations.
Sometimes they were enabling technologies.
And sometimes somebody simply built a resource so useful that an entire research field began organizing around it.
That was ImageNet.
What Does Her Story Tell Us About the Future?
There’s a temptation to imagine scientific breakthroughs as sudden flashes of genius.
A researcher wakes up one morning.
Has an extraordinary idea.
Writes down a formula.
And by lunchtime, the world has changed.
Reality is usually less cinematic.
Li’s contribution required years of research, considerable skepticism, a large team, thousands of human workers, and an enormous amount of organizational persistence.
The breakthrough wasn’t simply a clever algorithm.
It was recognizing that the field needed something different.
Better data. A common benchmark. A problem worth solving.
And now, with spatial intelligence, she’s pursuing another ambitious question.
What if artificial intelligence could move beyond identifying objects and learn to reason more effectively about the physical world?
That may prove transformative.
Or today’s approaches may encounter limitations that require new techniques.
History hasn’t written the ending yet.
But there’s something remarkable about a scientist who helped reshape AI once and is now attempting to help reshape it again.
Especially when the second effort involves an $8.2 billion acquisition agreement.
Apparently, the ambitions didn’t get smaller.
Friday Night Laughs Mini — Friday the 13th Edition
Since today is Friday the 13th, it seems only appropriate to bring Jason Voorhees into our computer-vision discussion.
Jason decided to test a new AI-powered security camera at Camp Crystal Lake.
The camera recognized his hockey mask immediately.
“Excellent,” Jason thought. “Artificial intelligence has finally learned to see.”
Then the system announced:
“Welcome back, Mr. Voorhees. Your application for summer camp counselor has been approved.”
Even Jason was horrified.
Apparently, the algorithm had been trained on the hiring records from every Friday the 13th sequel.
Somebody really needs to clean that dataset.
Final Thoughts: Sometimes the Most Important Breakthrough Is Giving Everyone Something Better to Learn From
Fei-Fei Li helped change artificial intelligence by recognizing that one of its biggest obstacles was not a shortage of clever researchers.
It was a shortage of sufficiently large, diverse, and useful training data.
ImageNet provided an extraordinary resource.
The ImageNet Challenge provided a common measurement.
Deep-learning researchers provided powerful algorithms.
And together, they helped launch one of the most important technological transformations of the twenty-first century.
Not every part of that transformation has been perfect.
ImageNet had flaws.
Machine learning still has limitations.
And visual recognition is not the same thing as understanding the world.
But the overall impact is undeniable.
Li helped make computer vision into something vastly more capable than it had been.
And with World Labs, she’s now pursuing another level of visual and spatial intelligence.
That’s what makes her such an important figure in AI.
She didn’t merely help answer a question.
She helped the entire field ask a better one.
And sometimes that’s how revolutions begin.
Next up in Episode 5: Demis Hassabis and Google DeepMind — From Games to General Intelligence.
We’ll meet the chess prodigy and game developer who helped build a research laboratory that taught AI to master Go, predict protein structures, and pursue increasingly general forms of intelligence.
Apparently, beating humans at board games was just the warm-up.
If you’re enjoying this series, follow me for the next episode.
And here’s a question I’d genuinely love to hear your thoughts on.
What’s the bigger breakthrough: teaching AI to recognize everything it sees, or teaching it to understand how those things actually work together?
Drop your thoughts in the comments.
Bonus points if you can explain why my supposedly intelligent phone keeps showing me photographs of the same cat when I’m searching for a completely different cat.
Some problems may never be solved.
Art Prompt (Suprematism):
Create an extraordinary early twentieth-century geometric abstract composition that appears to occupy the mysterious boundary between painting, architecture, and weightless mechanical construction. Against a vast, softly textured ivory-white ground, arrange a complex yet precisely balanced constellation of monumental black rectangular beams, translucent smoky-gray planes, elongated ocher and dusty-blue bars, delicate graphite diagonals, and a commanding golden-yellow circular disk partially concealed behind intersecting structures. The composition should feel simultaneously flat and impossibly three-dimensional, as though an intricate architectural machine has been suspended in empty space without gravity. Introduce subtly overlapping wedges, narrow black vertical elements, offset rectangular fragments, and tiny flashes of vermilion that intensify the visual rhythm. Employ a restrained palette of warm ocher, pale mustard, slate blue, chalk white, charcoal black, parchment beige, and cool silver-gray. Recreate the exquisite material complexity of historical avant-garde mixed media through layered gesso, matte oil pigment, finely textured paper, sandpaper-like grain, delicate metallic foil reflections, and faintly visible geometric construction lines. Establish a strong diagonal movement through the composition while allowing quiet negative space to counterbalance the intricate central arrangement. The resulting artwork should evoke an intellectual, utopian vision of architecture freed from gravity, where geometric forms become floating structures and familiar physical perspectives dissolve into elegant mathematical uncertainty. Museum-quality finish, exquisite surface detail, sophisticated asymmetry, remarkable spatial tension, and a timeless atmosphere of radical artistic invention. No figures, recognizable buildings, text, or decorative borders.
Video Prompt:
Begin with an explosive diagonal movement as a massive black geometric beam snaps into position across a luminous ivory field, immediately followed by a brilliant golden-yellow disk spinning rapidly into view from behind a translucent gray plane. Thin graphite lines shoot across the composition like precision-engineered trajectories while overlapping rectangular structures pivot, rotate, extend, and slide past one another in perfectly synchronized mechanical choreography. Create astonishing three-dimensional parallax as flat painted surfaces suddenly acquire impossible architectural depth, with weathered paper textures separating into multiple floating layers. Dusty-blue and warm ocher bars sweep through the foreground, momentarily obscuring the golden disk before revealing it again through sharply angled openings. Introduce rapid perspective inversions, elegantly timed geometric collisions, subtle metallic reflections, and fragments that rotate independently while maintaining a balanced abstract arrangement. Alternate explosive directional movements with brief moments of perfectly suspended stillness, creating a hypnotic visual rhythm. Preserve the tactile appearance of hand-painted pigments, layered gesso, worn paper, graphite marks, and delicate foil surfaces throughout the animation. Use bold cinematic lighting, precise motion blur, dramatic spatial compression, and seamless depth transitions to suggest an imaginary architectural world assembling itself beyond ordinary gravity. Conclude with the entire structure rotating into a breathtaking symmetrical equilibrium before one slender black diagonal suddenly slips free, creating a perfect visual loop. Vertical 9:16 composition, extraordinary clarity, dynamic editing, and museum-quality visual refinement.
Songs to Pair With the Video
Inspector Norse — Todd Terje
An infectious, playful electronic groove with the perfect rhythmic energy for all those rotating planes, shifting structures, and wonderfully impossible geometric movements.
LesAlpx — Floating Points
Propulsive electronic percussion and intricate synthesizer textures that complement the sharp architectural movements and hypnotic mechanical precision.
Follow for more AI history, surprising scientific characters, creative art, and the occasional encounter with a computer that still can’t recognize a perfectly ordinary household object.