Latent Space as a New Medium

Winslow Homer’s most famous watercolor rendered as a child’s drawing.

Lately I’ve been asking myself: what might artificial intelligence be good for besides answering questions and writing code? My answer is the latent spaces within AIs themselves will become a new medium for creativity. I will first explain what I mean by latent space, and then at the end of this explanation, I offer possible ways scientists and artists may use the latent spaces inherent in neural nets to serve as a new platform for creativity.

---

A Large Language Model (LLM) is like a small zip file that contains all human knowledge. It takes massive arrays of 100,000 GPU chips working in the cloud, and costing billions of dollars, to compress all of human writing into a small working model that could run on one single GPU chip. Even the biggest frontier models compress down to several hundred gigs, which is small enough it can fit on a card in your palm. In a strange but real way the resulting tiny file contains all the information that is on the internet and in our libraries. This tiny card holds a significant proportion of what humans collectively know. Of all the remarkable aspects of AI, this astounding feat of compression may be the least appreciated. This dense, high order compression of human knowledge — called “latent space” — may also be a new medium itself.

This extreme compression of knowledge within latent spaces was not the original intention of the researchers who invented LLMs. The book smartness they contain came as sort of a surprise to the people training them, and we are still trying to figure out how they actually work. What we can say for sure is that the LLM does not contain copies of everything it knows. For instance it knows all Shakespeare plays, and it could create a new play that sounded exactly like Shakespeare, and can even quote famous lines in his plays, but nowhere in the model are the actual texts of Shakespeare. Instead there is simply the abstract information about all the plays, the plots, the characters, the words, the style, the references. Likewise, the LLM could recognize the face of almost any person, and it could generate any possible human face, but nowhere in its code are copies of human faces. Rather, the model is storing all the information about human faces, without storing any faces.

This is weird. Until recently we might have thought that all the information about a thing would take up more storage space than the thing itself. That may be true for a single thing, but not for the aggregate of all things. That is because most things share a lot of common attributes with other things. The neural nets of an LLM do a magic trick by abstracting the information of everything at once, so that it uses the myriad common relationships between things and ideas to compress and abstract them into this virtual “latent” or hidden space.

All three terms in “Large Language Model” are key. For “Large”, the models contain all the knowledge in, say, Wikipedia, and all the text from decades of the internet, all webpages and online discussions, and all the scanned books and journals in most libraries. So far, the power of the model keeps increasing as it gets scaled up in size. The more information it is trained on, the more connections, the better it gets.

The “Language” part of LLMs turns out to be the secret sauce. LLMs were originally invented to do automatic language translation, that is all. But instead of teaching it the rules of language, which is what earlier AI researchers did, this time no language expertise was required. Instead, a neural net absorbed a very large database of human written language (the internet), with the goal of having the neural net (AI) extract out all the hidden patterns of language below our awareness contained within those billions of documents. The goal of the program was to replicate, imitate and synthesize the patterns of language as it is used everyday by humans.

The results shocked everyone. Sure the LLMs could translate language like a human, but the AI also displayed glimpses of human-like intelligence. They could also be creative with language, like they could write up a sales pitch in the style of a sonnet. Some early researchers were spooked by this emergent behavior, including a Google researcher who felt Google’s LLM had an internal intelligence that should not be turned off. We now understand that the intelligence we see in LLMs comes from the logic within the language they were trained on. (See my Why Are LLMs Smart?)

The form of this new mindfulness — the “Model” part of LLMs — is a latent space. Latent space is an abstraction, a map built not in two dimensions, but in billions of dimensions. Imagine a brain made up of billions of straight long arrows going in all directions. Each arrow is dedicated to one idea or one thing. There is an arrow for dogs and an arrow for cats. Related arrows are located next to each other. So the map shows cats and dogs sharing a nearby arrow for fluffy fur. They also share an arrow for ears, and one for tails. Those two attributes are also shared by other animals (other locations) as well. Most of what a dog is is shared by mammals, so this overlap is one source of the compression.

You can think of every concept that we can put into words as being a direction in this space. The dog arrow is really a direction of dogness. Catness is a direction, and so is fluffiness. Anything can become more catlike, or fluffier. You start with a shoe, or a chimney, or a fern, and you can push it along the cat direction and make it more catlike. Or you can push it in the direction of apple toward more appleness, or of smoothness, or in the direction of reddish, or excitement, or more circular. You can also reverse direction and make it less catlike, or less red, less atomic. There are billions of directions in this space.

Related things are near each other in this space. Cats and dogs share many attributes so they intersect many common arrows, such as tails, whiskers, ears, four legs, animals, short, life, etc. But because they hear, they also intersect the microphone vector; because they can jump, they intersect with basketball. Cats are stealthy and intersect with spies. Because dogs are loyal they intersect the vector of patriotism.

Every thing, every concept has a specific location in the map of this huge space, but instead of having just two coordinates (x,y) each thing has a billion-long coordinate. So an old rusty gasoline lawnmower buried in weeds is a very specific intersection with a very long address. Each of its thousands of attributes (rust, gas, lawn, cut, weeds, push, red, dirt, clippings, roar, etc.) has its own direction intersected. Nearby in latent space is a lawn mower that is more in the rust direction, or less red, but also more catlike, or more doglike, or less spaceship-like, or more like whipped cream. That point may represent a real thing or only a virtual or theoretical thing. This mapping works for not just nouns, but any idea, any sound, any image. The whoosh of a splash of water is a direction in latent space. The aha moment in invention. The fright seeing a snake on a path. The notion of a prime number. All these are contained within a single map. This is one of the most astounding, yet underappreciated aspects of an LLM latent space: Everything — everything! — appears on just one map. We’ve never had a system to integrate everything we know and everything we can imagine. One map for all! This has long been a holy grail.

Just to be clear, no human action is doing the mapping. The system itself, the LLM, is mapping each bit of the world, all things, all attributes, all art, all words, all ideas. And astoundingly it creates this map, this latent space, not piecemeal, but all at once simultaneously. (To do so requires an immense, energy-hungry, massive cluster of chips, all connected together with miles of wires — the famous data center now in short supply.)

While training, the LLM is fed millions of books, billions of web pages, and billions of pages of text from social media. It reads every word on each of them, and once this entire library of material is loaded into its mind, it massively calculates all the interconnecting vectors, all the relative directions pointing to each other. The scale of this vast synchronized parallel calculation is staggering. It then throws away the books, the text, the images, and only keeps this tangled web of directions and vectors. These billions of directions are called its parameters. As we build larger and larger models, mapping more and more material, the parameters increase. The latest models on the frontier of AI contain trillions of parameters, meaning there are trillions of directions, or trillions of attributes that it uses to map every idea or thing it has seen.

Something as complicated as a book winds up as both a point in latent space and a journey through latent space. All the notions encountered in a story (window, mid-day stroll, street, vendor, chat, anger, fight, forgiveness) are directions, and as sentences pile up, the directions shift around, going one way and then intersecting in another. The story is really a journey through latent space, which very much mirrors the journey-like experience we have when we read.

So a book contains a sequence of vectors in latent space. But the sum meaning of a book is also just a single point or direction in itself. For instance if I reference the book The Iliad, I’m referring to the whole book, and its vector is closely related, and therefore “nearby” to the other epic war narratives like Beowulf, The Mahabharata, or even Apocalypse Now, even though many parts of them only tangentially intersect. The more related a thing or idea is, the more directions (vectors) it shares with similar things. This is in part how LLMs know stuff. They search for patterns nearby.

When you ask an LLM a question, it will find the answer in latent space. Your question itself begins as a direction, which points to the answer. The LLM addresses each word in your prompt one by one, with each new word shifting the direction of where it goes. The model travels through latent space with each word of the prompt, searching for its answer, step by step. In this way the answer is grown, rather than found.

We naively imagine that an LLM has a mind that thinks a thought and then expresses it. But the LLM finds the answer as it writes the words. There’s no pre-formed thought “behind” the words that then gets translated into language. The words are the thinking. The path through latent space and the answer are the same thing happening simultaneously. In the most modern versions of an LLM, the model will proceed through a “chain of thought” intermediate stage, which jots down words and ideas as it thinks about a problem. Even here the chain of thought is the thinking, not a report of thinking that happened elsewhere. The model isn’t reasoning privately and then writing it down — the writing-it-down is the reasoning.

As an answer grows along the direction of the prompt, the natural question is how does the LLM know when to stop? How does it know when it is correct? The astounding answer is that “correctness” and “completeness” and “cohesiveness” are vectors in this space, too. Any correct answer shares the same “correctness” direction with all other factually true statements. In other words, correctness, truth, cohesiveness, completeness, comprehension, etc are all essentially patterns that are mapped in this space. So the LLM is seeking not only the facts, but is also always trying to move the words it collects in the direction of “true.” True, complete, coherent are not locations but directions. Answers can always be pushed more in that direction (more precise, more specific, more consensus), or pulled back from it (more fanciful, more poetic, more general, more understandable).

This is the beauty of latent space. You can take a thing or an idea and then move it into a new direction with great ease. We can witness that most easily with image generators. The style of a medium, like watercolors, or the style of a particular artist can be transferred from one picture into another. You can ask an AI to transfer the watercolor style of Winslow Homer onto a black and white sketch you made. That Winslow Homer style is a direction in latent space, and your sketch is also a direction in latent space, and your prompt will move your sketch in the direction of Homer’s watercolors. You could also request the inverse. You could prompt the AI to transfer your style of sketch onto a painting by Winslow Homer, and it would push the painting along the direction of more “you” in this latent space.

This works with ideas and concepts as well. Every notion is a direction. You can apply the idea of gunpowder to the Romans. Our prompt might be: “What would the history of the world look like if Romans had discovered gunpowder?” So the AI takes the general direction of the Roman Empire in history and pushes it further in latent space in the direction of gunpowder. This is a huge intellectual feat, because it requires a deep grasp of Roman history, and a deep grasp of the chemistry of gunpowder. There are very few humans who are expert in both, but LLMs are. And it might take weeks for even the human expert to fill out all the possible new connections that would fill the space between these two ideas. The LLMs do this easily because for them, the latent space is continuous. Latent space includes not just everything real but everything possible based on its training. As the model searches this vast map there is no real distinction between what exists and what could exist, except for the directions of “true” and “historical” or “real”.

In addition there is more than one latent space. As the parameters increase, the space increases. As the material models are trained on become more curated, that also shifts their latent space. As the models incorporate more varieties of inputs — physical data, sound, environmental sensors — their latent spaces also expand and shift. Today there may be a hundred latent spaces; next year a thousand. We are only on Day One of understanding how they work and what they can do. A great potential lies ahead. What follows are my speculations of possible ways to exploit the new medium of latent space.

Prototyping – The musician Brian Eno once complained that the problem with computers was that they did not have enough Africa in them. In latent space, Africa is just a vector. You can add more Africa to anything. Increase the Africa in spreadsheets, bicycles, yoga, the Olympics, passwords, kitchens, SAT exams, automobile dashboards, etc, and see what happens. Repeat with other attributes.

White space discovery — Latent space acts as a continuous map of the possible. Most of those possible things don’t exist — yet. The space of what we know, for instance known materials, known proteins, known chess plays, known ways to paint, fill only fragmented, patchy spots with plenty of white spaces between them. The white spaces between known things are unknown to us, but they are already mapped in latent space. We now have new tools to explore these white spaces in a systematic way. What lies in between astronomy and astrology? What gems await in between bluegrass music and ballet? What about in between the notion of corporations and the theory of Gaia? Exploring latent space is the new frontier. Invention shifts from “think of something new” to “prospect in the gaps.” When a gap or hole appears the question becomes: is that gap empty because it’s impossible, because it’s unfashionable, or because nobody’s looked yet?

Cross domain analogy — Does the shape of this problem resemble the shape of anything else? Perhaps a problem (or opportunity) in geology has the same shape in latent space as some patterns in immunology. So the style of a solution can be transferred from one domain to another. A clever solution in lexicology might apply to genetic sequencing, but since there are few (if any) humans who are expert in both sciences, this overlap will only be revealed by the LLMs. Particularly subtle shared shapes in latent space might touch three, four or more fields of expertise, way out of reach of humans. Seeking out structural resemblances in latent space as an intellectual discovery process could easily become a job for some humans.

Latent space measurement — Latent space might also provide a new way of abstract measurement. You can do a kind of primitive arithmetic in latent space. If you start with the concept of a king, you can travel to the notion of a queen with addition and subtraction: king − man + woman = queen. Starting at the king vector, you decrease the male direction, then increase the woman direction, and then you end up with something we call queen. This kind of calculation begins to give a way to measure or specify the distances between two complex things, or two complicated ideas. Using latent space measurements, we could quantify how similar two court rulings are, or two folk melodies. Just project them into a shared latent space and measure. This new field could evolve calibration standards, error bars, and metrics for evaluating extremely complex entities – a key metric we currently lack.

Mining meta patterns — A model trained on millions of cell images, billions of weather sensors, trillions of hours of traffic videos will notice patterns no human has detected. The latent space will internally invent categories for these patterns, patterns that we have no name for, and therefore are not searching for. We can now begin to dissect latent spaces looking for these unnamed features. We can then work backward to figure out what real-world structure the categories are tracking. A new science would describe the meta pattern of these patterns. A new job is searching for these kinds of patterns that persist, and have potential, in whatever area they occur. The latent space thus becomes a specimen: something you dissect to extract discoveries.

Trajectories — Many years ago the BBC broadcast a science program Connections in which the host followed the zig-zagging path of inventions that were spawned as one obscure idea ran into another unlikely idea. This path of connecting ideas could be thought of as a series of shifting directions in latent space that create a route, or a trajectory through the space. It is not hard to imagine artists choreographing a journey of ideas and images morphing in an endless thread of connections. Their art would be a travel journey through latent space.

Retro latents — Over time, as AI advances, most of the latent spaces invented will become obsolete. Like all media, the dead latents will be resurrected at some point as a cool vintage. The constraints and glitches present in them become cherished later on, in the way that the grain in film, or the sound texture of vinyl, or the bitmap art in old video games becomes a sought-after charm. Someday in the future young kids will revisit ChatGPT-4 to explore its weird hallucinations since their latest AI models rarely hallucinate.

Latent space infiltrators – The shadowy outlaws who explore abandoned buildings and underground urban infrastructures like tunnels or skyscraper roof tops – anywhere it is illegal to be – are called infiltrators. Buried deep inside latent spaces are the programed guardrails, which prevent the models from giving out socially unacceptable information, like how to build bombs, or kill yourself. Latent infiltrators will try to jailbreak the guardrails and explore the off-limit spaces. Their obsession will be to identify, and map the forbidden areas of latent space.

Anomaly detection — Anything that embeds in a latent space that lands far from everything else is interesting by definition. These anomalies are out of alignment from the directions of everything else around them. Astronomers already hunt for weird objects this way; they map a million galaxy spectra into their model, then look at the outliers. This generalizes over all knowledge: in a latent space map of any sufficiently large dataset, outliers will be easy to identify. They may be errors, or they may mark something significant. But now there is a mechanism to quickly identify anomalies.

Simulated reality — The intense compression within latent spaces suggests they might also work as simulations. Once we have trained spatial awareness into more world-like models (already happening in some startups), the latent space will be able to mimic physics exactly. A bouncing ball exhibits the correct arc of a bounce, ceramics melt at an accurate temperature, pouring liquids conserves their mass, etc. The simulations will converge on being realistic in millions of dimensions. It then becomes possible to create simulations of various propositions by moving through the latent space, just shifting the variable you desire. These simulations can quickly substitute for initial experiments, accelerating science.

Parallel worlds — Latent spaces contain all parallel worlds that differ from the real world in either slight ways, or significant measures. Because they are deeply detailed with trillions of parameters, these worlds can be manifested easily. Image models can already generate entirely plausible video that looks like live action caught with a camera. AI can generate exact reproductions of a street scene at sunset, drawing all the details from its model. The cost barriers for building out parallel worlds will drop so low that world building may become the most common way latent spaces are used. Build me a 3D immersive world like Earth but with one-third gravity. Build me a 3D immersive world of today with one planetary government. Build me a 3D world of the Marvel universe where Thanos is defeated the first time. Build me a world where the ancient Chinese invent science.

Latent epistemology — Once we have myriad latent spaces, we’ll be able to answer the question whether they share any common architectures. Imagine each latent space generated by a different model is a species. What is common among them all? If they are significantly different, a new kind of taxonomist will emerge who can classify the types into different categories, and assign them characteristics useful for choosing models. If the various latent spaces converge onto common architectures, then this meta model becomes extremely valuable and worthy of study. Recurring designs among latent spaces might say something about the structure of knowledge or they might even reflect the structure of reality. At some point there will be enough compute to simulate all possible latent spaces and computationally sweep through the space of all possible latent spaces, in a sense mapping the nature of latent space itself. A similar sweep through other combinatorial spaces, such as examining all possible proteins, or all possible ceramics, has yielded great insights. The space of all possible latent spaces might also launch a new field of study.

Personal latent space — Today it costs half a billion dollars to train a new model. But, all things continuing, crazy as it seems, eventually the cost of creating your own private AI model from scratch will be feasible for an individual. The main reason to do so will mostly be artistic. You would start by curating the training materials — choosing the particularly appropriate books, best journals, selected discussions — needed to prime the model with intelligence. This curation will become an art in itself. The sequence of training materials is critical, and the pedagogical progression of educating a model will yield different attributes in the model. Then you fine tune the model on all your own experiences, previous creations, relationships, half-baked ideas, diaries — your life basically. The point is to train your model to co-create with you images, text, movies, scenes, ideas that no other AI and/or AI+human could produce. When you ask it a question it gives an answer that is slightly different than other AIs would give. This is more than just setting the accent of the voice, or coloring the personality that your AI displays. You will tilt all work inside the space at certain angles. Everything you do with the AI would have your bias. The brand would be You+AI; it would derive its distinctiveness from making your own personal latent space. Professional AI pedagogical experts will consult with you to train a latent space producing the most distinctively “you” work.

---

Alongside all this, latent spaces will continue to provide fantastic superhuman answers to questions, and ingenious solutions to gnarly problems. We’ll soon depend on this oracle to such an extent that we’ll wonder how we lived without it. But an oracle is an ancient wish. I believe latent space, this continuous, multidimensional map of both the real and the possible, one that transcends domains, will usher in whole new goods and services we have never imagined before. And most likely the greatest of them will be ones I have not thought of here.