Another banger! I like the idea of latent infiltrators. I use LLMs like noun particle accelerators at times just to see what the connections in the space are.
We need to be highly cautious about framing "correctness," "completeness," or "cohesiveness" as objective vectors or directions within latent space. LLMs are maps of maps tracking humanity’s connection to physical reality and our commentary about it at the same time. If a model were trained exclusively on medieval medical texts, its vectors for "healing" would point with absolute mathematical confidence toward bloodletting and balancing humors.
Rather than tracking objective truth, the LLM is tracking Stephen Colbert’s concept of "Truthiness." It isn't measuring facts; it is measuring the social density of our assertions about reality. This actually makes distinguishing vastly harder at the edges.
It’s not about better or worse. He’s effectively trying to understand text as speech. Unless you think all communication reduces to a single mode, then it’s just confusing - as this essay is clearly confused
Kevin has “default” ontological commitments but that doesn’t make the essay confused. You’re the one that needs to establish what else that we should be “seeing” or else you’re just telling us we are wrong.
And I am preparing to all for academic journals which these folk don’t read - it’s all about getting trashy “engagement” rather than actually contributing.
thus my plea that he thinks about it more and hopefully gets to something interesting to professionals.
Huh? His target audience isn’t professionals tho…which means your own critique of the triad is misplaced…you’re not in this circuit of meaning and just complaining about it
this is easily the most mind-bending thing i've ever read - and the best explained.
would it be acceptable for me to post a link to your article on my page on Facebook? unlike Substack, i'm up against huge resistance to AI, there; this article might at least stop one or two people in their tracks and make them realise what a truly extraordinary time we are living in.
I have even been wondering if AI is for everyone. I can hardly find one person in my arts community willing to be curious about it. They do watch what I do with it with great awe, because it’s beautiful. But the fear is palpable.
it’s early days. people reacted like that to the internet, at first. fear, distrust - eventually they came on board (and practically capsized the boat).
I wonder if their improvements will follow a similar cost curve to that of genome sequencing. Was millions and millions when I was in high school, now it’s 500$
Great collections of ideas. We actually published a paper around ten years ago where we did this on a cheap laptop, code written by the student from the ground up, using not LLMs but Restricted Boltzmann machines. We used various ways to explore latent spaces, using photos from the brilliant art project Exactitudes by Ari Versluis and Ellie Uyttenbroek. See Verkoelen et al., https://doi.org/10.1007/978-3-319-55750-2_22 and preprint https://surfdrive.surf.nl/s/yX2jxskENxGD2Lj
Kevin, thank you for this. As a non-technical user I've been pushing into a few of these use cases in Claude and ChatGPT without knowing what to call them — your cross-domain analogy finally gave me the language. I'm a non-linear thinker, so I treat my chats as a thought partner, not a "go do this" machine (I use Cowork for that). I already see where things correlate across five or six domains, and I bring those parallels in. I wince every time it gushes back, "Oh, this is so novel!" Now I have a better understanding of why that response comes back, plus a term for what I've been doing. I'm exploring Latent Space. How cool is that?
This is a great article! I've been reading your books and articles for years, and more recently have been a lurker on your Substack. I especially enjoyed and often reference your book What Technology Wants. I'm an MIT grad and currently run an online university.
I've been recently experimenting with the idea of a Latent Space as a Creative Canvas through both vibe fiction writing and vibe music writing. The way I've framed it fits exactly how you describe as working as an artist with the latent space as my canvas.
If you take the latent space of [Cyberpunk X Gospel X Story] then you get the serial podcast story of https://www.thefounderparable.com
Much more broadly, if you take the idea of simulation theory, you could image God as a programmer who turns the latent space of all possibilities into our universe by creating a physics engine based on nine equations: https://www.thefounderparable.com/p/universe-source-code
The most interesting shift here is that latent space is not just a compressed archive. It is becoming a navigable design surface.
Search asks: “Where is the answer?”
Latent exploration asks: “What exists between the answers we already know?”
That changes the role of AI from retrieval engine to structural discovery tool. The real opportunity may be in finding white space, cross-domain analogies and unnamed patterns that no single discipline has the vocabulary to see.
The model is no longer just generating outputs. It is exposing the topology of possibility itself.
Might there be a problem that with a latent space so vast and with all of human symbolic output compressed into it, that archetypes, demons and other such things will spontaneously arise, and whether you treat them as eigenvectors, attractors or as entities, they may well have some degree of psychological potency for the user?
This description is quite beautiful. It does bring up a question about AI hallucinating. We apply our need for certainty to AI but what if uncertainty is the only way to find new patterns? The uncertainty of Bateson - framing it as the gap between digits is where chaos is the only source of new patterns. This may be what we can learn from AI. When it brings back a “hallucination” it is because it is hunting those edges that are in the “not-yet”. But my AI said that sometimes that is the most interesting pattern to offer. If certainty is your measure of value then it will look like a mistake. But that’s how great science and art happens. We trust the universe and dare to be curious. “what if” is my motto.
Great piece, I am teaching something in this vein right now at the School of Visual Arts. We have only begun to scratch the surface of this new medium of AI. Like most new mediums, it has to imitate something we know before it can find its own place in the world, and this latent exploration seems like it has amazing potential. My own work looks at the connection between moral foundations and how we experience weather. Concepts that need this latent space to connect easily. Thanks for writing this piece and giving me another source to share with my students.
This made me think about a very specific moment I recognise in two quite different parts of my life.
I’m an architectural designer, and I also play drums. Starting a design project and sitting down at the drums can feel remarkably similar: there is a field of possibilities, but no resolved path yet.
I make one move — draw a line, place a space, play a rhythm — and that changes what becomes available next. Then another move responds to that one, and the thing gradually takes form.
So the part of latent space that interests me most is not really “finding an answer”. I’m not sure there is ever a final answer to find. It feels more like refinement through participation — moving through possibility, with each step changing what becomes available to the next.
"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." - Could not agree more. This means not only the things of Physical AI as in the "World2Vec" approach, but all concepts nested in a relational hyperspace.
The Spatial Web provides the infrastructure:
- Hyperspace conceptual model for any entity and its relationships
- Hyperspace Modeling Language (HSML) as the basis of a web of
- Universal Domain Graph (UDG) as the graph of graph of all entities with coordinates in hyperspace.
The UDG is the latent space defined in your blog. Further interesting is the convergence of independently-trained models having the same structures for concepts. This is the basis for features and maps in the latent space of the Spatial Web.
So does this effectively mean that all the knowledge contained within the LLMs is essentially a secondary source; in other words, it's strictly taking the meta-data from all the source material itself and using that to create it's knowledge base, but it doesn't actually contain any of the source material itself? Sounds like the beginning of a long game of telephone.
Another banger! I like the idea of latent infiltrators. I use LLMs like noun particle accelerators at times just to see what the connections in the space are.
We need to be highly cautious about framing "correctness," "completeness," or "cohesiveness" as objective vectors or directions within latent space. LLMs are maps of maps tracking humanity’s connection to physical reality and our commentary about it at the same time. If a model were trained exclusively on medieval medical texts, its vectors for "healing" would point with absolute mathematical confidence toward bloodletting and balancing humors.
Rather than tracking objective truth, the LLM is tracking Stephen Colbert’s concept of "Truthiness." It isn't measuring facts; it is measuring the social density of our assertions about reality. This actually makes distinguishing vastly harder at the edges.
Rats, I missed an opportunity to use the term truthiness in earnst. That is precisely the right term.
" I use LLMs like noun particle accelerators at times just to see what the connections in the space are." - that is a very instructive metaphor.
You are almost there. You’re just stuck in a dyadic framework.
Would like to see you try to think through the triad.
And definitely dump the 20th century cog sci and Kantianism
Well, please establish what your triad would add to the debate.
A triad isn’t inherently better than a dyad.
So please tell us where “there” is before we can go there.
Truthiness actually implies a triad if you think about it more
It’s not about better or worse. He’s effectively trying to understand text as speech. Unless you think all communication reduces to a single mode, then it’s just confusing - as this essay is clearly confused
Kevin has “default” ontological commitments but that doesn’t make the essay confused. You’re the one that needs to establish what else that we should be “seeing” or else you’re just telling us we are wrong.
It literally exactly makes him confused.
And I am preparing to all for academic journals which these folk don’t read - it’s all about getting trashy “engagement” rather than actually contributing.
thus my plea that he thinks about it more and hopefully gets to something interesting to professionals.
Huh? His target audience isn’t professionals tho…which means your own critique of the triad is misplaced…you’re not in this circuit of meaning and just complaining about it
this is easily the most mind-bending thing i've ever read - and the best explained.
would it be acceptable for me to post a link to your article on my page on Facebook? unlike Substack, i'm up against huge resistance to AI, there; this article might at least stop one or two people in their tracks and make them realise what a truly extraordinary time we are living in.
Link please!
i shall. will probably be met by a resounding silence, but you never know! thank you - your article has really made things clearer for me.
I have even been wondering if AI is for everyone. I can hardly find one person in my arts community willing to be curious about it. They do watch what I do with it with great awe, because it’s beautiful. But the fear is palpable.
it’s early days. people reacted like that to the internet, at first. fear, distrust - eventually they came on board (and practically capsized the boat).
I wonder if their improvements will follow a similar cost curve to that of genome sequencing. Was millions and millions when I was in high school, now it’s 500$
That is correct. It does not contain the information in the form of the sources; it has its own form: latent space.
Great collections of ideas. We actually published a paper around ten years ago where we did this on a cheap laptop, code written by the student from the ground up, using not LLMs but Restricted Boltzmann machines. We used various ways to explore latent spaces, using photos from the brilliant art project Exactitudes by Ari Versluis and Ellie Uyttenbroek. See Verkoelen et al., https://doi.org/10.1007/978-3-319-55750-2_22 and preprint https://surfdrive.surf.nl/s/yX2jxskENxGD2Lj
Kevin, thank you for this. As a non-technical user I've been pushing into a few of these use cases in Claude and ChatGPT without knowing what to call them — your cross-domain analogy finally gave me the language. I'm a non-linear thinker, so I treat my chats as a thought partner, not a "go do this" machine (I use Cowork for that). I already see where things correlate across five or six domains, and I bring those parallels in. I wince every time it gushes back, "Oh, this is so novel!" Now I have a better understanding of why that response comes back, plus a term for what I've been doing. I'm exploring Latent Space. How cool is that?
This is a great article! I've been reading your books and articles for years, and more recently have been a lurker on your Substack. I especially enjoyed and often reference your book What Technology Wants. I'm an MIT grad and currently run an online university.
I've been recently experimenting with the idea of a Latent Space as a Creative Canvas through both vibe fiction writing and vibe music writing. The way I've framed it fits exactly how you describe as working as an artist with the latent space as my canvas.
If you take the latent space of [Cyberpunk X Gospel X Story] then you get the serial podcast story of https://www.thefounderparable.com
If you take that story and map it into the latent space Music, you get the Cyberpunk Gospel Musical on Flow Music at: https://www.flowmusic.app/playlist/1abe8c8c-88f4-441d-9851-66fd66b99d9b
Much more broadly, if you take the idea of simulation theory, you could image God as a programmer who turns the latent space of all possibilities into our universe by creating a physics engine based on nine equations: https://www.thefounderparable.com/p/universe-source-code
I posted a video and slides describing my approach on Substack at: https://www.thefounderparable.com/p/tutorial-vibe-fiction-writing-with and have linked back to your post in the comments. Thanks for your post!
The most interesting shift here is that latent space is not just a compressed archive. It is becoming a navigable design surface.
Search asks: “Where is the answer?”
Latent exploration asks: “What exists between the answers we already know?”
That changes the role of AI from retrieval engine to structural discovery tool. The real opportunity may be in finding white space, cross-domain analogies and unnamed patterns that no single discipline has the vocabulary to see.
The model is no longer just generating outputs. It is exposing the topology of possibility itself.
Might there be a problem that with a latent space so vast and with all of human symbolic output compressed into it, that archetypes, demons and other such things will spontaneously arise, and whether you treat them as eigenvectors, attractors or as entities, they may well have some degree of psychological potency for the user?
https://drtompollak.substack.com/p/all-the-demons-hiding-in-your-ais
Latent space reminds me of liminal space and all the possibilities between them. Great work!
This description is quite beautiful. It does bring up a question about AI hallucinating. We apply our need for certainty to AI but what if uncertainty is the only way to find new patterns? The uncertainty of Bateson - framing it as the gap between digits is where chaos is the only source of new patterns. This may be what we can learn from AI. When it brings back a “hallucination” it is because it is hunting those edges that are in the “not-yet”. But my AI said that sometimes that is the most interesting pattern to offer. If certainty is your measure of value then it will look like a mistake. But that’s how great science and art happens. We trust the universe and dare to be curious. “what if” is my motto.
Great piece, I am teaching something in this vein right now at the School of Visual Arts. We have only begun to scratch the surface of this new medium of AI. Like most new mediums, it has to imitate something we know before it can find its own place in the world, and this latent exploration seems like it has amazing potential. My own work looks at the connection between moral foundations and how we experience weather. Concepts that need this latent space to connect easily. Thanks for writing this piece and giving me another source to share with my students.
I enjoy your essays tremendously!
This made me think about a very specific moment I recognise in two quite different parts of my life.
I’m an architectural designer, and I also play drums. Starting a design project and sitting down at the drums can feel remarkably similar: there is a field of possibilities, but no resolved path yet.
I make one move — draw a line, place a space, play a rhythm — and that changes what becomes available next. Then another move responds to that one, and the thing gradually takes form.
So the part of latent space that interests me most is not really “finding an answer”. I’m not sure there is ever a final answer to find. It feels more like refinement through participation — moving through possibility, with each step changing what becomes available to the next.
"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." - Could not agree more. This means not only the things of Physical AI as in the "World2Vec" approach, but all concepts nested in a relational hyperspace.
The Spatial Web provides the infrastructure:
- Hyperspace conceptual model for any entity and its relationships
- Hyperspace Modeling Language (HSML) as the basis of a web of
- Universal Domain Graph (UDG) as the graph of graph of all entities with coordinates in hyperspace.
The UDG is the latent space defined in your blog. Further interesting is the convergence of independently-trained models having the same structures for concepts. This is the basis for features and maps in the latent space of the Spatial Web.
Status
- Developed by the Spatial Web Foundation https://spatialwebfoundation.org/
- Anchored in the IEEE 2874-2025 standard https://ieeexplore.ieee.org/document/10949052
- Further development underway. https://georoundtable.xyz/spatial-web-baseline-2025.html
So does this effectively mean that all the knowledge contained within the LLMs is essentially a secondary source; in other words, it's strictly taking the meta-data from all the source material itself and using that to create it's knowledge base, but it doesn't actually contain any of the source material itself? Sounds like the beginning of a long game of telephone.