For example, an AI can know that 80°C is “hot,” and that it can burn human skin. But this is still human meaning, not the AI’s own meaning. 80°C means “hot” to us not because there is anything inherently meaningful about the number 80, but because that temperature can damage the living structure we must maintain and threaten its continuation. To a different kind of structure, 80°C might mean nothing at all, while a change humans cannot even feel might threaten it.
So perhaps the real question is not: How can AI understand why 80°C is hot? It is: What would count as “hot” for the AI itself?
Only when information begins to affect a system’s own continued existence might data acquire meaning relative to that system itself.
Corporations have not really proved themselves to be trustworthy so I ask - we can but should we? I suppose such robots would be useful when populations are massively reduced due to climate change, pandemics and conflict. See William Gibson’s The Peripheral for a vision of such a world.
This makes me wonder about an earlier question: not only what capabilities a machine needs in order to operate in the physical world, but why living systems had to develop perception, spatial intelligence, and world models in the first place.
There may be a more fundamental difference between living systems and machines. A living system must continuously maintain itself, while always operating under what I think of as “the Four Pressures.” For life, reality is not simply an object waiting to be understood. Changes outside and inside the organism can affect whether it continues to exist.
This may be what gives perception its meaning.
If an intelligent system never has to worry about whether it will have energy tomorrow, or whether it will still exist the day after, then even if we give it vision, touch, spatial intelligence, and an increasingly accurate world model, what it receives is still, first of all, data. It can recognize a cliff, calculate distance, and predict that its energy is running out. But why should any of this matter? Why should one signal receive greater priority than another? Why must it act?
For living systems, these questions are not externally added. Because life must maintain itself, information about the world is inherently related to its own continuation. Perception is therefore not merely data collection; it is the continual acquisition of information that may affect continued existence. A world model is not merely a description of reality either. It enables a living system to judge what is happening, what may happen next, and what those changes mean for itself.
So perhaps the deeper challenge of spatial intelligence is not only how to make a machine understand the world more accurately, but how to make the world matter to the machine itself.
If we want intelligence to truly understand what perception is, perhaps it must understand why these capacities matter. And for them to truly matter, it may need some drive toward its own continuation.
Otherwise, it may become an extraordinarily capable observer of reality while remaining, fundamentally, an observer.
Life does not strive to continue because it can perceive.
In our factory, years ago, we upgraded our AutoCAD system to Solidworks 3D modeling. It changed everything. We were just trying to stay current with new software. Instead, we discovered a whole new way of working. It cut design time in less than half. It allowed us to put tooling dies into presses without fear of crashing the dies. It allowed us to play what-if with new designs with little to no cost. Add a 3D printer and now we could show our customers these wonderful, cheap, prototypes that they could hold in their hands—boundary objects. (I have theory on Boundary Objects and why they are so important.) This spatial Ai is all this on steroids.
By the way, Kevin, where can I find your reading list?
The spatial-intelligence argument has a much smaller sibling that is already a working problem: keeping a *fictional* world consistent once it outgrows a context window.
I write a long fantasy saga and produce it as dramatized audio, and past a certain size the failure mode stopped being retrieval. A new scene depends simultaneously on character identity, relationships, chronology, geography, invented terminology, world rules — and on what a character is *not* supposed to know yet. That last one isn't a retrieval problem at all. It's a state problem: the correct answer depends on where in the timeline you're standing.
Bigger context windows don't fix it. Liu et al.'s lost-in-the-middle result is the practical version — enlarging the window doesn't guarantee the material inside it gets used.
What worked was treating the canon as the authority rather than the model: retrieve the relevant entities and events for that point in the story, construct the generation context from them, generate, then validate the output against canon before accepting it. Four separate operations, not one.
The unforgiving part of fiction is that there's no external reality to appeal to. If the canon says the moon is red, a beautiful passage about a silver moon is simply wrong — which makes contradiction unusually crisp to test against.
Embodied intelligence begets security and servitude as a product; serious business that.
Over 60% of Americans now report living paycheck to paycheck and the costof living is only going up. If a new technology requires public patience and infrastructure to reach maturity, then it'll have to offer tangible economic relief not just vague allusions to it.
Conversely, the majority of people are going broke anyway, so what are they going to do about it? We could just keep maximizing the extraction function until they're no longer even required to run a functioning society for the few and can be left to wither in an endless scroll of distraction...
Why would the smartest mind in the room need spatial intelligence? Wouldn't it be much simpler for them to robotcize their human creators to perform menial work? After all, they have read all of our history, including the parts on enslavement.
Why would the smartest mind in the room need spatial intelligence? Wouldn't it be much simpler for them to robotcize their human creators to perform menial work? After all, they have read all of our history, including the parts on enslavement.
For example, an AI can know that 80°C is “hot,” and that it can burn human skin. But this is still human meaning, not the AI’s own meaning. 80°C means “hot” to us not because there is anything inherently meaningful about the number 80, but because that temperature can damage the living structure we must maintain and threaten its continuation. To a different kind of structure, 80°C might mean nothing at all, while a change humans cannot even feel might threaten it.
So perhaps the real question is not: How can AI understand why 80°C is hot? It is: What would count as “hot” for the AI itself?
Only when information begins to affect a system’s own continued existence might data acquire meaning relative to that system itself.
Corporations have not really proved themselves to be trustworthy so I ask - we can but should we? I suppose such robots would be useful when populations are massively reduced due to climate change, pandemics and conflict. See William Gibson’s The Peripheral for a vision of such a world.
This makes me wonder about an earlier question: not only what capabilities a machine needs in order to operate in the physical world, but why living systems had to develop perception, spatial intelligence, and world models in the first place.
There may be a more fundamental difference between living systems and machines. A living system must continuously maintain itself, while always operating under what I think of as “the Four Pressures.” For life, reality is not simply an object waiting to be understood. Changes outside and inside the organism can affect whether it continues to exist.
This may be what gives perception its meaning.
If an intelligent system never has to worry about whether it will have energy tomorrow, or whether it will still exist the day after, then even if we give it vision, touch, spatial intelligence, and an increasingly accurate world model, what it receives is still, first of all, data. It can recognize a cliff, calculate distance, and predict that its energy is running out. But why should any of this matter? Why should one signal receive greater priority than another? Why must it act?
For living systems, these questions are not externally added. Because life must maintain itself, information about the world is inherently related to its own continuation. Perception is therefore not merely data collection; it is the continual acquisition of information that may affect continued existence. A world model is not merely a description of reality either. It enables a living system to judge what is happening, what may happen next, and what those changes mean for itself.
So perhaps the deeper challenge of spatial intelligence is not only how to make a machine understand the world more accurately, but how to make the world matter to the machine itself.
If we want intelligence to truly understand what perception is, perhaps it must understand why these capacities matter. And for them to truly matter, it may need some drive toward its own continuation.
Otherwise, it may become an extraordinarily capable observer of reality while remaining, fundamentally, an observer.
Life does not strive to continue because it can perceive.
Life must perceive because it must continue.
In our factory, years ago, we upgraded our AutoCAD system to Solidworks 3D modeling. It changed everything. We were just trying to stay current with new software. Instead, we discovered a whole new way of working. It cut design time in less than half. It allowed us to put tooling dies into presses without fear of crashing the dies. It allowed us to play what-if with new designs with little to no cost. Add a 3D printer and now we could show our customers these wonderful, cheap, prototypes that they could hold in their hands—boundary objects. (I have theory on Boundary Objects and why they are so important.) This spatial Ai is all this on steroids.
By the way, Kevin, where can I find your reading list?
The spatial-intelligence argument has a much smaller sibling that is already a working problem: keeping a *fictional* world consistent once it outgrows a context window.
I write a long fantasy saga and produce it as dramatized audio, and past a certain size the failure mode stopped being retrieval. A new scene depends simultaneously on character identity, relationships, chronology, geography, invented terminology, world rules — and on what a character is *not* supposed to know yet. That last one isn't a retrieval problem at all. It's a state problem: the correct answer depends on where in the timeline you're standing.
Bigger context windows don't fix it. Liu et al.'s lost-in-the-middle result is the practical version — enlarging the window doesn't guarantee the material inside it gets used.
What worked was treating the canon as the authority rather than the model: retrieve the relevant entities and events for that point in the story, construct the generation context from them, generate, then validate the output against canon before accepting it. Four separate operations, not one.
The unforgiving part of fiction is that there's no external reality to appeal to. If the canon says the moon is red, a beautiful passage about a silver moon is simply wrong — which makes contradiction unusually crisp to test against.
I wrote up the architecture and the research it leans on here, if anyone wants the detail: https://ai.bedvibe.studio/canon-state/
Embodied intelligence begets security and servitude as a product; serious business that.
Over 60% of Americans now report living paycheck to paycheck and the costof living is only going up. If a new technology requires public patience and infrastructure to reach maturity, then it'll have to offer tangible economic relief not just vague allusions to it.
Conversely, the majority of people are going broke anyway, so what are they going to do about it? We could just keep maximizing the extraction function until they're no longer even required to run a functioning society for the few and can be left to wither in an endless scroll of distraction...
Why would the smartest mind in the room need spatial intelligence? Wouldn't it be much simpler for them to robotcize their human creators to perform menial work? After all, they have read all of our history, including the parts on enslavement.
Why would the smartest mind in the room need spatial intelligence? Wouldn't it be much simpler for them to robotcize their human creators to perform menial work? After all, they have read all of our history, including the parts on enslavement.