Umwelt is the right word for what's missing, and it goes straight at the word "universal" in the title. All five roads currently assume one external world, approximated more or less well by different means. Von Uexküll's actual claim is sharper: the model isn't an approximation of a shared world, it's constitutive of the organism's world. A tick and a robot aren't holding degraded copies of the same master model, they're inhabiting different worlds entirely.
Distributed ecologies push further still, into territory closer to a different book of mine on swarm intelligence, where understanding lives in the interaction structure rather than in any single head.
There is a striking recursion here. Human language is already a model of a model: perceptual experience becomes concept; concept becomes speech or text; text becomes training data; training data becomes distributed parameters; parameters become a new generated description.
Each stage can preserve important structure, but each can also discard it. The LLM’s cave is therefore built from human caves—our perceptual and conceptual projections, compressed once more into language.
Yet the same condition gives LLMs an unusual power. They can integrate shadows cast by millions of observers, instruments, disciplines, and historical periods—far beyond the experiential range of a single person. The challenge is to bind that immense inherited map to mechanisms for calibration, experimentation, and responsible action, so that it remains corrigible by the territory rather than becoming an eloquent archive of the map.
That's a precise way to put it, and it's the question this series keeps circling back to.
The recursion you describe, perception to concept to text to parameters to output, is specific to language-trained systems. Several traditions covered later, video prediction, physics simulation, robot contact, exist specifically to shrink that chain, learning straight from pixels and actions instead of through the human cave you're describing.
Whether that escapes the recursion or just swaps one set of shadows for another is still open. But "corrigible by the territory" names exactly what's at stake. Calibration against reality, not more eloquent compression, is what every road in this series is quietly trying to buy.
Leaning into my love of Borges layered onto the Plato analogy with a little AI help.
I have been fiddling around with the mental image of shadows cast by millions of data points, far beyond the experiential range of a single person (my words), and depending on the angle the light is shown through them shifting shapes and ideas appear on the wall. Kinda like how our queries pass through the AI process to emerge as answers.
Trying out some poetic descriptions. Poetry is another type of compression that represents the highest form of human language.
What of the Umwelt road and distributed robotic ecologies?
Umwelt is the right word for what's missing, and it goes straight at the word "universal" in the title. All five roads currently assume one external world, approximated more or less well by different means. Von Uexküll's actual claim is sharper: the model isn't an approximation of a shared world, it's constitutive of the organism's world. A tick and a robot aren't holding degraded copies of the same master model, they're inhabiting different worlds entirely.
Distributed ecologies push further still, into territory closer to a different book of mine on swarm intelligence, where understanding lives in the interaction structure rather than in any single head.
Neither road is here yet. Both should be
The Recursive Cave
There is a striking recursion here. Human language is already a model of a model: perceptual experience becomes concept; concept becomes speech or text; text becomes training data; training data becomes distributed parameters; parameters become a new generated description.
Each stage can preserve important structure, but each can also discard it. The LLM’s cave is therefore built from human caves—our perceptual and conceptual projections, compressed once more into language.
Yet the same condition gives LLMs an unusual power. They can integrate shadows cast by millions of observers, instruments, disciplines, and historical periods—far beyond the experiential range of a single person. The challenge is to bind that immense inherited map to mechanisms for calibration, experimentation, and responsible action, so that it remains corrigible by the territory rather than becoming an eloquent archive of the map.
That's a precise way to put it, and it's the question this series keeps circling back to.
The recursion you describe, perception to concept to text to parameters to output, is specific to language-trained systems. Several traditions covered later, video prediction, physics simulation, robot contact, exist specifically to shrink that chain, learning straight from pixels and actions instead of through the human cave you're describing.
Whether that escapes the recursion or just swaps one set of shadows for another is still open. But "corrigible by the territory" names exactly what's at stake. Calibration against reality, not more eloquent compression, is what every road in this series is quietly trying to buy.
Leaning into my love of Borges layered onto the Plato analogy with a little AI help.
I have been fiddling around with the mental image of shadows cast by millions of data points, far beyond the experiential range of a single person (my words), and depending on the angle the light is shown through them shifting shapes and ideas appear on the wall. Kinda like how our queries pass through the AI process to emerge as answers.
Trying out some poetic descriptions. Poetry is another type of compression that represents the highest form of human language.