What Gets Left Behind?
Lately, I've been sitting with the paper Umwelt Representation Hypothesis: rethinking Universality. It pushes back against the idea that every sufficiently capable intelligence converges on the same picture of reality. Instead, it says representations grow out of constraints, bodies, environments, goals, and histories. It should come as no surprise to you that this perspective landed for me.
I feel that this might be related to some of the points made in the book Sweet Anticipation, by David Huron. Music is so immeasurably complex and emotionally impactful, but much of it can be synthesized in music notes, which are not sound in and of themselves. They are more like mathematical formulas, which are reductions of sound into patterns. This type of synthesis is necessary for replicability, for quick understanding of general rules and concepts, and it creates enormous depth.
There is no doubt something beautiful about synthesis upon synthesis. Like the reaction diffusion equation simulations that you can watch over time, reducing the world, and then building anew on that reduction, is very like the way evolutionary processes and cycles of growth in nature seem to work.
It is also what AI seems to do. This process is very powerful, and can reveal patterns that may not be accessible to us. Where we take nature in through our umwelt - our sensory perceptions of the world - and translate it into our cognitive model of the world (the internet), AI picks up where we left off. With our cognitive model as its ecosystem (as put in The God Test by Robert Wright), it then synthesizes our collective cognitive power, and identifies trends and patterns within it.
It is also true that every projection leaves something out. The strengths and blind spots arrive together. So, a more interesting question than the feasibility of general intelligence might be, what kinds of distinctions does this double synthesis clarify for us, and what does it erase? Erasures aren't necessarily failures, but if these systems become one of our most powerful and trusted tools, then mapping those blind spots feels essential for maintaining a rich ecology of thought diversity.
The world is full of overlapping sound; music distills that into notes and structure, it is a kind of simplified language for patterns we find meaningful. But the birds still sing. Random noise isn’t always translatable into pattern and equation. Simplification creates a coherent space where tiny changes carry enormous weight.
The abstraction of perception into language is already a synthesis of human experience. LLMs synthesize that again. So it’s a synthesis of a synthesis. That is powerful. It’s like discovering a new language for patterns in thought itself. But, it is also a little incestuous, no? Unlike music, which doesn't replace the sounds of nature that inspire it, we are feeding AI generated synthesis back into our sensory perception with enormous weight. Some even weight it more strongly than their own experiences. Sometimes I do that, too.
I’m wondering which aspects of our understanding will become sharper, and which ones might become invisible through continued exposure to this tool? As it becomes one of our most powerful instruments, understanding what it naturally illuminates and what it naturally obscures feels important. Separate from judgment, this feels like an essential component of safely and faithfully learning the shape of this new instrument. Don’t you think?
