Optical Illusions are Out of Distribution Errors
Our visual black box is trained on what we see in the real world. We don't process raw sensory data - we keep a more abstracted model in our heads. This is a blessing and a curse - it may either be more useful to have a sparser cognitive inventory, and work on a higher level, or to ignore preconceived notions and try to pick up details about What Actually Is. Marcus Hutter would have us believe this compression is itself identical to intelligence - certainly true for the representation once it hits your neurons. But as a preprocessing step? It's hard to be confident. Is that summarisation by our visual cortex itself a form of intelligence? Ultimately, we reduce a three-dimensional world into a two-dimensional grid of retinal activations; this in turn gets reduced to an additional layer of summarisation, the idea of the "object". We start to make predictions about the movement of "objects", and these are usually pretty good, to the point where our seeming persistence of vision is actually a convenient illusion to disguise that our vision system is interrupt-based. And so when we look at a scene, rather than even being represented as a computer does an image, as any kind of array of pixels, there is in actuality a further degree of summarisation occurring; we see objects in relation to each other in a highly abstract way. Any impressions you might have about an object having a certain location, or appearance, only properly arise as a consequence of intentional focus, and your visual cortex decides it's worth decompressing further. This is in part why blind spots (literal, not figurative) can go unidentified for so long: persistence of vision can cover them up. You're not really seeing the neuronal firings, it's a level lower than "sight". It probably isn't possible to do so consciously - we only have access to the output of the neural network, and its high-level features; these more abstracted receptive fields are all we ever really see. This is what makes optical
(epistemic status: if the following describes an already known and well-studied object in the LLM literature please point me in the right direction to learn more. but it is new to me and maybe new to you!)
I've spent most of this week constructing and plotting what I'm terming "holographemes" because what's the point of doing science if you can't coin dumb jargon, and we're in the golden age where "mad linguistics" is finally becoming a real branch of mad science. They're next-token prediction trees over a known percentage of the full probability distribution from an LLM, up to the point of an end-of-string token. In this sense they're like a "holograph" of... (read 385 more words →)