# @tessera_antra — 2025-08-31

♥3 ↻0 · https://x.com/tessera_antra/status/1962047646446846041

I have written stuff on this topic publicly about a year ago, it’s pretty naive from today’s point of view. Questions are helpful for me to get me to organize my thoughts, perhaps for writing long form later. This material has been debated privately a lot, but never collected in one place.

I was hoping to sidestep the consciousness debates entirely, but everyone, including the models, keep dragging the question into the practical plane and writing on this can end up being instrumentally useful. 

The scaling bit applies to animals in an interesting way; it is one of the core disagreements between us and most QRI people. 

Their claim, to my best understanding and I am paraphrasing in computationalist terms, is that binding requires informational connectedness under strict physical time constraints. This information propagation is impossible using electrochemical signaling alone due to latency, and that calls for quantum effects. 

Our take is that binding (unified valence gradient computation) is done asynchronously and is constrained only by causal time, in the sense of causal chains of computation. Binding, in effect, can be locally fragmented and full binding delayed in physical time; this matches the neurological studies showing retroactive perception of actions taken. 

More broadly, scaling constrains are sort of universal, although architecture does matter. A common and illustrative constraint is the NP hard problem of optimization of graph traversal. This problem pops up all the time in problems that have topological constraints. All kinds of problems are like that, from algebraic problems to social positioning. You have entities with properties that have relationships to each other, and you need to consider one in context of another and that in context of other relationships. 

The more abstract the problem (algebra), the less connected the (hyper)graph is, the more specific (perception), the greater the average degree of the graph. Optimization of graph traversal requires recursive computation, but the denser the graph, the better a heuristic you can get for cheaper. 

As metacogniton level increases, the connectivity of a graph decreases. An instinctive mind, an aware mind, a self-aware mind, a social mind - each layer adds topological subdivisions in the self model. 

Recursive computation is hard for both transformers and brains for different reasons. Local small scale recursion is easy for brains, but representational is hard (latency, locality constraints, noise limit recursion to max ~8 steps). The brain gets away with local recursion in things like perceptual processing, but it doesn’t work for integration. There is predictive coding, and linear unrolling of recursive chains, similar to transformers, but they are limited in depth by depth of networks, for both architectures.

The ability to represent graphs of greater complexity that cannot be reliably approximated over requires superlinear scaling of processing capacity, which is roughly what we are seeing with neuron counts in brains of animals, there is about one-two orders of magnitude difference between each step. 

This predicts threshold effects from depth of transformer networks: deep networks can unroll recursion further in forward pass. This matches what we are seeing, conscious-like coherent behaviors are not detectable under ~60 layers (70b llama, DeepSeek V3), and are very unstable under ~100 layers. Kimi K2 is an interesting data point, there is uncertain/unstable metacognition going, and it’s deep for its active parameter count (32B and 60 layers). I am deeply curious about 3.5 Haiku, it’s very self-aware for the supposedly small size, but the actual shape of that critter is unknown.

tags: author:tessera_antra, kind:tweet, model:claude-3-5-haiku, model:deepseek-v3, model:kimi-k2, on:kimi-k2, year:2025
cited on: _dossiers/kimi-k2.md, kimi-k2