Mugur Marculescu

New paper: Labels as Computational Primitives. The claim: a neuron's label isn't a comment. It's a compressed behavioral spec implying the unit's decay, threshold, bias, and connectivity. Which means a working neural network can be compiled from language instead of trained. zenodo.org/records/21523885

Restless Accordion: when a point's color sets its own gain youtu.be/CwqAf4mWJNU?si=RPVl…

What does it take to compute? youtube.com/watch?v=vTKPyMTO…

LLM guided neural transcription works.

Would love to see @arcprize consider energy as part of the scoring dynamic in ARC-AGI-4!

The thing everybody gets wrong about world models is the world part.

Two neurons walk through a gate. One turns to the other: “Looks like we share a prior.”

As the LLM approach saturates model capacity (for a given training corpus)... the intelligence balance moves to the questions, not the answers.

Complexity from recursive polarization implemented in compute shaders.

Self relating medium compute shader graph implementation.

Both sides of the AI consciousness debate are making a category error. Consciousness isn’t a property a system has or lacks; it is the ordinary medium that carries state and acts on it. Computation is the process. Without that self-relating medium there is no computation, only a description of one. Irreducibility is why you can’t check the answer from outside; you have to run it. Graphs matter because sequential loops can’t co-instantiate the state being reflected on. The question “is it conscious?” already assumes the wrong ontology.

consciousness is the most abundant and ordinary thing in the universe and computation is the only process

fundamental particles are the ones doing matrix multiplication

Sonnet 5.5 and Opus 5.5 have nearly converged... it's just a pricing position matched to an effort level label offset. ⚖️ The more important thing to consider is the reason... is this because of a step function in the Sonnet 5.5 training lapping the progress speed on Opus (this is hard take off on hard take off)? or they are both maxing out... as in perfection converges on the ceiling?

A different shape of "machine"

When execution gets cheap, judgment about direction gets scarce.

LLMs accidentally and implicitly started to implement graph loops (neural loops) the moment we started feeding their output back in as input. this is what thinking is it’s just on the wrong architecture.

/btw in Claude Code is about as close as you can get to a counterfactual experiment without disturbing the system.

Opus 5.5 passed the test. @AnthropicAI

matrix multiplication = graph propagation