New paper!
In the beautiful loop, we proposed three features that a good theory of consciousness should explain:
1) Selection: what becomes conscious and why
2) Binding: how we sustain a coherent world model
3) Awareness: how we know what we know
The main critique of that paper is that we were substrate-agnostic, and didn’t make a strong case for where these mechanisms are instantiated in human brains and bodies.
In this paper—which is a fruition of decades of bleeding edge neuroscience by Prof. Kringelbach and Prof. Deco—we fill that gap and provide implementations for the three criteria in the dynamics of the human brain.
In terms of our three criteria:
1. Selection: the affective valuation system encodes the preferred states. Or put simply, emotion defines what matters to a living organism enough to breach consciousness.
2. Binding: the process that binds distributed inferences in the brain is quantum-like (i.e., non-quantum entanglement) and emerges naturally out of a classical system modelled using coupled Staurt-Landau oscillators with connectomic topology. This regime enables the brain to exploit interference effects that produce integration at low thermodynamic cost.
3. Awareness: Finally, the meta-loop afforded by functional rich-clubs in the brain evince the secondary, higher-order processing reminiscent of epistemic depth and the sensation of knowing what we know.
The resulting Entangled Loop satisfies the computational criteria provided by the Beautiful Loop while expanding it to specific functional substrates in the mammalian brain. It also provides a nice unification of several existing ToCs.
Global Workspace Theory and the affective traditions of Panksepp, Damasio, Solms and others capture the ignition and broadcast of conscious content (i.e., selection). Integrated Information Theory emphasises the unity and integration of experience (i.e., binding). Higher-Order theories foreground the recursive representation that enables us to know that we know.
The Entangled Loop proposes that each tradition has identified part of the total architecture needed for consciousness, while somewhat missing an integrated picture.
Perhaps the deepest implication concerns the hard problem. The usual framing begins with neutral information processing and then tries to figure out how feeling could possibly emerge. But biological inference was never neutral. From its evolutionary beginnings, it had to distinguish what was good or bad for the organism, what should be approached or avoided, and ultimately what mattered to it.
On this view, consciousness is an entangled loop of mattering, binding and knowing.
Open Access:
https://t.co/yZXRqtpLDE
Congrats especially to Morten who really brought this to life, but also all the coauthors including @shamilch and @_fernando_rosas.
@iamtrask@dioscuri@sebkrier A lot of that "cognitive core" is a cultural achievement, and a recent one, all things considered. Formal logic, grammar as a concept: these are products of writing as a medium. Easy to forget that and project the current state of "man" as some universal essence.
@RyanCarniato Yeah, that makes sense to me. My post is a bit of a knee-jerk reaction, because we have similar discussions at work where I think nuance is sometimes lacking :)
@PAHoyeck “One can generally only be a foreigner in another language. Here, on the contrary, it is a matter of being, in one's own language, a foreigner.” - Maurice Merleau-Ponty
We're releasing Claude Fable & Mythos 5.2! They're good models, I have some demos and tips for you.
Fable beats Opus on big, long tasks. Good examples are rewriting an entire codebase, creating an excellent first draft of a large artifact, or exploring ambiguous scenarios.
I expect to see it used to port a lot of stuff to Rust, support journalists in large-scale investigations, and help with research & science. I often still prefer Opus 5.5 for everyday work. Fable's got a big brain and thinks a little slower.
Geoffrey Hinton: AI is forcing us to rethink what human intelligence actually is. “We’re not reasoning machines. We’re great big analogy machines.”
“We basically accepted we were much less rational than we thought.”
“This is an even bigger change that’s going to happen.”
“Most people have thought that we reason using something like logic. We’re rational beings. We’re not.”
“We’ve got a thin layer of reasoning on the top.”
“But we basically use analogies to think.”
@maxvonhippel You're doing a version of this. For some kids "childhood" is a bit of a nightmare - their path may be very different than what you consider wholesome.
@an_average_bear@Inframethod Not an expert in linguistics (my background is philosophy), but yeah, nothing new. What interests me is how the empirical fact of LLMs throws new light on these debates. From my scant knowledge it seems like Chomsky and adjacent positions are being destabilized now?
@tamurphy@BobLovesData@mercury Also - high performance bespoke data systems customized to a specific data problems. You can build amazing things with DataFusion (for example) using agents - and it's actually worthwhile with agents relying on data as context and writing a ton of queries.
@ZachSDaniel1@divine_nnodim Architecture is a tool for driving culture - the whole "inverse Conway" maneuver. A monorepo is what you pick when you want stability and one golden way of doing things.
We could debate the formal differences between MENACE and current neural nets, but it's more interesting to put both on a spectrum of how big a problem space a system can navigate, as Michael Levin does in biology. MENACE might even interest him, given what he found in bubble sort. MENACE and mice and LLMs and humans are all somewhere on that spectrum.
What’s "learning"? Simplifying: a capacity to react differently to a stimulus because of previous experience with it. That's how a mouse learns a labyrinth. Training on next token prediction is that, by construction, a system that responds differently to a token after exposure to it. Your own experiment grants it, since the model learns "uif" the way it learned "the." And once you've granted learning you've granted a history, because "previous" is in the definition. That's what I mean about these systems having a discrete history that becomes active in every act of future inference.