Maybe I am misunderstanding the paper. But my understanding after staring at the equations for a long time is that it claims (without evidence) that an intelligent system undergoes some kind of phase transition to self learning when it uses less than 10^-7 joules per bit per second with 70% efficiency and the information propagation velocity squared exceeds 1 meter per second. Is this correct? And where do these numbers come from?
-- Matt Mahoney, [email protected] On Sun, Aug 31, 2025, 8:44 PM Rob Freeman <[email protected]> wrote: > Dorian, > > Ah, well I'm sympathetic to an approach looking at this as a problem of > dynamics (though I'm not sure what "memory-energy" coupling could mean.) > > Colin Hales, who I think has corresponded here, has an electrodynamic > field model which would seem to be more strictly embodied. He says it must > be implemented as electrodynamic fields. Lee Cronin seems to have a similar > argument, but from a chemistry perspective. George Lakoff, embodiment... > basically as neurons, anyway. (The whole embodiment thing became something > in the fields of linguistics which retained a basis in data after Chomsky: > Functional, Cognitive Linguistics, and indeed generally in the "corpus > based" fields most directly connected with machine learning. And quite > rightly. They saw something. Insisting on basis in a "corpus" is an > embodied form of non-compression. "Corpus" = body. Embodiment is a form of > the non-compressibility argument, as I say. You need the whole corpus, said > Corpus Linguistics. This as opposed to Chomsky, who argued language could > be... not compressed... the whole point became that it could not be > compressed, that's why Chomsky still rejects machine learning. Chomsky > insisted any abstraction must be innate, exactly because it could not be > compressed/learned. At one level, observable compressions contradicted. So > linguistics resolved to embodied, or innate. But then LLMs ignored > linguistics and went and "learned" over corpora anyway!) > > But you're not trapped by the embodiment interpretation of this. Good. Let > alone Chomsky's "unlearnable" innate structure. You see a solution in > dynamics. Good. (Once a dynamical system becomes chaotic, it does become > embodied in a sense, but chaos is not limited to one embodiment.) > > So what are the parameters of this dynamics? You say "certain > energy-structured conditions (e.g., coherence, memory capacity, signal > velocity) are necessary". > > "Coherence" I'm sympathetic to. > > I've been pushing the idea of a dynamical system solution parameterized by > the "coherence" of oscillations. Driven by network symmetries of > prediction. A dynamical system parametrized by similarity of context, > anyway. A dynamical system in the sense the patterns grow and change. > Actually I think its evolution is probably chaotic on some level. > Concurring with Walter Freeman on that, from the neuroscience field. > > Checking your link, you try memristors. > > "emergent, quantized intelligence, analogous to a phase transition" sounds > good. Phase transitions being a big theme of Walter Freeman. > > "adaptive behavior arises intrinsically from the physics of the > substrate". Yes. > > Personally I think the crucial parameters are the ones already identified > by LLMs: shared context or shared prediction. If you can harness dynamics > which depend on those, it doesn't matter what your substrate is. I got > focused on the potential of (neural activation) oscillations in a network > of neurons representing language sequences. Just because synchronized > oscillations struck me as dynamics to capture those parameters of shared > context/prediction. > > From a casual glance at your paper I'm unable to tell if the parameters of > your dynamics are also shared context/prediction. If they are, then it > might be good. > > I feel you might be looking at static attractors in some sense though. > Some kind of "bubble memory". Which then has meaning... how? I can't find > the word "meaning" anywhere in your paper. Instead you have "adaptive, > feedback-driven reconfiguration". So it seems you make something of the > novelty of reconfiguration, but this has value only because of "feedback". > So the system will give meaning to these reconfigurations by some kind of > feedback from the environment? > > That sounds to me like George Edelman's Neural Darwinism. Endless random > reconfigurations, which are selected for meaning by the environment (like > his Nobel Prize winning immune system insight.) > > Of interest to me, I recently came across Eugene Izhikevich's > "polychronizations". Stable sequences appearing (from the dynamics) > spontaneously in networks of neurons. But Izhikevich didn't attribute > meaning to these based on the way the sequences shared contexts either. > > It looks to me like you have the insights about dynamical systems, and the > power of reordering/reconfiguration. But (like Edelman and Izhikevich) you > may be missing shared context/prediction, as the key parameter. Making > "meaning" internal to the system too, and not needing to be imposed by any > external feedback. > > -R > > On Mon, Sep 1, 2025 at 2:23 AM Dorian Aur <[email protected]> wrote: > >> Electrodynamic Intelligence doesn’t argue that learning is tied to a >> specific biological embodiment or that there’s only one valid substrate. >> What it does argue is that *learning occurs from the physical dynamics >> of signal propagation and memory-energy coupling,* not from abstract >> optimization rules. This isn't an anti-compression or anti-symbolic stance, >> it is a move toward *physically embedded, self-organizing systems *that >> learn because of how they are built, not because of what we train them to >> do. >> We agree that “can’t be compressed” doesn’t imply “must be embodied in >> one way.” EDI doesn’t claim there's only one valid realization of learning >> or intelligence. Rather, it highlights that certain energy-structured >> conditions <https://zenodo.org/records/16997063>(e.g., coherence, memory >> capacity, signal velocity) are necessary for physically grounded >> intelligence to occur.. >> Propagation-driven learning is not about memorizing complexity but about >> allowing local physical dynamics to shape the system's functional structure >> over time. Compression is secondary, what's primary is whether the system >> can self-organize adaptively through physical interaction to generate >> self-learning and active memory consolidation. >> >> --- Dorian Aur >> > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/Ta9b77fda597cc07a-M2ce345fbb62815fcb868ce18> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta9b77fda597cc07a-M637794d719218156ae6efa55 Delivery options: https://agi.topicbox.com/groups/agi/subscription
