Agency, Attractors, & Observer-Dependent Computation in Biology & Beyond

The assumption that computation has a single objective truth—that something either 'is' or 'is not' a computation independent of observers—is wrong; computation is dependent on external observers who view a set of physical events as a computation and benefit thereby, making it fundamentally relative rather than absolute.

“the computation is really in the mind of the Observer or or multiple observers more than it is a feature of some particular you know set a set of a set of physical events and so on the one hand it makes things quite relative because it means that there could be multiple observers that look at the exact same set of uh physical interactions see different things being computed interpret them in different ways”

Traditional systems of ethics based on distinguishing engineered vs. naturally evolved systems, presence vs. absence of big brains, and phylogenetic relatedness are going to become obsolete in coming decades and will be replaced by frameworks based on recognizing and respecting unconventional minds (artificial, biological, hybrid, alien-like systems).

“none of these criteria are going to survive the next decades this is all all this stuff has to go and and new new uh systems of Ethics has to come because all of this is really about how do you relate to others other systems how do you relate to things that look like you to things that look nothing like you things that don't share the same evolutionary tree with you”

Modern molecular medicine approaches gene regulatory networks by trying to rewire them (gene therapy, promoter modification) to change behavior from disease to health state; but this requires changing hardware and is therefore limited.

“if you want this thing to behave in a particular way meaning let's say to go from disease state to a healthy State you're going to have to rewire it somehow meaning you might add nodes or subtract nodes or change the weights or something like that you're going to interact with the hardware and this is what modern molecular medicine does it's it's Gene it would be gene therapy you would have to add new new you know let's say if it was a gene regulatory Network you'd add a new Gene that hits the promoter of some other gene or you know you'd change the promoters to be stronger a weak or something like that”

One of the things Evolution does is not just change the hardware but also provides new observers and new capabilities that extract more benefit from the exact same piece of hardware activity by viewing them in different ways; it's the perspective that can change, allowing you to add different ways to interpret what a machine is doing and thereby benefit and achieve adaptive fitness increases without changing the physical substrate itself.

“what Josh and I proposed in this paper is that one of the things Evolution does is not just change the hardware but it also provides new observers and new capabilities that extract more benefit from the exact same piece of uh Hardware activity by viewing them in different ways it's the perspective that can change so so you don't change the machine you add different ways to interpret what that machine is doing and and thereby benefit and get and get all you know adaptive in increases in adaptive Fitness”

The Giordano Bruno tradition (dating back centuries) held that competency and agency are not zero in non-living systems like rocks; this philosophical tradition has been overshadowed by modern mechanisms but contains important intuitions about universal agency.

“this is this is a super old idea Giordano Bruno you used to say stuff like this this is this is very old um that that it isn't zero the competency of these things is not zero it's very low but it isn't zero”

Eric Hoel's work (referenced as 'the map is better than the territory' framework) quantifies that for certain kinds of systems you gain more predictive and control power by taking seriously the higher levels (bodies, organisms, organs) rather than reducing to molecules and atoms; this demonstrates that higher-level descriptions are not just convenient but literally more powerful.

“what he basically developed was a way to quantify and make rigorous a debate that's been going on for uh probably thousands of years which is uh are there any higher levels like bodies organisms uh you know people things like this or is it or should it really be reduced to talking about molecules and atoms and whatever else is underneath I mean people have been discussing that for a long time and I so he's found a way to quantify that and his analysis shows that there are and this is of course also the work of Julio tanoni and and other people like that um uh that would have been many others that have contributed to this since then that uh basically have figured out that for certain kinds of systems you actually gain more power power means uh better ways to control and predict the system”

Chris Fields has made a rigorous argument that the only way to have a universe with literally zero agency/competency everywhere is to have a static universe where nothing ever happens; active inference and variational principles are built into the structure of physics at the most fundamental level.

“I once asked Chris um if uh if it get given all of that I said is it is it possible to have a universe with no least action principles so basically could we could we I you know our world doesn't to me our world doesn't have zero um zero agency anywhere but but the question is could could there be one could there be a universe with that had literally zero and uh Chris said that the only way to do that is to have a universe where nothing ever happens a static world where nothing changes”

As we move from systems with low agency to high agency on the spectrum of persuadability, we need to shift from ideas of control (which work for machines and thermostats) to ideas of relationship and respect, because high-agency systems benefit from and deserve ethical consideration based on their autonomy.

“but as you head towards I mean I think it's very important and then I'm writing stuff about this that'll come out at some point as you get the further you get to the right side of that Spectrum you have to shift over from ideas of of control to ideas of relationship and this is where a lot of the um you know uh ethics aspects come in”

The concept of 'hacking' in biology is morally neutral—it means one system sending signals that get another system to do things it otherwise wouldn't; parasites hack hosts, cells hack each other, and evolution has optimized these hacking relationships; the body exists because cells are constantly hacking each other.

“and then of course the better you understand some system the better you get at hacking it and hacking I don't mean the negative thing where you know you sort of abuse the the system although that certainly happens right with parasites and and various other things that absolutely happens but but it doesn't have to be negative the the the reason we have a body instead of a bag of amoebas is that these cells are constantly hacking each other they are constantly sending out signals that get each other to do various things that they otherwise wouldn't do it's Behavior shaping all you know up and down the the the the the scale hierarchy”

AI serves as a lens or translation device that will help us see the intelligence that already exists all around us in systems we don't yet recognize as intelligent, making parallel forms of intelligence suddenly communicable and visible.

“that you know one of the things I love about AI is that it's it's like a lens or or a translation device that's going to help us see a lot of the intelligence that's all around us all the time it's like being able to suddenly being able to communicate with an intelligence with a parallel set of intelligence so they were here all along we just never knew we never knew how to recognize and we never knew how to relate to them”

Intelligence and agency are not binary but exist on a spectrum called the 'spectrum of persuadability,' where different tools and protocols are needed to interact with systems at different points: mechanical clocks require hardware modification; thermostats require set-point changes; animals require behavioral rewards/punishments; humans require rational reasons and can operate with high autonomy.

“I call it a Continuum of persuadability and then that's on purpose again it's because the reason is that I'm not talking about um uh I'm not trying to pin it on on a kind of a single Universal objective fact about the system I'm looking at it from the perspective of an observer who wants to relate to it somehow to and to understand it to to um change how it works to get it to do something to manipulate it somehow to make a new one you know whatever that”

Things are real to the extent that you must pay attention to them to survive and function successfully; if you can afford to ignore something, you can reasonably treat it as not real, and if ignoring it leads to suffering or death, you must treat it as real regardless of whether you can hold it in your hand.

“which is can I afford to ignore it that that's it it's again and I'm taking I'm once again taking very technological engineering approach to this I'm saying that um things are real to the extent that you need to pay attention to them and they may be things you can hold in your hand and they may not be but to the extent that you will you you suffer by not paying attention to them that's what tells you if they're real or not”

Different planaria species have different head shapes because they occupy different attractor basins in morphological space; under normal circumstances acorns make oak trees and frogs stay froggy because evolution has selected for those attractors, but the xenobot attractor also exists in morphospace and can be reached through bioelectric manipulation.

“different shape planaria heads so we work with these flatworms these planaria different shape heads correspond to and that's the figure that you're talking about correspond to different attractors in that space because all things being equal plus or minus various um environmental influences you still you know the typically the acorn ends up in the oak tree attractor and the the you know the Frog a ends up in the froggy attractive not always because we can push it into the xenobot attractor but that's the difference that's a different that's a different story”

The benefit constraint on computation is critical: you can only legitimately treat something as a computation if you can specify what practical utility it gives you—what you can do with that interpretation that you couldn't do otherwise.

“because in treating it as a computation you have to say what does it enable you to do that you couldn't do otherwise in other words you have to benefit from that lens of looking at this set of if you're going to treat a set of events as some kind of computation you have to be able to say what is the Practical import how did you how did you benefit”

The prediction/control approach to understanding systems is incomplete; what matters more is the inventive or generative capacity—the ability of a framework to help you invent new systems and solutions, not just predict what already-designed systems will do next.

“I think prediction is is just part of the story what we're really interested in is maybe I I don't know what a what a good word for is um maybe maybe it's pre-invention or something it's it's the idea that we're not just going to predict exactly what this clever system is going to do next we actually need to be in a place to make one and the next one and the one after that and so now I'm interested in Frameworks that facilitate that I don't want to just be able to say you know somebody hands you a complicated thing and now here here's how we know what it does next I want to know how we get to the next one”

Organs like the liver and kidneys should be recognized as intelligent agents navigating physiological state space—they respond to perturbations in blood chemistry in sophisticated ways that keep you alive despite harmful inputs.

“and you would also know that there are these things called liver kidneys and so on that are really clever about navigating that space that when they get perturbed by specific things that happen in terms of your blood chemistry they can find ways around it they keep you alive despite all sorts of you know terrible things that you do to them with you you know with your lifestyle and all that um we would have no problem recognizing them as beings that live in this space we would know we live in this space we just don't we just we're very bad at recognize this recognizing this”

Life is fundamentally any system that is really good at scaling the non-zero competencies of chemistry and physics into much larger competencies, novel problem spaces, and cognitive light cones that we recognize as being on the agency spectrum; aggregating particles without this scaling (like making a rock) preserves minimal agency, but aggregating with scaling produces life.

“there are other ways to aggregate them which actually amplify what's going on here and the systems that do that that's what we call life that's that's what when when people you know I don't spend a whole lot of time trying to come up with a definition for life but but but but that's what I think life really refers to life what we call life is any system that is really good at scaling these fundamentally non-zero competencies of chemistry and physics into much larger competencies larger cognitive light cones novel problem spaces and things that we begin to recognize as being on the Spectrum”

Gene regulatory networks (which are completely deterministic sets of nodes corresponding to genes or proteins that upregulate or downregulate each other) demonstrate associative learning capacity and can be conditioned like animals when treated as learning systems rather than pure mechanical systems, showing the same network can exhibit different learning capabilities depending on how it is observed and tested.

“what we did is we decided okay could we could we do this to the to the network and so what we would do is we would stimulate one of the nodes that means you know upregulate one of the genes um do the same thing to another one that where normally the first one has no effect on whatever the response we're looking for is so we choose one of the nodes as we call it our response maybe it's you know maybe it's something that controls blood pressure or maybe it's something that is some kind of enzyme so something that we care about and so that we choose a node that always turns that on we choose some other node that's the neutral node that's like the Bell which normally has no effect on it and we just present them together we present them together and then we pause and then we present them together and then we pause and we present them together so when you do this what you find out is that uh for certain for certain networks and for certain choices of the conditioned stimulus know the unconditioned stimulus list node and the response it will actually learn to associate them”

The typical objection that observer-relativity about intelligence means 'anything goes' is wrong; claims must be constrained by engineering protocols with demonstrable benefits—you must specify what you gain from that lens.

“so this isn't the kind of um uh kind of uh loosey-goosey sort of talk where you say everything is uh you know everything is alive everything I mean you can do that but that's but that's pretty useless okay that doesn't Advance you at all what I'm talking about is a very specific uh way to to to drive research agendas which is to take tools that are useful in one field for dealing with one kind of system and ask what other kinds of systems are these tools useful for and to not be able to I mean I'm arguing against having these sort of armchair philosophical um preconceptions about what things have to be somebody somebody said to me once um we were talking about this thing and I was somebody said well well surely you don't think the weather has any intelligence to it and being perfectly serious I said has anybody tried training the weather”

A strict reductionist examining a microprocessor can correctly predict everything it will do using physics (Maxwell's equations, Schrödinger's equations) without believing algorithms are causally real, but this view makes them useless for writing new algorithms or inventing new things, showing prediction and causal understanding are different.

“so the thing is that uh that perspective it's it's it's not wrong in the sense that empirically yes we have equations that guide all these things but would you would you hire that person for your new software company I'm gonna say I'm gonna say you wouldn't because because right because anybody that doesn't believe that the algorithm is what makes the electrons dance is is it's it's not that they're factually wrong but but they're not going to make anything new they're not if they're not going to write any new algorithms if they don't if you don't fundamentally believe that there is such a thing as an algorithm that has causal power that determines what happens next you you can't participate in this whole stack of what happens afterwards”

Intelligence and agency are not properties you can decide about systems philosophically from an armchair; they are empirical questions requiring testing of whether systems exhibit habituation, sensitization, associative conditioning, or other behavioral markers under appropriate behavioral protocols.

“this is an empirical question you and I cannot decide whether weather patterns meaning patterns of air movement in the atmosphere do or do not exhibit habituation sensitization associative conditioning of some sort you and I sitting here in our chairs are not going to be able to decide that and and and a lot of people don't don't think that's the way to do it because they feel that if the empirical um you know if if those kind of empirical things uh are are the prediction of your world view then you then you're then the then the worldview must be wrong”

The question of whether there is a zero point on the spectrum of agency/persuadability is genuinely difficult; the minimal requirements for agency are goal-directed behavior (systems that tend toward particular outcomes despite perturbation) and some degree of indeterminacy (behavior not fully determined by immediate local causes), both of which are present in elementary particles through least action principles.

“when I think about the absolute minimum I want it to have two features the first feature I wanted to have any system that's on that Spectrum the first feature I wanted to have is some degree of goal directed uh Behavior goal directed Behavior means that it uh tends to achieve a particular outcome in some kind of uh State space and if it is deviated from that it will expand some minimal degree of capacity to still get there okay and so we so so and and so that's that's the first thing some kind of goal directed activity and the second thing I would expect some minimal degree of indeterminacy”

Random gene regulatory networks do not exhibit associative learning capacity to nearly the extent that biological networks do, suggesting evolution selects for (directly or indirectly) the capacity for associative learning.

“the reason I bring this up one is that uh random networks don't do this very much biological networks do this so that means that Evolution sort of uh selects for this or directly or indirectly but but the you know life life likes this capacity”

The statement 'nothing in the universe has zero agency' makes many people uncomfortable because it challenges fundamental assumptions about the inert nature of matter and the special status of life; but this discomfort is misplaced because the statement is not metaphysical poetry but an operationally-grounded engineering claim about what kinds of tools work to interact with different systems.

“so I I don't spend a whole lot of time trying to come up with a definition for life but but but that's what I think life really refers to life what we call life is any system that is really good at scaling these fundamentally non-zero competencies of chemistry and physics into much larger competencies... and I say I'm not trying to pin it on on a kind of a single Universal objective fact about the system I'm looking at it from the perspective of an observer who wants to relate to it somehow to and to understand it”

A ball rolling downhill has a non-zero level of competency on the intelligence spectrum (it minimizes energy without being pushed) but very minimal agency—you can only interact with it by changing the landscape or pushing it, not by changing its beliefs or internal state.

“if you tell me it's a bowling ball on a hill what I hear is it has it's it's not zero competency it actually has by by by a virtue of least action principles it actually has a little bit of Competency to minimize things like that with you know kinetic and the total energy and all that kind of stuff so so but but I also know that if I want to change the way the ball rolls I don't have any I'm going to gain zero benefit by thinking about um internal uh State you know internal how does the ball feel about this landscape that's not going to be a useful way for me to go and the only tools I have are to modify the actual landscape”

The reason that mathematics, geometry, topology, and navigation-based concepts are useful for understanding biological systems is not necessarily because they are 'the true' fundamental frameworks, but because they are the tools we have that allow us to make progress; other frameworks (Don Hoffman's spacetime-free approaches) may ultimately be better, but we should judge frameworks pragmatically by what progress they enable, not by armchair philosophy.

“I I lean very heavily on concepts of navigation and and and you know spaces and geometry and topology and all of this stuff I not because I know that this is the best frame this is these are the tools that we have and so we can push the ball forward a little further than than prior work which didn't use those tools that's fine but that's not to say that these are this is the best way to think about this entirely possible you know I mean I mean Don Hoffman and and other physicists tell us that space-time is doomed”

Different fields have developed different tools and frameworks for understanding systems (behavioral science, cognitive science, thermodynamics, etc.), and the question is which tools are useful for which systems, not which tools are 'objectively correct'.

“taking uh tools from cognitive and Behavioral Sciences from from the kinds of things that Carl and Chris are doing that is is is super rich it drives all kinds of new research agendas so so that's why I stand behind that statement”

The 'thousand brains' theory (Jeff Hawkins) is interesting but should be expanded: the entire body, not just the brain, builds models of the world using reference frames, and the body is continuously constructing multiple representations in various problem spaces (morphological, physiological, transcriptional, etc.).

“I think it's I think it's very interesting and important and I think that everything is great except instead of the brain you can put the entire body right so so I think I think the whole body's doing this there are there are not that many things that your brain is doing that the whole body doesn't that other parts of the body aren't doing in various spaces”

The 'tame' acronym stands for 'Technological Approach to Mind Everywhere' and reflects the commitment to empirical testing and operational protocols rather than armchair philosophy.

“but it's a very that's why tame stands for technological approach to mind everywhere the technological aspect is it's it's we're not arguing philosophy here we are trying to get to very uh specific and testable ways in which other other beings observers and they could be scientists like us they could be um Biologicals they could be parasites they could be con specifics it could be the entire evolutionary process all of those are observers uh are able to optimally interact with that system”

We are not yet at the point where we have fine-grained control over the bioelectric interface in systems like planaria; control is currently coarse enough to produce stochastic rather than deterministic outcomes, explaining why bioelectrically-perturbed flatworms produce a probability distribution of head shapes rather than a specific predetermined form.

“we don't have you know the the technology for um uh controlling that bioelectrical interface is still in its infancy we have we have a fairly limited control over the richness of that bioelectric interface and so what that means is that in the planaria case we can't really choose specifically which head shape it'll get”