New Words


Preface

Originally, this essay was meant as an addendum to previous work. But I had more to say on the topic than I thought, and thus I have separated the two. That said, this is barely an essay, and certainly contains no explicit thesis; rather, it is a collection of various terms and connotations that have crept into my thinking. It would not be wrong to interpret it term by term, without any sort of throughline. It is, however, also the case that the terms I define very literally embody the ways I’ve been refining common ontologies. As a collective, then, their shared assumptions betray the nature of the ontological exploration that birthed them; one could reasonably claim that the definitions in isolation do convey an implicit thesis. I intend this not as reading instructions, but rather as an observation on definitions that this list provides occasion to mention. After all, I intend to make the implicit thesis explicit before long.

A Few New Terms

Though the following terms are primarily intended to be used epistemologically, many of them have useful real-world analogues. The two are notably related; in general, ontology dictates behavior (and often the opposite is true as well), and thus most of the following are both literal and theoretical terms. In fact, often things will act as both literal and metaphorical demonstrations of terms — I will provide examples later. For now, however, I will treat them as purely epistemological; the literal interpretations (which I claim follow fairly easily) will be left to the reader.

The framework I put forward is a mathematical one; as such, it lends itself to map-territory errors. Take care to avoid this mental slip — the known properties of these objects align with their mathematical equivalents, but this does not mean their unknown properties do as well: the nature of the mathematical objects often does not inform that of the theoretical ones.

It has occurred to me in writing these that they are perhaps too mathematical; often these abstract notions are simple intuition pumps, but once rigorized, they become unwieldy. I must confess that these definitions are, at times, self-indulgent. But I do claim that sufficient mathematical rigor forces the reader to interpret these notions from a different perspective, which has value outside whatever use the terms themselves may or may not bring.

Value Function:

A value function is, broadly speaking, what generates an ontology (more on this later). It takes (almost) anything — a potential action, a past event, an idea — and evaluates it. It exists both culturally and individually. It need not be (and rarely is) self-consistent (except in the sense that, as a black-box function, it is well defined). Indeed, on an individual scale, it is perhaps best thought of as the weighted sum of the value functions of mental subagents. Value functions are, except in the black-box sense, somewhat recursively defined; a cultural value function can similarly be thought of as some nuanced version of the weighted sum of individual value functions; similarly, the value function of a subagent can be broken down further. The realistic range probably lies from global to neuronal, with both extremes lacking practicality. But I use this term almost exclusively at either an individual or subagent scale, for which the definition does not need much rigor; its intuitive meaning is fairly obvious.

Ontology:

An ontology is a collection of beliefs through which the world is parsed; equivalently, it is a way of organizing the world. It is thus quite nuanced to say a value function generates an ontology. I explore this idea in a footnote,1 and I find it interesting, but there is not much clarity to be found — suffice to say that the two are deeply interconnected. Approximately, an ontology is a collection of beliefs, while a value function is the application of those beliefs. To unify them, one might claim that the value function is precisely the facet of an ontology that evaluates and facilitates the changing of said ontology. I personally think they are best left separate — both notions are useful.

Local Maximum:

A local maximum, somewhat like the analytical term, is a state where an arbitrarily small perturbation in any direction lowers the output of the value function. Intuitively, let our value function be $F_v:\mathbb{R}^n\rightarrow \mathbb{R}$, and let the local maximum be the $n$-dimensional vector $v$. Then $\exists \delta>0$ s.t. $F_v(v) > F_v(v_1)$ $\forall v_1: 0<|v-v_1|\leq\delta$.

For a mental state $s$, let $P_s = \{x : \text{gradient flow from } x \text{ converges to } s\}$ be the basin of attraction of $s$. Let $\alpha_s$ refer to the infimum of the set $\{|u| : s+u \notin P_s\}$. Note that $\alpha_s$ is nonzero only if $s$ is a local maximum. For some local maximum $l$, let us define $\alpha_l$ as the escape distance of $l$. Call the boundary of the basin of attraction, $ES_l = \partial P_l$, the escape set of $l$; note that $\alpha_l$ is then the distance from $l$ to $ES_l$. Finally, define the infimum of $\{\frac{F_v(l)-F_v(p)}{|l-p|}: p\in ES_l\}$ as the escape slope of $l$.

Obviously a local maximum with an escape slope of $5$ doesn’t mean anything — I haven’t given any numbers. But I still find these notions useful; it’s clear what I mean when I say something has a gentle or harsh escape slope.

The choice to work with local maximums as opposed to local minimums is intentional, but somewhat arbitrary. Maximums work very nicely with the idea of a value function, and impart the core idea of getting stuck in a set of beliefs. Minimums, however, capture the natural inclination to move towards these local extrema. I don’t think gravity is actually a helpful intuition in epistemology, but it certainly applies to real world models at times — homelessness, for example, is a local minimum, and gravity has some meaning in that case. Suffice to say that for this exploration maximums made the most sense, but minimums are symmetrically defined, and similarly useful.

Stable States

A stable point is a mental state where the net force exerted is zero. Belief-space can be decently modeled with rudimentary kinematics — we have clear notions of position, velocity, acceleration, force, inertia (mental states are point masses, thus $m=I$; I have chosen inertia, as it is a much more intuitive term), static friction, kinetic friction, work, etc. It is worth noting a separation between internal and external forces, but they behave similarly from a purely mathematical perspective.

A stable state, then, is a non-empty, closed, connected region $S$ s.t. $\varphi_t(S) \subseteq S$ $\forall t \geq 0$ (with $\varphi$ being the net-force flow), and every $p \in S$ is recurrent: its trajectory returns arbitrarily near $p$ at arbitrarily late times. Additionally, if you care, let $S$ be Lyapunov stable: $\forall \text{ open }U\supseteq S$, $\exists \text{ open }V\supseteq S$ s.t. $\varphi_t(V) \subseteq U$ $\forall t \geq 0$

Note the difference between net-force flow and value-function gradient — the value function gradient exerts a force, but depends on how much an individual trusts and places import on their own beliefs. Though many local maximums are stable states, and vice versa, they are by no means the same. Intuitively, being at a local maximum means your beliefs are self-justifying, self maintaining — it implies a set of self-consistent axioms. A stable state, however, indicates a stagnation in changing of beliefs, regardless of self consistency; often the desire to correct beliefs is balanced by an external force imposing an oppositional belief system. Actually, this applies to the notion of being stuck, but you can have very large stable states which don’t deserve to be called stagnation — an example is arguably the oscillation of American politics. But the definition is intuitively useful regardless.

Stable states also imply the notion of escape velocity (and relatedly, escape energy). This is less well defined in the case of a stable point, which relies heavily on the notion of static friction instead; for this, escape force applies instead. Regardless, the same intuition underlies all of these terms. Do note that escape velocity (and other such terms) isn’t constant — it varies over the space. Thus it is useful to have notions of minimum, maximum, average, etc. escape velocity (as well as what escape speed looks like along a given direction). Moreover, all of these notions can be very meaningfully restricted to cycles and substates within our stable state (cycles are always stable states unless you require Lyapunov stability, in which case it depends).2

As a final interesting notion, I do intuitively feel like a stable state should also need to be topologically equivalent to a point, but I think this requires me to say far more about the structure of our space than I have thus far (or am able to).

Attractors/Influence Nodes

An attractor is, approximately, a point with gravity. There are mathematical justifications, but after a brief attempt at rigorization, I have instead decided to leave it to intuition. Generally, an attractor exhibits force based on distance, with closer points experiencing a greater effect. It is one of three similar objects: attractors, repulsors, and pipelines. While attractors exert a force towards themselves, and repulsors exert one away, pipelines have orientation; they attract or repel based on relative position, much like an attractor-repulsor dipole. The notion of direction proves useful even for attractors and repulsors — intuitively, the canonical example of attractors, advertising, is almost always targeted towards an audience with a specific starting set of beliefs. Thus we give attractors orientation (and thus orientation alignment falloff), such that they act on a cone, exhibiting the strongest effect along its center.

Note that it need not be the case that the sphere of influence of an attractor is centered at the destination of attraction (e.g. Jane Street advertising isn’t mainly targeted towards people who are already quants).3 It doesn’t seem to particularly matter if we let the sphere of influence be away from the attractor, or if we instead let the force be directed somewhere other than the attractor. The force exerted on the space is the same either way — the notion of the attractor as an object is decently arbitrary. When you think about stacking multiple, they take on different meanings — multiple attractors at the same point directing force elsewhere might represent a reading group with varied opinions (depending on the scope with which you apply this framework), while disparate attractors directed to the same point seems like the best model for Jane Street advertising. It doesn’t really matter; both of these notions will be deprecated in a moment. I have left this section solely to force you to think about the underlying intuition a bit.

Let us at once strengthen and simplify our definition. Forgetting the idea of pipelines and repulsors, think of an attractor for a moment as a more abstract notion with a direction, and any continuous function that gives force from non-zero distance and alignment (in the force-cone sense). Let an influence node then be the collection of an arbitrary number of these at a single point (they should ideally converge). This is a very powerful notion, in the sense that it allows three influence nodes to model almost anything (one can do anything radial, and two can do anything except on collinear points). This can be useful — precision is good. But they’re as unwieldy as they are useful. Thus along with this very strong definition, I’m inclined to return to simpler definitions for attractors, repulsors, and pipelines. Forget de-localization, forget conical falloff, forget direction (except in the case of pipelines). It is much better to have a weaker notion of an attractor; strong notions are only useful in precise spaces, and these terms are meant to be general. For example, in a social context, a person may be an attractor, a repulsor, or even (in some cases) a pipeline. But calling them an influence node is probably more trouble than it’s worth.

I have taken you on a circuitous path, and I have changed definitions on you multiple times. To summarize, we have defined nodes of influence, a powerful general way to think of force-exerting objects. But these are unwieldy. Thus we have three intuitively special cases: attractors, repulsors, and pipelines, which attract, repel, and push through, each based on distance (except pipelines).

Event Horizons:

The idea of an event horizon does not lend itself trivially to our mathematical model, but it is useful nonetheless. It can perhaps best be thought of as a shell topologically equivalent to an $S^{n-1}\subseteq \mathbb{R}^n$, each point of which is a pipeline oriented inwards (which is approximately equivalent to an attractor inside the event horizon; this is usually also present). Event horizons also often weaken forces directing points out of their interior. Though the model is nebulous, the intuition is clear: an event horizon is a mental state (or location, framework, etc.) that traps its victims. I like to describe it as removing one’s ability to see outside of it (this feature is technically more appropriate for the event horizons of white holes, but this is a little meaningless).

Black Holes:

An epistemological black hole is a powerful stable-state attractor with an event horizon. The stable-state usually has a very high escape energy, and, relatedly, (to borrow a term from the local maximum lexicon) a substantial escape slope. It also usually (but not always) permeates the space, that is, the attractor at the center is either exceptionally strong, or, more commonly, has a very slow falloff. A classic example is the alt-right — along with tendrils throughout thought-space that pipeline you into its center, its absolutist ontology makes it very difficult to escape, along with dimming any ideas from outside its purview.

A danger arises in combining definitions: simpler definitions are nonetheless quite useful, and thus I have been inclined to give them detail, but in doing so, synthetical definitions become overspecified. It is therefore worth taking a moment to discuss the most notable properties of black holes. The specific architecture of their event horizons, attractors, and stable-states are unimportant (though occasionally interesting) — more notable are the collective properties: exhibiting a massive attractive force across a large area of thoughtspace, pulling its victims in, and warping their ontologies to trap them inside. Indeed, the one crucial aspect of a black hole is its inescapability — its extreme escape energy.

Singularities:

A singularity is a black hole which covers a finite segment of each line in $\mathbb{R}^n$ except one (which it contains an infinite length of). Call this line the ontological asymptote of our singularity. A singularity also applies a force to its entire area parallel to the direction of its asymptote, as well as another orthogonal ring of force attracting points towards the asymptote (note that an asymptote only extends to infinity in one direction — the parallel force is in the infinite direction).4

Intuitively, this should be thought of as a black hole with the added property that as it immerses its victims, it accelerates them away from the rest of the space, until any external ontology is unreachably far away. This models a single concept overwhelming the ontology until everything is placed in relation to this core idea — the classic example is AI safety. A victim might come upon the ideas on forums, then very educated people explain very reasonable sounding EV math to them, and the magnitude of the issue overwhelms any smaller considerations. I’m not saying this is bad (in this essay), rather, it is a notable pattern worth modeling.

Examples:

As I write, I worry that in each definition I put more thought into math than philosophy, and thus the Overton window of my subagents shifts; by the end, then, I have lost sight of the initial purpose of this endeavor in pursuit of ever more precise models. I do hope I have mitigated this somewhat — as the author, I’m too close to the piece to tell. Regardless, it thus seems worthwhile to end by running through a few examples of each definition at various scopes.

The epistemological version of a local maximum is a set of self-reinforcing beliefs — a trivial example is utilitarianism based on axiomatic assumptions. Alternatively, consider an individual who believes self criticism in any capacity is unhealthy, and avoids it. The escape distance isn’t particularly far in this case, but it does have a notably steep slope — the value function this state induces strongly penalizes any deviation. Note that while the value function determines the location of local maximums, the location in thought space also itself determines the value function, and so a point is a local maximum precisely if it is a local maximum on the function the ontology it represents induces. This interpretation allows for a complication of the space into one with built in value forces, as both the value function and its weight are determined by location. But this gets a bit into the ontology/value-function distinction mess. More importantly, it restricts the allegorical value of the notions which build off it to specifically epistemology, which I would rather not do. Indeed, this notion is also quite useful in practical contexts: consider the path of a soulless Bay Area math kid, consisting of grinding, camps, comps, applications, etc. — I claim this fits the definition quite well, and all the auxiliary terms follow somewhat trivially. Still, local maximums are perhaps the notion here with the worst non-epistemological applications, as they rely on the rather specific concept of value functions.

Stable states are far more general; epistemologically, whatever you currently believe is probably a stable state. Nihilism is a stable state. Burnout is a cyclic stable state (sometimes). More practically, college is a (very large) stable state, and most friendship dynamics come to rest in stable states.

Attractors are common in epistemology: consider rationality, fascist ideology, or externalizing blame during introspection. They are similarly commonplace in the real world; almost any end can be modeled as an attractor in some framework, and most means can be framed as pipelines. Advertisements, charisma, and prestige are all attractors in different senses. Repulsors can mostly be found in opposition to attractors (e.g. externalizing blame vs. accepting blame), but one notably interesting example is war propaganda (particularly on either side of Israel-Palestine, interpreted as repulsors from sympathy with the other camp). I will not give specific examples for influence nodes or pipelines; they can easily be used to add nuance to the examples already provided.

Event horizons are almost exclusively seen in the presence of black holes. The exceptions usually lie around things that aren’t particularly attractors, and are instead just really hard to escape — the framework of assumptions underlying cultural discourse has an event horizon (is an event horizon?), as does the set of ontological assumptions imported by a lexicon. These concepts are nebulous, to say the least. The clear examples come with black holes: political extremism, nihilism (again), and AI safety.

Indeed, AI safety is the only thing I currently think of as a singularity (though various things trivially become singularities under intentional constructions of the underlying space). One of my insidious ideas in creating these definitions is to build a lexicon around a set of axioms that define AI safety as problematic. This is perhaps a bad take. I intend to explore it in more depth in the future. For now, please criticize these definitions, add new ones, or otherwise progress them; in many ways they represent the current state of my thinking, and thus refining them would prove invaluable.


← All Inktober entries

1. It is true that a value function evaluates states, and so while the inclination towards certain states is often attributed to the ontology, it is, in fact, closer to taking the gradient of a value function. Yet a value function seems in one sense very clearly subservient to ontology, simply by virtue of being an element (or subset). There are, in fact, two kinds of ontologies, the difference being inclusion of the unconscious. If I was somehow capable of listing every belief I hold at this moment, it is hard to say this collection is not my ontology. Yet this set would realistically not perfectly describe how I parse the world, and thus it is similarly hard to say it is my ontology. There are similarly two kinds of value functions: a true value function, which includes desire, emotion, etc., and the self-perceived/conscious value function. The latter more or less generates the conscious ontology, in the sense that it captures the ways in which the subconscious shapes held beliefs, but it seems to me quite difficult to truly impose any sort of hierarchy on these objects.

It would perhaps be possible to nail these definitions down and clear up this confusion, but this seems unwise; I find the description I opened with to be the most useful way to think about ontology — to further specify it would necessarily decrease its scope. In discussing the relationship between value functions and ontologies, it seems best to view them as counterparts — an ontology is a collection of beliefs, while a value function is the application of those beliefs. The set theorist in me twitches upon reading this, but it will have to suffice; this is, after all, mathematical in intuition only. ↩
2. There are a lot of interesting mathematical properties here; how about a stable state where any subset fitting all the properties except Lyapunov stability also has Lyapunov stability? I spent some time thinking about what this would mean allegorically, and it’s not entirely vacuous, but it’s definitely more trouble than it’s worth. ↩
3. It turns out that the math I hand-waved allows force to be an arbitrary continuous function of (non-zero) distance. Though this is radial, the cone notion restricts it, which mostly gives us arbitrary point forces anywhere in the space. When I introduce the notion of influence nodes, which allow arbitrary composition of attractors and repulsors, we can replicate arbitrary force functions $f: \mathbb{R}^n \rightarrow \mathbb{R}^n$ from anywhere. So the notion of de-localized attraction is a bit useless in a purely mathematical sense. But these arbitrary force functions are somewhat allegorically meaningless, and thus would never come up; their flexibility is solely to improve their ability to model intuitive objects. From this perspective, de-localized attraction, though impractical, acts as a useful piece of intuition. ↩
4. Unfortunately, this asymptotic property means that the black hole inside the singularity can’t contain a stable-state, as it can’t have pseudo-cyclic force-flow trajectories. Worrying about this too much is a bit of a map-territory error, but it is still unfortunate. One potential solution is to wrap our space into a high-dimensional torus. This actually does have a nice intuitive backing (horseshoe theory), but I don’t personally like it — I think it defeats the point of the key notion of a singularity, which is that it’s hard to get back to the origin. ↩