The Carnot Belief Engine

Is Twitter the intellectual equivalent of the combustion engine?

Recently

Venkatesh Rao

published a call for more work into the physics of intelligence — as contrasted from the philosophical and engineering discourses that pervade discourse today.

I am a dilettante, not a physicist, but I like dabbling in these concepts, and so today I want to sketch out an argument that humans have been creating “belief engines” that persuade humans of one thing or another for ages. In the same way that heat engines have dozens of engineering applications, but operate under specific physical laws, perhaps belief engines are constrained by some underlying dynamics that translate across types of intelligence.

Belief engines

One obvious example is the Twitter feed. The Twitter feed is propaganda, distilled. Its structure is designed to efficiently:

1. Present a claim.

2. Enrich it with context, such as author, links, views, and certification.

3. Introduce commentary that interprets the claim.

4. Facilitate a decision: like, comment, block.

Then, present a new claim, and repeat. Product people might call this an “engagement loop,” and it is. But there is plenty of psychological drama playing out in every one of these loops, and it leaves people burnt out and sad. And it’s hardly unique.

Take the old-school daily newspaper. It created a similar dynamic at a product level:

1. Present the headline news.

2. Enrich it with deeper discussion, on pages A2-A16.

3. Introduce commentary, on A17, that interprets the headlines.

4. Facilitate a discussion — did you read what happened?

Business analytics is no different. the entire purpose of business intelligence is to facilitate belief — that metrics are a very efficient way to convince large amounts of people of what is important, what is happening, and what should be focused on. But it operates not at the individual level, but on an organizational level.

What if there is something to this, at an intelligence level?

Conversion

Generically, what the belief engine does is make intelligence change its mind. Let’s just call the intelligence Hal — could be a psycho robot, could be Bryan Cranston, could be your dog.

Here’s what conversion looks like for Hal:

At 12:00, Hal believes the proposition: “Grapes are good.”

At 1:00, Hal believes the proposition: “Grapes are not good.”

This is a conversion. Let’s set aside a few of the possible concerns — can Hal spontaneously change his belief? Can Hal maintain two beliefs simultaneously? Which of the propositions is true? — and just ask:

What does it mean to believe?

Here’s a possible answer: belief is when there is little discrepancy between default expectation and the proposition. In other words, you can measure belief in terms of surprise. If I think dogs are great, and I hear someone say, “Dogs are horrible”, I get defensive (although this is a human reaction), because the new information disrupts my default view.

This notion of distance is important, since different propositions can be further apart than others. “Reality is a simulation,” is a fun and interesting proposition because it is so disruptive. Even if the distance calculation is not Euclidean — if, for example, it requires some measurement of the whole constellation of expectations, there ought to be some analogous state which can be converted to, and whose distance can be measured.

Hypothesis: The distance between belief states is measurable.

Next, we have to move the intelligence from point-in-latent-space A to point-in-latent-space ~A. In other words, we need to do psychic work to convert its default belief.

Let me propose two things:

1. The intelligence can transiently occupy point “~A” through concentration. This requires attention to an external belief state.

2. The intelligence must return to a default introverted state, because attentional resources are limited.

3. With repeated effort, it becomes easier for the intelligence to occupy point “~A”. In other words, the default state begins to incorporate point “~A” into its worldview.

So, the process of “expanding your mind” is, literally, a process of expansion to include new beliefs, followed by compression towards a default viewpoint.

Movement from belief state A to belief state ~A.

The theory, then, is that all learning — training, education, conditioning, etc. — all work on intelligence through this expansion-compression mechanism.

The Carnot Belief Cycle

In thermodynamics, the heat engine operates on such a compression-expansion cycle, leveraging heat (q) to manipulate the system and generate work on the surroundings. I won’t go into details here (see Wikipedia for a refresher), but I’ve drawn the chart below for comparison.

If we want to do work on the psychic system, we need to make it cycle between periods of very high concentration (i.e. attentive, external) and low concentration (i.e. diffuse, receptive) states. This process generates work that can be used, if directed, to move an individual from one belief to another.

Mapping of Carnot Cycle to Persuasion Cycle (Learning Cycle?)

Just as heat mediates the engine, information mediates the pscyhic expansion.

1. New information comes into the system. Attentional resources are initially diffuse but are allocated to process this information.

2. Attention concentrates on the information, incorporating additional information, especially from external sources. The intelligence transiently occupies this new space.

3. Attention is displaced, allowing information to leave the system. The net outcome is a contrast between the default and new information — a cost function of sorts. This may result in a feeling (“boo!”) or simply a knowledge difference, but it destablizes concentration.

4. Finally, the intelligence returns to a default, introverted and low-concentration state, where it is once again receptive to new information.

This is rough, and I can’t quite nail down all the different factors here.

The most interesting part of the cycle is from position 3 (attentive, external) to position 4 (attentive, internal). This is where information leaves the system, a counterintuitive fact.

But this is what happens in reality! When scrolling through Twitter, that most hellish of belief engines, you can spend an hour on the platform and explore hundreds of different belief states. When you walk away, what are you left with? Tiny intellectual souvenirs, accompanied by self-disgust.

I think there is an active forgetting that takes place when these engines push your brain around. Perhaps the intelligence seeks to displace the information by connecting it to some other proxy. Humans exchange the concentrated knowledge of a semantic space for the feeling it induces, which allows us to forget about the thing itself and store it in the body for more intuitive access later. The machine learning model simply updates its weights and discards the outputs entirely.

Dog training, machine learning, metrics reviews

One thing I like about this analogy is that the four-step process seems to fit the model of machine learning training, as well as other forms of training, quite well.

In many machine learning problems, we adjust the weights of the model (train) to make it approximate some cost function. This drives the weights to produce outputs that line up with our desired goals, and drive the output, for a given proposition, towards some space.

How do we do it with machine learning models?

1. Feed the model information

2. Let it propagate across the network

3. Calculate a cost for each input

4. Backpropagate this cost to update the weights

Then, repeat for many epochs.

Let’s just assume that the model is a form of intelligence. The model is a persistent, stable artifact (although it does not have to be deterministic). It’s not the brain, to use Rao’s analogy, just as the wing of an airplane is not a bird’s wing. But it learns like a brain (maybe).

Let’s also assume there are other intelligences. Two interesting ones come to mind: animals and social groups. Without going deep into their substrates, and how they differ, I think we can agree these groups have ways of learning, that are mediated by information.

Maybe this learning, too, follows an expansion-then-compression cycle?

Below are some examples mapping different learning processes (engines) onto this four-step framework.

This does / does not make sense

Here’s why this exercise interests me: we’re moving into a world where human intelligence is not the only intelligence on the block. That appears to be unnerving — but I think we’ve been here the whole time.

Human societies don’t have brains, but they do act in certain predictable ways. They have well-understood of propagating information and adjusting corporate actions. Businesses — to varying degrees — have always possessed a sort of artificial intelligence.

What’s been interesting to me about data work, and analytical work in particular, is fundamentally intelligence. In governments, this is actually what they call it: intelligence.

You can change the opinion of a business. This can manifest as both a strategic action — pivots — as well as internal moods — maybe a big deal fell through. Effective leaders driving their “vision” of the business can, literally, change the opinion of the business. This is belief conversion.

So if we want to understand the physics of intelligences, we ought to draw lines between what we know about organizations, humans, and perhaps, large language models.