Addendum: Amusement, Awe, and Curiosity

Earlier some feelings were left out of the model…

Amusement, awe and curiosity didn’t fit the same way. If emotions are the feelings the controller outputs when there is variation off the setpoint, then these (and some others) don’t really work that way.

They do fit, but they fit in the estimator half of the model.

Why They Didn’t Fit…

Hunger, fear, shame. Each one reports a gap between where you are and where you’re supposed to be, and acting on it closes the gap. That’s what an error signal does.

Amusement does none of that. I don’t think a set point gets defended by a joke and nothing gets corrected when something is funny. Awe is the same. Curiosity is different as well because acting on curiosity often produces more curiosity and drive for exploration, which runs backwards from every loop in the framework (a hunger loop that made you hungrier would be a broken hunger loop).

What are the three of them reporting on, then? None of them is holding a value where it’s supposed to be. These are responsible for reinforcing learning and growth in different ways.

How Do They Fit?

The framework has three parts behind it. Part 1 is the controller side, where an emotion is the error signal against a set point. Part 2 is the estimator, working out what’s actually happening. Part 3 puts the two together and argues the estimate becomes the set point. Two functions with two gains, and the emotions we already had fit with the controller.

What would a feeling attached to the estimator even look like? It wouldn’t have a target, because the estimator is figuring out the target. It is determining a change (or not) to your model of the world.

That fits these three, exactly.

Curiosity Already Had a Formal Answer

Friston and colleagues published a formal answer in 2015. They call it epistemic value, the expected information gain from an action, and it sits inside the expected free energy alongside the goal-directed term (Friston et al., 2015).

The system weighs two things when it picks what to do. How much closer does this get me to what I want, and how much does this tell me about the world? That second one is curiosity.

That also answers what the Slime Mold Time Mold authors flagged, in their own words, as an enigma: “acting on your fear should make you less afraid, acting on your thirst should make you less thirsty, but acting on your curiosity often seems to make you more curious” (Slime Mold Time Mold, 2025b). They identified it as a probable non-emotion.

OK you ask, isn’t a signal that grows when you act on it just a broken loop? For hunger, yes. For an information-seeking loop it’s correct behavior, because reducing uncertainty in one place usually shows you what else there was to be uncertain about (I’ve watched this happen in a first session – you answer one thing and three more open up). The reward gets paid for the reduction. Let’s not delve into curiosity killed the cat examples but it is reminiscent to the famous Socrates quote: “The only true wisdom is in knowing you know nothing” explaining the drive for more knowledge.

Kidd and Hayden reviewed the psychology and neuroscience of curiosity and found the same information-gain account running through the empirical work (Kidd & Hayden, 2015).

Amusement Is Payment for Catching Your Own Error

Hurley et al. (2011) made this argument in Inside Jokes, and it fits the estimator reading with no adjustment. Your brain can’t predict everything so it takes shortcuts, and humor is what the system pays you for finding one of those shortcuts that was wrong, in your own live model.

Here’s one of mine, from a set I ran in 2020:

"I got a shot the other day. I’m traumatized by shots… The last time at the doctor the nurse was very patronizing. ‘Oh, come on now it’s just like a little bee sting.’

‘Oh well that’s the problem then, I’m allergic to bees… and their stings… and having sharp metal objects plunged into my body!’"

The room hears “just like a little bee sting” and takes it the way the nurse meant it, as a thing so small it’s the standard for small. They build that model themselves in about a second without noticing they’re doing it. “I’m allergic to bees” is where they catch it – the reassurance was the threat the whole time. The laugh is their payment for their own catch.

Nothing there corrects the world. The correction is to their map of it, which is estimator work.

Information you see coming isn’t funny, because there’s no estimator error left to catch. Explaining a joke kills it for the same reason. And a joke usually stops working the second time you hear it, because the error can only be made once. Think of amusement as when we have a big correction in the estimate and it needs to be a positive response because it allows us to get positively reinforced when we get this kind of new information shock which could result in a better understanding of the world. My comedy coach for my stand-up comedy career Tony Vicich used to describe the structure of a joke as

Setup (the opening)

Premise (building tension)

Punchline (includes a “twist” that shatters the premise)

Mark Twain talked about the secret to humor is truth and pain. But it can’t be truth you already know (the sun is in the sky at noon is completely boring) it has to be truth you just discovered through the twist.

…and that is reestablishing a new estimate.

Awe Is What a Big Revision Feels Like

Keltner and Haidt found two features in every clear case of awe: perceived vastness, and a need for accommodation, which they define as not being able to fit the experience into your existing mental structures (Keltner & Haidt, 2003).

Read that second one in this framework and it stops being an appraisal and relates to our estimator. Prediction error small enough to absorb gets absorbed and the model updates inside its current structure. Prediction error too big (vast=awe) for that forces the structure itself to change. What they’re calling accommodation is the model getting rebuilt or at least a large change in estimation.

Those two core features are what every clear case shares. What makes one awe experience feel different from another is a set of five peripheral appraisals that flavor it (Keltner & Haidt, 2003).

Threat introduces danger, and pushes the experience toward dread or terror. A natural disaster, or a menacing leader.

Beauty brings aesthetic pleasure and positive valence. A mountain range at first light, or a piece of art.

Exceptional ability is watching supreme skill, and it moves toward admiration.

Virtue responds to moral beauty or extreme human goodness, and it moves toward elevation.

The supernatural brings perceived divine or supernatural causality, and adds the uncanny or a kind of sacred ecstasy.

Keltner and Haidt list threat and beauty among the features that vary, and on this reading that stops being strange at all! A forced rebuild of your model of the world can arrive as wonder or as dread depending on what the new model says about where you stand in it, and the size of the revision is the SAME either way.

The Three of Them Together

The common factor in all of these is they adjust the way the estimator adjusts. This means how much of the model is getting changed.

Curiosity pulls toward information the model doesn’t have yet. Amusement rewards you for a burst of new information you didn’t expect, and awe is what a rebuild of the internal structure feels like while it’s happening.

The control half handles regulation and its feelings are error signals. The estimator half builds and maintains the model, and these three are the feelings attached to that.

What’s Still Open

All of this is theoretical. Our work here isn’t to prove anything but to suggest a new model.

Aesthetic pleasure is the next thing to explore. It may be a combination of the estimator, the way awe is. Or it may have a function of its own that feels good when the mind is captivated and is not focused on other controller-related things. If the mind is managing all kinds of input and estimation and control at once, then maybe (like stimming, or like cutting) the input provides relief from all of that. That’s for the next paper.

How You’d Prove It Wrong

How do you separate the two halves of the model for research? Change what somebody is uncertain about and amusement, awe and curiosity should move while fear and hunger sit still. Change how hard the system is set to act (gain, damping, whatever you can manipulate cleanly) and you should get the reverse.

If both sets move together under either handle, then there probably aren’t two jobs here at all, and the central claim of Part 3 has a problem a long way past this addendum.

In the End

This means there are different categories of feelings. The set described earlier (hunger, fear, sensory input) are from the controller and are designed to initiate behavior. This other category (including amusement, awe and curiosity) are from the estimator and are designed to update the model). The latter is still a drive that changes behavior (I want more amusement, I love awe and curiosity feeds itself) so we continue to update our internal models. In the end the framework didn’t need stretching to reach them.

What it explains is why a joke works at all, not which joke works on which person. Your priors aren’t mine, so the same twist catches you and misses me. I mean I think “The Naked Gun” is funny and I have no idea why my wife doesn’t! (I tell her it’s OK to be wrong)

References