(Last revised: June 2026. See changelog at the bottom.)

5.1 Post summary / Table of contents

Part of the “Intro to brain-like-AGI safety” post series.

In the previous post, I discussed the “short-term predictor”—a circuit which, thanks to a learning algorithm, emits an output that predicts a ground-truth supervisory signal arriving a short time (e.g. a fraction of a second) later.

In this post, I propose that we can take a short-term predictor, wrap it up into a closed loop involving a bit more circuitry, and we wind up with a new module that I call a “long-term predictor”. Just like it sounds, this circuit can make longer-term predictions, e.g. “I’m likely to eat in the next 10 minutes”. This circuit winds up being a kind of hybrid between Temporal Difference (TD) learning and supervised learning.

I will argue that there is a large collection of side-by-side long-term predictors in the brain, each comprising a short-term predictor in the Learning Subsystem (more specifically, in the extended striatum—putamen, nucleus accumbens, lateral septum, part of the amygdala, etc.) that loops down to the Steering Subsystem (hypothalamus and brainstem) and then back up via a dopamine neuron. These long-term predictors make predictions about biologically-relevant inputs and outputs—for example, one long-term predictor might predict whether I’ll feel pain in my arm, another whether I’ll get goosebumps, another whether I’ll release cortisol, another whether I’ll eat, and so on. Moreover, one of these long-term predictors is essentially a value function for reinforcement learning.

All these predictors will play a major role in motivation—a story which I will finish in the next post.

Table of contents:

• §5.2 starts with a toy model of a “long-term predictor” circuit, consisting of the “short-term predictor” of the previous post, plus some extra components, wrapped into a closed loop. Getting a good intuitive understanding of this model will be important going forward, and I will walk through how that model would behave under different circumstances.
• §5.3 applies this model to the special case of actor-critic reinforcement learning (RL), where we’re predicting whether a plan will have a good or bad outcome. This connects RL concepts like reward and value / critic, to everyday psych concepts like pleasure and motivation, in the context of the “long-term predictor” motif. At a high level, the connection is strong, but I use some concepts and assumptions that will be unorthodox and confusing for RL experts. I try to help by walking through the common stumbling blocks.
• §5.4 will relate long-term predictors to the neuroanatomy of the striatum and brainstem.
• §5.5 will offer five lines of evidence that lead me to believe this story: (1) It’s a sensible way to implement a biologically-useful capability; (2) It’s introspectively plausible; (3) It’s evolutionarily plausible; (4) It’s neuroscientifically plausible; (5) It’s compatible with the psychology literature on “Pavlovian conditioning”, “revaluation”, “incentive learning”, and more.

5.2 Toy model of a “long-term predictor” circuit

A “long-term predictor” is ultimately nothing more than a short-term predictor whose output signal helps determine its own supervisory signal. Here’s a toy model of what that can look like:

• The blue box is the short-term predictor of the previous post. It optimizes its output signal such that it approximates what the supervisor signal will be in 0.3 seconds (as an example).
• The purple box is a 2-way switch. The toggle on the switch is controlled by genetically-hardwired circuitry (gray oval), according to the following rules:
  1. By and large, the switch is in the bottom setting (“defer-to-predictor mode”). This setting is akin to the genetically-hardwired circuitry “trusting” that the short-term predictor’s output is sensible, and in particular producing the suggested amount of digestive enzymes.
  2. If the genetically-hardwired circuitry gets a signal that I’m eating something right now, and that I don’t have adequate digestive enzymes, it flips the switch to “override mode”. Regardless of what the short-term predictor says, it sends the signal to manufacture digestive enzymes.
  3. If the genetically-hardwired circuitry has been asking for digestive enzyme production for an extended period, and there’s still no food being eaten, then it again flips the switch to “override mode”. Regardless of what the short-term predictor says, it sends the signal to stop manufacturing digestive enzymes.

Note: You can assume that all the signals in the diagram can vary continuously across a range of values (as opposed to being discrete on/off signals), with the exception of the signal that toggles the 2-way switch.^[1] In the brain, smoothly-adjustable signals might be created by, for example, rate-coding—i.e., encoding information as the frequency with which a neuron is firing.

Next, let’s walk through the properties of this toy model.^[2]

5.2.1 Toy model walkthrough: preliminaries

Here’s a little summary table for this subsection, that I’ll explain in the text below.

5.2.1.1 What is the short-term predictor getting trained to do?

The short-term predictor is a learning algorithm. Given infinite time, we normally expect learning algorithms to settle into some steady-state configuration where they no longer update. We can think of this configuration as “what we are training it to do”. So, in this toy model, what are we training the short term predictor to do?

Well, first, if there’s an override with ground truth R, we are training the short-term predictor to be already outputting R, immediately before that override.

And second, while we’re in defer-to-predictor mode, we are training the short-term predictor to keep its output steady. After all, any change in its output would create a (transient) discrepancy between the short-term predictor’s output, versus its supervisor, thus causing an update.^[3]

(Clarification: I’m using the phrase “keep its output steady”, but that glosses over the key detail that we are training it to “keep its output steady” in a way that privileges one temporal direction over the other: the output now is treated as what the output should have been in the immediate past. Whereas in the other direction, if the short-term predictor was making a bad prediction in the past, but now it has new information that reveals the error, we are not training it to “keep its output steady” by sticking with the old bad prediction. That’s just not what the learning algorithm updates would do.)

So putting those two observations together, in this toy model, we are training the short-term predictor to always output what the next override will be, even if that override won’t happen for a very long time. (Or if there’s randomness that makes the next override unpredictable, we are training it to predict the expectation value of the next override.)

5.2.1.2 AI intuitions: supervised learning, and Temporal Difference (TD) learning

If you’re an AI reader, think of the overrides as like sporadic “injections of ground truth”, turning the setup into a kind of supervised learning.

And separately, think of defer-to-predictor mode as a version of Temporal Difference (TD) learning.

AI people might be wondering: In AI textbooks, TD learning is notoriously bad at long-term planning, because the correct behavior “walks backwards” one (fraction-of-a-second) timestep per attempt. And yet, humans will plan a vacation a month in advance, without having previously gone on (1 month) / (0.3 seconds) ≈ 10,000,000 vacations. Is that evidence that TD learning can’t be involved? No! The trick is, in the brain, the context data going into the short-term predictor already contains temporally-extended latent variables. For example, the mental concept “I’m getting lunch” is active in your mind 10 minutes in advance of eating, as you get on the bus to the restaurant, and is also active in your mind at the moment when you bite into the sandwich. So nothing has to propagate back one-timestep-at-a-time over the course of thousands of lunches; instead it takes as few as one iteration for the short-term predictor to learn that the “I’m getting lunch” latent variable predicts digestive enzymes.

(Note that the term “TD learning” is usually used in reinforcement learning (RL) in the context of going from a reward function to a value function. But TD learning is really a general-purpose algorithm that takes a quantity X and spits out a long-term expectation of X. We can use that algorithm for whatever we want, including for things like “digestive enzyme production” which are neither rewards nor values.^[4] That said, we will in fact apply these ideas to rewards and values, as a special case, when we get to §5.3 below.)

5.2.1.3 Neuro intuitions: innate reactions, and self-fulfilling prophecies

Meanwhile, neuro readers should think of override mode as innate reactions to innate triggers, often related to homeostasis, like salivating in response to salty food in the mouth, or physiological arousal in response to pain. As discussed in Post 3, these reactions generally happen entirely within the Steering Subsystem (brainstem, hypothalamus, etc.).

And then think of defer-to-predictor mode as “self-fulfilling prophecies”. For example, as you hold the cracker in your hand, you expect that you’re about to salivate, and that expectation turns into reality. Similarly, as you go to start a painful procedure, you expect that you’re about the have physiological arousal, and that expectation turns into reality. The Steering Subsystem has no direct evidence either way, so it “defers” to the guesses coming from the Learning Subsystem.

5.2.2 Toy model walkthrough: the case of static context

Next, let’s walk through what would happen in this toy model, in the specific case that the “context” is static for some extended period of time. For example, imagine a situation where some ancient worm-like creature is digging in the sandy ocean bed for many consecutive minutes. Plausibly, its sensory environment would stay pretty much constant as long as it keeps digging, as would its thoughts and plans (insofar as this ancient worm-like creature has “thoughts and plans” in the first place). Or if you want another example of (approximately) static context—this one involving a human rather than a worm—hang on until the next subsection.

In the static-context case, let’s first consider what happens when the switch is sitting in “defer-to-predictor mode”: Since the output is looping right back to the supervisor, there is no error in the supervised learning module. The predictions are correct. The synapses aren’t changing. Even if this situation is very common, it has no bearing on how the short-term predictor eventually winds up behaving.

The times that do matter for the eventual behavior of the short-term predictor are those rare times that we go into “override mode”. They inject ground truth into the system, and produce an error signal in the short-term predictor’s learning algorithm, changing its adjustable parameters (e.g. synapse strengths).

After enough life experience (a.k.a. “training” in ML terminology), the short-term predictor should have the property that the overrides balance out. There may still be occasional overrides that increase digestive-enzyme production, and there may still be occasional overrides that decrease digestive-enzyme production, but those two types of overrides should balance out to an average of zero. After all, if they didn’t balance out, the short-term predictor’s internal learning algorithm would gradually change its parameters so that they did balance out.

And that’s just what we want! We’ll wind up with appropriate digestive enzyme production at appropriate times, in a way that properly accounts for any information available in the context data—what the animal is doing right now, what it’s planning to do in the future, what its current sensory inputs are, etc.

5.2.2.1 David-Burns-style exposure therapy—a possible real-life example of the toy model with static context?

As it happens, I recently read David Burns’s book Feeling Great (my review). David Burns has a very interesting approach to exposure therapy—an approach that happens to serve as an excellent example of how my toy model works in the static-context situation!

Here’s the short version. (Warning: If you’re thinking of doing exposure therapy on yourself at home, at least read the whole book first!) Excerpt from the book:

For example, when I was in high school, I wanted to be on the stage crew of Brigadoon, a play my school was putting on, but it required overcoming my fear of heights since the stage crew had to climb ladders and work near the ceiling to adjust the lights and curtains. My drama teacher, Mr. Krishak, helped me overcome this fear with the very type of exposure techniques I’m talking about. He led me to the theater and put a tall ladder in the middle of the stage, where there was nothing nearby to grab or hold on to. He told me all I had to do was stand on the top of the ladder until my fear disappeared. He reassured me that he’d stand on the floor next to me and wait.

I began climbing the ladder, step by step, and became more and more frightened. When I got to the top, I was terrified. My eyes were almost 18 feet from the floor, since the ladder was 12 feet tall, and I was just over 6 feet tall. I told Mr. Krishak I was in a panic and asked what I should do. Was there something I should say, do, or think about to make my anxiety go away? He shook his head and told me to just stand there until I was cured.

I continued to stand there in terror for about ten more minutes. When I told Mr. Krishak I was still in a panic, he assured me that I was doing great and that I should just stand there a few more minutes until my anxiety went away. A few minutes later, my anxiety suddenly disappeared. I couldn’t believe it!

I told him, “Hey, Mr. Krishak, I’m cured now!”

He said, “Great, you can come on down from the ladder now, and you can be on the stage crew of Brigadoon!”

I had a blast working on the stage crew. I absolutely loved climbing ladders and adjusting the lights and curtains near the ceiling, and I couldn’t even remember why or how I’d been so afraid of heights.

This story seems to be beautifully consistent with my toy model here. David started the day in a state where his short-term-predictors output “extremely strong fear reactions” when he was up high. As long as David stayed up on the ladder, those fear-reaction short-term-predictors kept on getting the same context data, and therefore they kept on firing their outputs at full strength. And David just kept feeling terrified.

Then, after 15 boring-yet-terrifying minutes on the ladder, some innate circuit in David’s brainstem issued an override—as if to say, “C’mon, nothing is changing, nothing is happening, we can’t just keep burning all these calories all day. It’s time to calm down now.” The short-term-predictors continued sending the same outputs as before, but the brainstem exercised its veto power, and forcibly reset David’s cortisol, heart-rate, etc., back to baseline. This “override” state immediately created error signals in the relevant short-term-predictors in David’s amygdala! And the error signals, in turn, led to model updates! The short-term predictors were all edited, and from then on, David was no longer afraid of heights.

Here’s the diagram for this situation; make sure you can follow all the steps.

5.2.3 Toy model walkthrough, assuming changing context

The previous subsection assumed static context lines (constant sensory environment, constant behaviors, constant thoughts and plans, etc.). What happens if the context is not static?

If the context lines are changing, then it’s no longer true that learning happens only at “overrides”. If context changes in the absence of “overrides”, it will result in changing of the output, and the new output will be treated as ground truth for what the old output should have been. Again, this seems to be just what we want: if we learned something new and relevant in the last second, then our current expectation should be more accurate than our previous expectation, and thus we have a sound basis for updating our models.

5.3 The RL value function (a.k.a. critic, a.k.a. “valence guess”) as a special case of long-term prediction

Just as we need to produce digestive enzymes based on an expectation of future meals, we likewise need to make decisions based on an expectation of future good outcomes that would result from those decisions. Thus, long-term predictors are a key part of planning and reinforcement learning in the brain.

More specifically: If you’re a human, you need to decide whether to watch TV or go to the gym. If you’re some ancient worm-like creature, you need to “decide” whether to dig or to swim. Either way, this “decision” impacts energy balance, salt balance, probability of injury, probability of mating—you name it.

The only practical way for an algorithm to output good foresighted decisions is to put everything into a common currency: a scalar signal of overall “goodness”. RL practitioners call this signal “reward”, and it loosely^[5] corresponds to the everyday notion of “pleasure”. Once you have this signal, you can insert an algorithm that learns to predict it; RL practitioners call this prediction “value” or “critic”, and in everyday life it’s related to motivation. And exactly how do we get this prediction? With a long-term predictor, of course!

(…And then another algorithm needs to turn those predictions into good decisions—i.e., the “actor” part of actor-critic RL. But I’ll hold off on that until the next post.)

I’ll explain my proposal through an example:

5.3.1 Example: Walking upstairs to get a sweater

Suppose I’m downstairs and feeling cold. An idea pops into my head: “I’ll go put on my cozy sweater”. So I walk upstairs and put it on.

What just happened? Here’s a diagram:

This is the same kind of diagram as §5.2 above, but instead of predicting a digestive enzyme signal, the blue box is predicting a special signal I call valence (more on which in a minute).

Here the “override mode” is the pleasure from putting on the cozy sweater, which can be grounded in innate brainstem circuitry related to warmth, tactile feelings, and so on. At that point, when the sweater is on, the brainstem “knows” that something good has happened.

But the more interesting question is: what was happening during the thirty seconds that it took me to walk upstairs? I evidently had motivation to continue walking, or I would have stopped and turned around. But my brainstem hadn’t gotten any ground truth yet that there were good things happening. That’s where “defer-to-predictor mode” comes in! The brainstem, lacking strong evidence about what’s happening, sees a positive valence guess coming out of the striatum and says, in effect, “OK, sure, whatever, I’ll take your word for it.”

And the short-term predictor will (after some life experience) output a good guess for the next override valence, whenever that may occur. And since the cozy sweater will be a positive valence override, the valence guesses preceding it will likewise be positive. And therefore I will be motivated to get the sweater, including spending 30 seconds walking up the stairs.

(Recall from §5.2.1.2 that the short-term predictor can learn quite quickly, because of temporally-extended latent variables in the context data. In this case, the latent variables in question would be related to the explicit idea, in my head, that I’m going to put on a sweater.)

5.3.2 How exactly does this setup relate to textbook RL?

There’s a high-level similarity-in-spirit between this setup and the kind of conventional actor-critic RL that you’ll find in an AI textbook. But when we drill into details, the correspondence between the two is pretty convoluted. Indeed, if you have a strong background in the conventional RL literature, you’re probably at a disadvantage in trying to parse my proposals! Here are some of the common points of confusion:

5.3.2.1 “Valence”, and ground-truth-iness

A number of differences between my model versus textbook actor-critic RL are related to the particulars of the “valence” signal, and related RL reward.

• Valence is a bit like a textbook RL reward function in the sense that, in override mode, it can be “ground-truth-y”—e.g. if I stub my toe, that seems bad (negative valence).
• …But also, valence is a bit like a textbook RL value function in the sense that, in defer-to-predictor mode, it becomes “forward-looking”—in the example above, valence was positive while I was walking up the stairs, because of the expectation of a sweater.
• Even when valence is “ground-truth-y” in override mode, it’s still substantially less ground-truth-y than we’re used to from textbook RL. For example, RL setups often calculate rewards based on external things like Atari scores, whereas in biology “all rewards are internal, and the internal/external distinction is not a useful one” (Singh et al. 2009). In particular:
  • Valence (and pleasure) are generally a function of what you’re thinking about, and also what’s happening in your body, not (directly) what’s objectively happening in the world. For example, it’s totally possible to like golf when you think about it in a certain way, but also dislike golf when you think about it in a different way, or in a different mood, etc.
  • Even in override mode, valence (and pleasure) can be quite non-ground-truth-y, because the “genetically-hardwired circuitry” upstream of the override signal also has other inputs, and those other inputs can include other long-term-predictors that are currently in defer-to-predictor mode. For example, laughter is pleasant (positive valence override), and laughter in turn is caused by a certain combination of signals related to physiological arousal and calm (see A Theory of Laughter), and those signals can in turn be triggered by thoughts and memories. So if I’m alone in bed, I can make myself laugh by thinking of something funny, and then I’d be having a positive-valence override despite no “ground-truth-y” external trigger.

5.3.2.2 Miscellaneous circuit elaborations

The circuit above is a toy model, so here’s a few words about what it’s leaving out, at least in the important special case of valence.

First, the short-term predictor in the circuit needs learning rate modulation. For example, suppose I’m doing math homework outside. At time t=0.0 seconds, I’m thinking “what if I take the square root of both sides of the equation?”, and then at time t=0.2 seconds, I’m thinking “aah my cell phone alarm is going off, I have to leave for my appointment!”. Thinking about the appointment might be positive or negative valence, but either way, it doesn’t provide any evidence about whether it would have been a good idea to take the square root of both sides.

Normally, my brain would do the TD learning thing, by comparing the actual valence at t=0.2 to the valence guess at t=0.0, and doing a weight-update on the short-term predictor to make them more consistent. However, in this case, it would be better to just skip that update altogether, i.e. to transiently set the learning rate to zero.

The end-result is what I sometimes call “TD learning with sporadic power outages”. These “power outages” enable the value to shift in the absence of a reward; this feature seems necessary to explain the fact that people are able to make progress on a plan (but without consummating the plan), and then do something else for a while, and then go back to the plan later. Neuroscientifically, I believe this feature is implemented by the striatal cholinergic interneuron “pause” (cf. Zhang & Cragg 2017) that can be triggered by a (generalized) orienting reflex.

Other differences between my model versus textbook actor-critic RL relate to various ways that RL practitioners tweak the circuit. Actually, some of these might be used in the brain too.

One example is time-discounting. Another is: we could introduce an asymmetry in how we treat positive (overshoot) versus negative (undershoot) errors in the short-term predictor—see my discussion of “distributional learning” in a footnote of the previous post. Probably other things too.

Also, I put an all-or-nothing two-way switch in the toy model, but realistically, it’s probably more like a weighted average with dynamically-adjustable weights.

To keep things simple, I will be ignoring all these possibilities (including time-discounting) in the discussion here.

…But there’s one more very obvious difference that’s worth diving into:

5.3.2.3 Switch (value = expected next reward) vs summation (value = expected sum of future rewards)

My model amounts to a value function that approximates the next “override” reward, whenever it should arise. However, textbook RL does something different: the value function approximates the (exponentially-discounted) sum of all future rewards. This amounts to swapping out our purple-box 2-way switch with a 2-way summation:

(The correct answer could also be “something in between switch and summation”. Or it could even be “none of the above”.)

RL papers universally use the summation version—i.e., “value is the expected sum of future rewards”. What about biology? And which is actually better?

It doesn’t always matter! Consider AlphaGo. Like every RL paper today, AlphaGo was originally formulated in the summation paradigm. But it happens to have one and only one nonzero reward signal per game, namely +1 at the end of the game if it wins, or -1 if it loses. In that case, switch vs summation makes no difference. The only difference is one of terminology:

• In the summation case, we would say “each non-terminal move in the Go game has reward=0”.
• In the switch case, we would say “each non-terminal move in the Go game has a reward of (null)”.

(Do you see why?)

But in other cases, it does matter. So back to the question: should it be switch or summation?

I don’t have a strong opinion that it must be one or the other, from first principles. But I do think the RL value function (“valence guess”) calculation in brains is approximately the switch one, not the summation one. Again, I think it’s really “neither of the above” in all cases; just that it’s much closer to switch.

Why do I think that? Well here’s an example: sometimes I stub my toe and it hurts for 20 seconds; other times I stub my toe and it hurts for 40 seconds. But I don’t think of the latter as twice as bad as the former. In fact, even five minutes later, I wouldn’t remember which is which. (See the peak-end rule.) This is the kind of thing I would naturally expect from switch, but is an awkward fit for summation. It’s not strictly incompatible with summation; it just requires a more complicated, value-dependent reward function. (As a matter of fact, if we allow the reward function to depend on value, then switch and summation can imitate each other.)

You might be wondering: isn’t summation essential for long-term planning? My answer: yes it is essential … in textbook RL. But hey, so much the worse for textbook RL! I think brains can plan ahead just fine with the switch circuit. For example, suppose there’s a yummy jelly bean in front of me, but if I eat it, then a monster will bite my arm off tomorrow. I’m not gonna eat the jelly bean! Why not? Because I’m explicitly thinking ahead to tomorrow, and that makes the plan seem bad. See the discussion of temporally-extended latent variables in §5.2.1.2 above.

Anyway, in upcoming posts, I’ll be assuming switch, not summation. I don’t think it matters very much for the big picture. I definitely don’t think it’s part of the “secret sauce” of animal intelligence, or anything like that. But it does affect some of the detailed descriptions. This will be an additional source of confusion for people with a conventional RL background, on top of everything else mentioned above.

5.3.3 Teaser for my “Valence” series (2023)

Again, “valence”—and here I mean “actual valence”, not “valence guess”—signals whether a thought is good (so I should keep thinking it, and think follow-up thoughts, and take associated actions), or bad (so I should start thinking about something else instead), all things considered.

My 5-post “Valence” series notes that, given the obvious central importance of this “valence” brain signal‚ it should illuminate many aspects of our everyday mental life. And indeed, it does! I argue in the series that valence is critical to understanding desires, values, vibes, brainstorming, motivated reasoning, the halo effect, social status, depression, mania, and more.

5.4 An array of long-term predictors involving the extended striatum & Steering Subsystem

Here’s the long-term-predictor circuit from above:

I can lump together the 2-way switch with the rest of the genetically-hardwired circuitry, and then rearrange the boxes a bit, and I get the following:

Now, obviously digestive enzymes are just one example. Let’s draw in some more examples, add some hypothesized neuroanatomy, and include other terminology. Here’s the result:

Excellent! We’re halfway to my big picture of decision-making and motivation. The rest of the picture—including the “actor” part of actor-critic reinforcement learning—will come in the next post, and will fill in the hole in the top-left side of that diagram. (The term “Steering Subsystem” comes from Post #3.)

Here’s one more diagram and caption for pedagogical purposes.

In the next two subsections, I will elaborate on the neuroanatomy which I’m hinting at in this diagram, and then I’ll talk about why you should believe me.

5.5 Five reasons I like this “array of long-term predictors” picture

5.5.1 It’s a sensible way to implement a biologically-useful capability

If you start producing digestive enzymes before eating, you’ll digest faster. If your heart starts racing before you see the lion, then your muscles will be primed and ready to go when you do see the lion. Etc.

So these kinds of predictors seem obviously useful.

Moreover, as discussed in the previous post (§4.5.2), the technique I’m proposing here (based on supervised learning) seems either superior to or complementary with other ways to meet these needs.

5.5.2 It’s introspectively plausible

For one thing, we do in fact start salivating before we eat the cracker, start feeling nervous before we see the lion, etc.

For another thing, consider the fact that all the actions I’m talking about in this post are involuntary: you cannot salivate on command, or dilate your pupils on command, etc., at least not in quite the same way that you can wiggle your thumb on command.

(More on voluntary actions in the next post, when I talk more about the cortex.)

I’m glossing over a bunch of complications here, but the involuntary nature of these things seems pleasingly consistent with the idea that they are being trained by their own dedicated supervisory signals, straight from the brainstem. They’re slaves to a different master, so to speak. We can kinda trick them into behaving in certain ways, but our control is limited and indirect (see §6.3.3 of the next post).

5.5.3 It’s evolutionary plausible

As discussed in §4.4 of the previous post, the simplest short-term predictor is extraordinarily simple, and the simplest long-term predictor is only a bit more complicated than that. And these very simple versions are already plausibly fitness-enhancing, even in very simple animals.

Moreover, as I discussed a while back (Dopamine-supervised learning in mammals & fruit flies (2021)), there is an array of little learning modules in the fruit fly, playing a seemingly-similar role to what I’m talking about here. Those modules also use dopamine as a supervisory signal, and there is some genomic evidence of a homology between those circuits and the mammalian telencephalon.

5.5.4 It’s neuroscientifically plausible

The details here are out-of-scope, and somewhat more complicated than I’m making it out to be, but for (a major subset of) Thought Assessors, I believe the correspondence between my toy models and real neuroanatomy is as follows:

• The “short-term predictor” is usually^[6] a set of medium spiny neurons (a.k.a. spiny projection neurons) of the extended striatum (caudate, nucleus accumbens, lateral septum, part of the amygdala, etc.)
  • Most of the extended striatum is dedicated to just a few Thought Assessors, namely the “valence guess” Thought Assessor (part of the brain’s “main” / “success-in-life” RL system), or the analogous Thought Assessors for any of the brain’s narrow-RL systems (those are off-topic for this series but see §1.5.6 of my Valence series (2023)).
  • A smaller portion of the extended striatum—specifically, parts of the amygdala, lateral septum, and nucleus accumbens shell—comprises all the dozens-to-hundreds of “visceral” Thought Assessors, related to things like cortisol, freezing, immune system activation, and so on.
• The “Output” signal from the short-term predictor (a.k.a. the “guess”, or “scorecard” entry) goes to one of various Steering Subsystem areas (e.g. GPi, GPe, SNr, SNc, various cell groups in the hypothalamus).
• The “Supervisor” signal going into the short-term predictor is one or more dopamine neurons coming back (directly or indirectly) from the Steering Subsystem.^[7]

Thus, for example, my claim is that if you zoom into (say) the central amygdala, you’ll find dozens of little genetically-determined subregions, each projecting to a specific set of genetically-determined Steering Subsystem targets, and with a genetically-determined narrowly-targeted supervisory signal coming back (directly or indirectly) from often the same Steering Subsystem target. Thus, for example, maybe thus-and-such little group of amygdala neurons is universally associated with getting goosebumps. But each individual brain would learn within its lifetime a different set of patterns that trigger this group of goosebumps-related neurons.

Is that hypothesis consistent with the experimental evidence? My impression so far, from everything I’ve read, is “yes”, although I admit I don’t have super-solid evidence either way. More specifically:

For the valence-guess Thought Assessor, the gory details are out-of-scope, but I think my story is elegantly compatible with textbook basal ganglia circuitry, particularly the famous motif of “cortico-basal ganglia-thalamo-cortical loops”.^[8]

For the “visceral” Thought Assessors: On general priors, there’s certainly no question that the genome is capable of wiring up dozens-to-hundreds of little cell groups in dozens-to-hundreds of specific and innate ways—this is ubiquitous in the Steering Subsystem. What about more specifically? I’ve found a few things. As one small-scale example, Lischinsky et al. (2023) studied two little embryonically-defined amygdala subpopulations, and finds that these two subpopulations wind up triggering in very different circumstances and connecting to very different downstream cells. As another example, at a larger scale, Heimer et al. (2008) says that the bundle of outputs of (what I’m calling) the “visceral” Thought Assessors are just a horrific mess, going every which way, as would be expected from hundreds of cell groups projecting to hundreds of different Steering Subsystem targets:

In the case of some macrosystem outputs, such as, for example, those from the accumbens core, the fibers course among these structures in tightly fasciculated fashion, giving off robust “bursts” of terminations only within or in the vicinity of particular structures. In other cases, good examples being the accumbens shell and extended amygdaloid outputs, the fibers are more loosely, if at all, fasciculated and have countless varicosities and short collaterals presumed to be sites of synaptic or parasynaptic transmission that involve not only defined structures but also points of indeterminate neural organization all along the course of the medial forebrain bundle. … As a consequence of their descent into this poorly defined organization, it is exceedingly difficult to conceive of precise, definitive mechanisms that outputs from the forebrain, including from the macrosystems, might engage to orchestrate purposeful, adaptive behavior from the extensive and sophisticated repertoire of autonomous, albeit restrictively programmed, hindbrain-spinal motor routines. But it is equivalently hard to dispel the notion that this is precisely what happens, and it seems advisable to presume that we are at present unable to perceive the relevant functional-anatomical relationships rather than to think they don’t exist.

As another piece of suggestive evidence, Lammel et al. (2014) mentions so-called “‘non-conventional’ VTA [dopamine] neurons” in “medial posterior VTA (PN and medial PBP)”. These seem to project to roughly the “visceral” Thought Assessor areas that I mentioned above, and it’s claimed that they have different firing patterns from other dopamine neurons. That seems intriguing, although I don’t have a more specific story than that right now. Relatedly, Verharen et al. (2020) discusses certain dopamine neurons which burst when aversive things happen, and four of the five regions these neurons seem to “supervise” are home to visceral Thought Assessors.^[9] This makes sense because if the mouse freezes in terror, then that’s a positive error signal for the freeze-in-terror visceral Thought Assessor, as well as the raise-your-cortisol-level visceral Thought Assessor, and so on.

5.5.5 It offers a nice way to make sense of a wide variety of animal psychology phenomena, including Pavlovian conditioning, devaluation, and more

See my post Incentive Learning vs Dead Sea Salt Experiment (2024) for details on that. See also the last section of that post for some contrasts between my model and a couple others in the psychology and neuro-AI literature.

5.6 Conclusion

Anyway, as usual I don’t pretend to have smoking-gun proof of my hypothesis (i.e. that the brain has an array of long-term predictors involving extended striatum-Steering Subsystem loops), and there are some bits that I’m still confused about. But considering the evidence in the previous subsection, and everything else I’ve read, I wind up feeling strongly that I’m broadly on the right track. I’m happy to discuss more in the comments. Otherwise, onward to the next post, where we will finally put everything together into a big picture of how I think motivation and decision-making work in the brain!

Changelog

July 2024: Since the initial version, I’ve made a bunch of changes.

The biggest changes are related to neuroanatomy. I was just really confused about lots of things in 2022, but I hope I’m converging towards correct answers! In particular:

I’ve gone through several iterations of the diagrams in this post, particularly including which neuroanatomy corresponds to which boxes.

There used to be a much longer discussion of neuroanatomy in §5.4, but much of that was incorrect, and the rest got moved into the newly-rewritten §5.5.4.

I originally had a discussion of dopamine diversity. I mostly deleted that, apart from a couple sentences, having been convinced that at least one of the pieces of evidence I discussed (Engelhart et al. 2019) was not related to “visceral Thought Assessors”, but rather had a different explanation. (Thanks Nathaniel Daw for the correction.) I also moved the discussion of distributional RL into a footnote of the previous post.

More minor things:

I added a note that all the diagrams in this post should be understood as the kinds of diagrams you see in FPGA design—i.e. at every instant, every block is running simultaneously, and every line (wire) is carrying a numerical value. It’s not like the kinds of diagrams you see in descriptions of serial program control flow, where we follow a path through the diagram, executing one thing at a time.

I’ve also changed the wording on the labels of the various diagrams a few times since the initial version. For example, where it used to say “Will lead to reward?” and “It was in fact leading to reward”, it now says “Valence guess” and “Actual valence”, respectively. The previous wording was OK, but I think the connotations are a bit better here, and also more consistent with my later Valence series.

I added a second concrete example story (involving a cozy sweater) as §5.3.2.

I replaced a brief discussion of the “dead sea salt experiment” with a link to my more recent post dedicated to that topic, in §5.5.5.

January 2026: I added a new §5.2.1 (“preliminaries”) to hopefully explain the toy model better. (The old §5.2.1 is now §5.2.2, and the old §5.2.2 is now §5.2.3.) Plus various other minor copyedits. Thanks Eli Tyre for helpful criticism that spurred the rewrite.

May 2026: I rearranged and rewrote §5.3 for clarity (along with related edits to §5.2.1.2). Among other things, the sweater example moved from §5.3.2 to §5.3.1, while the rest of the discussion moved from §5.3.1 to §5.3.2. Thanks Rif A. Saurous for helpful criticism that spurred the rewrite.

June 2026: Added a footnote referencing Lee et al. 2024. Added a discussion of learning rate modulation to §5.3.2.2. Moved “Teaser for my ‘Valence’ series” from §6.4.1 to §5.3.3 (and shortened it).

1. ^{^}

   To be clear, in reality, there probably isn’t a discrete all-or-nothing 2-way switch here. There could be a “weighted average” setting, for example. Remember, this whole discussion is just a pedagogical “toy model”; I expect that reality is more complicated in various respects.

2. ^{^}

   I note that I’m just running through this algorithm in my head; I haven’t simulated it. I’m optimistic that I didn’t majorly screw up, i.e. that everything I’m saying about the algorithm is qualitatively true, or at least can be qualitatively true with appropriate parameter settings and perhaps other minor tweaks. 

3. ^{^}

   To be clear, this transient discrepency has nothing to do with signal propagation delays. In this toy model, I declare that signal propagation is instantaneous. Rather, it’s from the 0.3 second (or whatever) time-offset internal to the short-term predictor.

4. ^{^}

   Examples in the literature where people use the terminology “TD learning” in a context unrelated to RL reward functions include: “TD networks”, the Successor Representations literature (example), or this paper, etc.

5. ^{^}

   The term “pleasure” doesn’t quite have the right connotations, but it’s the least-bad word I can think of. For example, if I allow myself to be tortured to death to save the lives of my countrymen, is that “pleasure”? I would say: “Actually, kinda yes! Specifically: When you think of yourself as a brave hero, that thought comes with a great deal of pleasure. On the other hand, when you think about the feeling of pain, that thought comes with a great deal of displeasure. At the end of the day, you might or might not endure the torture to keep the secret.

6. ^{^}

   Not always—certain parts of cortex layer 5 are also Thought-Assessor-ish. But meanwhile, most of the cortex is doing something different, as discussed in the next post.

7. ^{^}

   As in the previous post, when I say that “dopamine carries the supervisory signal”, I’m open to the possibility that dopamine is actually a closely-related signal like the error signal, or the negative error signal, or the negative supervisory signal. It really doesn’t matter for present purposes.

8. ^{^}

   Lee et al. 2024 is definitely not a perfect match to my beliefs on this topic, but it’s probably closer than anything else in the literature, and reading it was helpful for my thinking.

9. ^{^}

   The fifth area where that paper found dopamine neurons bursting under aversive circumstances, namely the tail of the striatum, has a different explanation I think—see here.