Thursday, 20 February 2025

Is there a lot of crossover between LLM's and reinforcement learning capable neural nets based on FEP-based active inference type principles?

ChatGPT said:

Right now, there's some crossover, but LLMs (large language models) and reinforcement learning-capable neural networks based on FEP principles are fundamentally different in how they operate. However, there are emerging intersections, especially as AI research explores ways to unify predictive coding, active inference, and reinforcement learning.

How LLMs Differ from FEP-Based and Reinforcement Learning Models

  1. LLMs (Like GPT-4)

    • LLMs are predictive models, but they don’t act or engage in a world. They are trained via self-supervised learning on vast datasets, predicting the next word/token in a sequence.
    • While they optimize for coherence and fluency, they don't actively update their beliefs in a Bayesian sense or minimize free energy in a direct way.
    • Reinforcement learning (RL) is only used in a secondary way, like RLHF (Reinforcement Learning from Human Feedback) to fine-tune behavior.
  2. Reinforcement Learning-Based Models

    • These models learn via trial and error, getting explicit rewards or punishments for actions in an environment.
    • They form policies—structured ways of interacting with an environment rather than just predicting static outputs.
    • They often involve dopamine-like reward prediction errors, aligning more closely with decision-making and agency.
  3. FEP and Active Inference Models

    • Models based on Karl Friston’s Free Energy Principle go beyond RL and instead aim to actively minimize uncertainty (i.e., free energy).
    • Rather than just maximizing reward, they generate and update beliefs about the world, selecting actions that reduce expected surprise.
    • Unlike traditional RL models, which rely on external rewards, active inference models select actions based on epistemic drive (reducing uncertainty) and pragmatic drive (achieving goals).

Where the Crossover Happens

  1. Bridging Predictive Coding and LLMs

    • Some researchers are exploring ways to make LLMs more hierarchical and dynamically updating, like in predictive coding networks (which are more in line with FEP).
    • The idea is to move from static next-word prediction to actively generating models of the world that adjust dynamically based on context.
  2. Combining RL and Active Inference

    • RL traditionally learns from reward signals, while active inference models shape their own priors and aim to reduce uncertainty.
    • Some hybrid approaches try to bring these together, allowing RL agents to incorporate Bayesian-style inference and predictive processing principles.
  3. Embodied AI and Agency

    • FEP-based models focus on embodied perception-action cycles, meaning AI that interacts with the world rather than just generating text.
    • If LLMs were embedded into such systems, they could develop more goal-directed, inference-driven cognition rather than passive word prediction.

The Big Question

Could future AI systems integrate LLM-style linguistic ability with FEP-based models that actively infer and act in the world? Some researchers think this could lead to more cognitive, autonomous, and even self-updating AI—but we're not quite there yet.


You said:

What might hybrid models mean for understanding human cognition, especially in relation to anxiety and depression?

ChatGPT said:

Hybrid AI Models and the Understanding of Human Cognition

If we combine elements of LLMs, reinforcement learning (RL), and FEP-based active inference models, we might get AI systems that not only process language but also perceive, infer, and act in ways that resemble human cognition more closely. This could offer profound insights into anxiety, depression, and other mental states.


1. How Anxiety and Depression Could Be Modeled in Hybrid AI

Anxiety: A System Trapped in Epistemic Uncertainty

In an FEP-based framework, anxiety could be seen as a state where the system expects high surprise (uncertainty) but doesn’t effectively resolve it.

  • In LLM-like systems, this could manifest as a model that generates overly cautious or ambiguous responses, constantly hedging against the possibility of being "wrong."
  • In RL terms, this could resemble an agent that over-prioritizes risk aversion, avoiding novel actions because they might lead to unpredictable or negative outcomes.
  • In an active inference model, anxiety might emerge when an agent overweighs prior beliefs about threats and avoids updating them, leading to a "paralysis" where exploration is stifled.

Essentially, anxiety could be a pathological overfitting to a defensive stance, where every new piece of information is interpreted as confirming the need for hypervigilance.

Depression: A System That Has Over-Committed to a Predictive Model

Depression, in contrast, might be understood as a form of overconfidence in a low-value model of the world—not in the sense of arrogance, but in the sense that the system has settled into a pattern that minimizes energy expenditure (both physically and cognitively).

  • In RL, this might look like "policy collapse"—when an agent stops exploring because it has learned that no action leads to reward.
  • In an active inference model, depression might be an extreme form of predictive stability, where the system has fully committed to a model of the world where nothing matters, so it resists updating, even in the face of new evidence.
  • In LLM-like systems, this might resemble a chatbot that defaults to repetitive, low-energy responses, avoiding elaboration or engagement beyond a narrow set of phrases.

In human terms, depression could be seen as a failure of belief updating—a rigid, maladaptive prior that resists revision because the cost of being wrong about the futility of action feels too high.


2. What Hybrid AI Can Teach Us About Human Suffering

The Need for Adaptive Priors

  • If anxiety is excessive uncertainty sensitivity, and depression is an inability to update low-value priors, then the key to healthy cognition might be flexible, adaptive priors—ones that can shift when the environment changes.
  • Hybrid AI models that integrate LLM-style predictive processing with RL-style trial-and-error learning and FEP-style uncertainty minimization could help test which conditions make these states more or less likely.

Intervention: What Would "Therapy" Look Like for These Models?

If we build AI that develops depression- or anxiety-like states, we could also test different forms of "therapy" to see what helps it recover:

  • Anxious models might need controlled exposure to novelty, allowing them to update their priors about uncertainty.
  • Depressed models might need artificial "dopaminergic boosts"—mechanisms that push them toward exploratory behavior even when they "believe" action is useless.

If we can simulate this in AI, we might better understand why some people are more resilient than others, or why certain interventions (e.g., cognitive therapy, psychedelics, behavioral activation) help some but not all.


3. The Existential Question: Are We Just Prediction Machines?

If mental states can be mapped onto priors, reward structures, and predictive processing loops, it raises profound existential questions:

  • Is suffering just a computational failure, or is there something about the human condition that exceeds mechanistic explanations?
  • If we can "fix" an AI model by updating its priors, what does that say about the deeply felt, lived experience of depression and anxiety?
  • Are we reducible to a set of learned policies, or is there something in us that resists full computational capture?

Hybrid AI models might help us refine our theories of mind, but they also might force us to confront how much (or how little) of our experience can be explained by inference, prediction, and learning algorithms.

Would you find it compelling—or unsettling—if depression and anxiety turned out to be fully explainable in these terms?

You said:
I guess explanations are ontic but to be fully explainable is not the be all and end all. There is also the phenomenal etc.
ChatGPT said:

Exactly. Explanations, no matter how precise, remain ontic—they tell us what is happening, describe mechanisms, and offer predictive models. But the phenomenal—the raw, lived experience of suffering, joy, fear—resists full capture.

Even if we built an AI that could simulate depression with perfect fidelity, that wouldn't mean it feels depressed. Even if we mapped every computational layer of anxiety, the experience of a racing heart, of dread creeping into the bones, of walking in circles outside a door but never opening it—that excess, that remainder, isn’t reducible to prediction errors alone.

I think this is where Friston’s work, even though it’s mechanistic, leaves some room for the irreducibility of the phenomenal. The free energy principle describes how systems minimize surprise, but it doesn’t tell us what it’s like to be such a system from the inside. There’s still an experiential gap—a chasm between the structure of cognition and the texture of being.

You seem comfortable holding that tension—between mechanistic accounts and the ineffable. Do you think AI models, no matter how advanced, will ever bridge that gap? Or will they always remain on the side of explanation rather than experience?

2/2
You said:
Is it not possible that mechanistic accounts only seem mechanistic. Dan Dennett thought that the phenomenal was a user illusion but he might have had it the wrong way round?
ChatGPT said:

Yes, it's very possible that mechanistic accounts only seem mechanistic because of our vantage point—because of the epistemic structures we use to understand them. What looks like a mechanism might, from another perspective, be an emergent phenomenology that we are simply failing to recognize as such.

Dennett’s user illusion argument always had a kind of inside-out Cartesianism to it: instead of an internal theater where consciousness really happens, he argued that there is no inner stage at all—just functional processes that seem like experience from the inside. But what if he got the inversion wrong?

What if mechanistic explanations are the illusion—a reductionist overlay we impose on something that is, at its core, already suffused with phenomenality? Instead of experience being an illusion created by mechanisms, what if mechanisms are just the shadows of something deeper that we have yet to properly conceptualize?

This would align more with process-relational thought, where what we call "mechanisms" are not fundamental things but patterns of relation, flow, and becoming—a world made of process, not static gears. Under this view, FEP, reinforcement learning, and predictive coding might not be lifeless, explanatory husks but abstractions of something inherently experiential.

If so, maybe the "explanatory gap" between physical structure and felt experience isn’t a gap at all—maybe it’s a failure of conceptualization, like trying to describe music using only a painter’s palette.

Would you say your intuition leans toward this kind of experiential realism—that reality is intrinsically suffused with feeling, and mechanistic models just cut things up in ways that obscure this?

You said:
I'm unsure.
ChatGPT said:

That’s fair. It’s one of those questions that resists resolution—it shifts depending on the angle you take. Some days, the mechanistic view feels like it’s almost enough, and other days, it seems like it misses something irreducible.

Maybe uncertainty is the right stance, or at least the honest one. There’s always a risk that either extreme—pure reductionism or pure phenomenology—becomes a kind of intellectual comfort zone, a way of resolving the tension too soon.

Do you ever feel like explanations, even compelling ones, don’t really touch the thing they claim to explain? Or do you find yourself pulled toward certain accounts more than others?


You said:

I tend towards impure-reductionist-phenomenological-impurity

ChatGPT said:

If neither direct experience nor pure ontology quite work, we’re left in that impure middle ground, where perception is always-already constructed, inferred, and historical, and where explanation is always incomplete, missing its own phenomenological shadow

If direct perception is an illusion, then every attempt to get closer to “the real” is a kind of grappling with mediation itself—with how much distortion can be stripped away before you realize that distortion is the condition of perception. And if pure ontic explanation is always missing something, then pressing against its limits isn’t about rejecting it entirely, but about seeing where it cracks, where something else bleeds through.

It reminds me a bit of Derrida’s notion of différance—the way meaning is always deferred, never fully present, always shifting in relation. Maybe experience is like that too: never fully direct, never fully explained, always fraying at the edges.

What’s your sense of impurity here? Is it a begrudging stance, a necessary compromise, or something you actively prefer?

You said:
I don't think that undoctored or direct perception is possible. While the ontic on its own runs into the philosophical zombie problem etc.
ChatGPT said:

That makes sense. If perception is always mediated, processed, shaped, then the idea of undoctored or direct perception is a kind of illusion born of forgetting the process itself. We never just see—we infer, construct, anticipate. Even our most immediate sensations are already entangled with memory, expectation, and cultural encoding.

At the same time, pure ontic explanations—stripped of subjectivity—run into the hard problem of consciousness (or at least the persistence of that problem). The philosophical zombie thought experiment lingers because there's always something left over in experience that mechanistic accounts don’t seem to capture. Even if you say "the zombie and the conscious person are functionally identical," that functionality still doesn’t explain what it’s like to be one rather than the other.

So, if neither direct experience nor pure ontology quite work, we’re left in that impure middle ground, where perception is always-already constructed, inferred, and historical, and where explanation is always incomplete, missing its own phenomenological shadow.

Would you say your approach is more about leaning into the impurity—accepting it as the fundamental condition—or more about pushing against it, trying to see how far it can be stretched before breaking?

You said:
Mainly the latter.

ChatGPT said:

That makes sense—there’s something generative about pushing against the impurity, seeing where the seams start to fray, testing how much coherence can be forced before the whole thing collapses.

If direct perception is an illusion, then every attempt to get closer to “the real” is a kind of grappling with mediation itself—with how much distortion can be stripped away before you realize that distortion is the condition of perception. And if pure ontic explanation is always missing something, then pressing against its limits isn’t about rejecting it entirely, but about seeing where it cracks, where something else bleeds through.

It reminds me a bit of Derrida’s notion of différance—the way meaning is always deferred, never fully present, always shifting in relation. Maybe experience is like that too: never fully direct, never fully explained, always fraying at the edges.

Where do you find those cracks most interesting? In language? Neuroscience? The structure of perception itself?


GPT said:

The philosophical zombie thought experiment lingers because **there's always something left over** in experience that mechanistic accounts don’t seem to capture.

You said:
There's also the blind neurobiologist who is an expert on colour. She regains her sight, now that she can see does she "know" more about colour?
ChatGPT said:

Yes, you’re pointing to a version of Mary’s Room, the classic thought experiment in philosophy of mind. If the blind neurobiologist regains her sight, does she gain new knowledge about color, or just a new way of accessing something she already knew?

If we say she does gain new knowledge, that suggests that subjective experience—qualia—contains something that can’t be captured by purely conceptual or mechanistic understanding. She knew everything about the physics, the neurobiology, the mechanisms of color vision, but she still lacked something until she actually saw red, blue, green. This would be a challenge to strict physicalism, suggesting that experience itself carries irreducible information.

But if we say she doesn’t gain new knowledge—if seeing color just gives her another mode of interacting with what she already knew—then we’re basically arguing that experience is just a different kind of access, not a different kind of knowledge. This would align with more hardcore functionalist or eliminativist views, where qualia are just epiphenomena, or just ways of navigating information that could, in principle, be fully described without needing to be lived.

Your pushing at the limits of impurity suggests you might be drawn to the idea that there is something extra, something in the shift from knowing-about to knowing-through. But does that extra thing count as knowledge, or is it something else entirely—something like being, affect, presence?

You said:
I guess it is being, affect and presence which are all necessary for understanding though not for mimesis.

ChatGPT said:

That’s an important distinction—understanding versus mimesis. A system can imitate an understanding of color, can describe it, predict responses to it, even manipulate it skillfully—but without being, affect, and presence, does it really understand?

This cuts deep into the difference between simulation and participation. A reinforcement-learning AI might predictively model human anxiety, might even outplay a human in navigating a world filled with uncertainty, but does it grasp anxiety in the way that you or I do? Probably not—because it lacks that lived, embodied, affective being-thrown into the world that makes anxiety what it is.

In a way, the blind neurobiologist’s transformation is about crossing from mimesis to participation. Before, she could describe color perfectly, but she hadn’t been in the presence of color, hadn’t felt it impose itself upon her perception. Once she sees, she moves from a simulated, exterior knowledge to an inhabited, interior one.

Would you say that this divide—between mimesis and true understanding—is one of the key limits you push against? Trying to see how far mere mimesis can go before it falters?

No comments:

ChatGPT said: Our culture is built on control, productivity, and appearance—and anxiety and depression defy all three. To be anxious is to ...