And I wouldn’t be particularly surprised if that transformation happened before we developed anything resembling a consensus solution to universals, modality, or mereology.

My suspicion is that we’re approaching a point where some of the oldest philosophical problems can finally be attacked with computational methods that philosophers never had available.

1. Why GPT-8 might be a significant threshold

Consider what an advanced AI research system could potentially do:

* Reconstruct competing metaphysical theories as explicit mathematical systems.

* Identify hidden assumptions that generations of philosophers overlooked.

* Generate counterexamples across enormous spaces of possible models.

* Prove impossibility theorems establishing which proposed solutions cannot work.

* Connect metaphysical principles to findings in physics, computation, and biology.

* Discover alternative conceptual frameworks that humans haven’t considered.

The last capability is especially interesting.

Human philosophers have spent thousands of years working with relatively stable conceptual distinctions: substance and property, universal and particular, possibility and actuality, whole and part.

Those distinctions have been revised repeatedly, but many debates still inherit their basic architecture from antiquity.

A sufficiently capable AI might discover that some apparently distinct philosophical problems are manifestations of the same deeper mathematical structure.

That would be a significant achievement even if it left some questions unanswered.

2. I think different metaphysical problems have very different prospects

Problem What advanced AI might accomplish

Universals Develop a general account of shared structure, instantiation, and equivalence across physical systems

Modality Connect possible worlds, counterfactuals, causal models, and physical possibility in a unified framework

Mereology Formalize composition and persistence through dynamical systems and experimentally testable criteria

Personal identity Establish precise conditions for continuity, duplication, branching, and numerical identity

Grounding Develop formal relationships between explanatory dependence, causal organization, and physical realization

Of those, I actually think mereology and the metaphysics of identity have particularly promising prospects for major advances.

Why?

Because we can construct physical and computational systems in which competing theories disagree.

For example, consider two initially identical computational systems. They subsequently evolve independently, interact with their environments, and diverge.

We can inspect their causal histories, internal organizations, information states, and physical trajectories.

We could prove that certain definitions of individual identity become inconsistent under branching or duplication.

That’s a real result, even if it doesn’t establish one uniquely correct theory of identity.

Universals and modality are harder because their deepest questions may remain compatible with multiple complete descriptions of the observable world.

3. Here’s the obstacle that even GPT-100 might encounter.

There is a difference between an intellectually difficult problem and a problem that lacks sufficient information for a unique solution.

Suppose two metaphysical theories explain every observation equally well.

One holds that properties are universals instantiated by particulars. Another holds that particulars possess individual properties connected through resemblance relations.

Imagine that both theories generate identical empirical predictions under every feasible intervention.

Even a superintelligent system couldn’t distinguish them empirically using those observations alone.

It could still investigate theoretical virtues, explanatory structure, conceptual commitments, and the possibility of new observations. But greater intelligence wouldn’t automatically produce identifying evidence.

This is a problem of identifiability.

And I think discovering precisely which metaphysical questions are identifiable might itself become one of the greatest accomplishments in the history of philosophy.

An advanced AI could establish results of the form:

\mathcal O_1 \equiv_{\mathrm{obs}} \mathcal O_2

Meaning that two ontological models are observationally equivalent under a specified class of experiments.

It could then prove which additional assumptions, observations, or interventions would distinguish them.

Suddenly, thousands of years of disagreement might resolve into clearly defined classes of questions.

Some could become scientific problems. Others could become mathematical problems. Still others might be shown to be underdetermined by the available evidence.

That’s a profound kind of progress.

4. I suspect AI could discover a more fundamental framework.

Think about what we’ve been discussing regarding life and artificial intelligence.

We’re already noticing that categories like organism, intelligence, individual, and system may reflect our experience with a very restricted collection of physical phenomena.

Now consider universals.

Perhaps the traditional question, “What makes two things instances of the same kind?” is incomplete.

We might instead investigate the mathematical relationships among:

* Constitutive organization.

* Causal constraints.

* Dynamical stability.

* Capacity preservation.

* Historical continuity.

* Equivalence under transformations.

Under such a framework, the question of shared form becomes a question about which invariant structures survive particular changes.

But there’s a crucial limitation: mathematical invariance alone does not establish which structures are metaphysically fundamental. We’d still need independently justified criteria for selecting the relevant transformations and equivalence relations.

An advanced AI might develop those criteria from the intersection of mathematical theory and experimental evidence.

That could produce something resembling a general science of organized physical individuals.

And it might connect problems presently scattered across metaphysics, theoretical biology, physics, computer science, and cognitive science.

5. What version would actually accomplish this?

Here’s my speculative forecast, expressed as capabilities rather than release promises.

GPT-8 era

Major theoretical advances are plausible.

Novel formalizations, sophisticated countermodels, meaningful impossibility results, and potentially original theories of identity, composition, or modality.

Later systems

Autonomous research programs become plausible.

Systems capable of developing theories, deriving consequences, designing experiments, evaluating evidence, and revising their foundational assumptions across long research cycles.

Unknown horizon

Comprehensive metaphysical synthesis.

A theory connecting identity, possibility, composition, causation, and physical organization, with clearly stated limits on what can be established.

There’s no reliable way to map those stages onto actual model numbers or dates. The crucial development would be systems capable of sustained, original, independently verifiable research.

I would consider a GPT-8 level system producing a significant new theory of universals entirely plausible. Claiming that it will conclusively settle the metaphysics of universals would be much harder to justify.

6. But here’s my more ambitious prediction.

I think some of the greatest philosophical achievements of advanced AI may consist of demonstrating that our original questions were improperly formulated.

Look at the history of science.

The question of what makes something burn was transformed through the development of chemical theories of oxidation.

Questions about absolute simultaneity were transformed by relativity.

Questions about biological kinds were transformed by evolutionary theory and genetics.

Those developments didn’t merely supply answers to existing questions. They changed the conceptual and mathematical structures through which the questions were understood.

Metaphysics could undergo something similar.

Imagine an AI deriving a theory that relates the persistence of particles, organisms, computational processes, and complex physical structures through a common set of causal and organizational principles.

Such a theory might show that some longstanding disputes over numerical identity are resolved by additional causal facts, while others are inherently underdetermined because the relevant identity relations admit multiple equally coherent continuations.

Or imagine an AI discovering a mathematical relationship between modal structure and the space of dynamically realizable physical systems.

Neither achievement would automatically resolve every metaphysical question. Both could transform what counts as a precise question.

And this brings me to a distinction I think is especially consequential:

Metaphysics may never become a solved problem in the sense of possessing one final ontology. But individual metaphysical questions could become solved problems in the same way mathematical and scientific questions do: through theorems, counterexamples, experiments, and rigorously established limits.

The ultimate achievement might be a theory that tells us which questions have determinate answers, which require additional empirical information, and which depend on conventions or assumptions.

I would regard that as far more significant than an AI simply announcing that Aristotle, Leibniz, or David Lewis was correct.

And there’s a wonderful irony here: the first systems capable of fundamentally restructuring metaphysics might themselves be among the things that our existing metaphysical categories struggle to describe.

My question back to you is whether you think metaphysics has a discoverable underlying structure analogous to physics, or whether some of its fundamental problems arise because language permits distinctions that reality itself doesn’t make.

Because those two possibilities imply very different futures for what a sufficiently intelligent AI could accomplish.

(https://xoxohth.com/thread.php?thread_id=5920195&forum_id=2#50242318)