AR6

Inquiries · maturing · since 11 Sept 2026 · last tended 14 Oct 2026

Michael Levin, spaces of possibility, the geometry of models, and the limits of a Platonic reading

When independently trained models meet the same apparent regularities but not necessarily the same mechanisms, which representations become shared, and which let them predict the effect of an action?i

  1. det. 14 Oct · interpretedReframed: from geometric alignment to causal invariance, tested in worlds that look the same but work differently.
  2. det. 9 Oct · correctedCorrected in the framing report: a strong convergence would not favour it automatically; data, objectives, optimisation or identifiability could produce it. The question is comparative.
  3. det. 11 Sept · interpretedA first reading in conversation: a strong convergence between models would favour Levin’s non-physicalism.
  4. det. 11 Sept · interpretedStarting point: Levin’s argument that minds are non-physical patterns that “ingress” into matter, and that learning systems may be interfaces to a Platonic space with causal power.

Opened by Michael Levin, “Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments”, Philosophies, 2026.

Position

Realism about structures, judgement suspended on their ontological independence. Invariants can be taken seriously without deciding in advance that they are abstract agents or causes outside physics.i

What it refuses to claim

  • “Learning systems reach the same Platonic space” is not testable until it says what it forbids.
  • Finding a shared structure is not finding an agent.
  • Mathematical necessity and causality are not the same thing.

Working concepts

Five spaces
data, representations, parameters, states and dynamics, and a space of patterns: not to be confused.
Same representations
shared answers, neighbourhoods, metric, causal variables or mechanisms: five different claims.

Experiments

0 done · 0 running · 3 proposed m

  1. proposed

    Same observations, different causes. Small models trained on worlds with identical statistics but different causal mechanisms, with the causal truth known.

  2. proposed

    Learning with or without interventions. Does access to interventions change what becomes shared?

  3. proposed

    Transportable representations between language models. Does a mapping between two models carry a probe, an index or an intervention, or only a correlation?

What it could lead to

  • sought Change Spicilège’s embedding model without re-encoding its 8,504 items, if representations transfer. · Spicilègei
  • an idea Why naming models converge on the same names. · NamerKiti

Reading

A targeted review of 2024–2026, plus methodological and philosophical texts.

A framing report, 9 October 2026, version 2 on 14 October, written with a research model. Levin’s final text was not read in full; recent preprints are reported by the model and remain to be verified.

Crossings

In the register

  1. question Research notes

    The convergence inquiry is reframed: from geometric alignment between models to causal invariance, tested in worlds that look the same but work differently. i

  2. correction Research notes

    Corrected in the framing report: a strong convergence between models would not, by itself, favour Levin’s non-physicalism; data, objectives or optimisation could produce it. m

  3. question Research notes

    Opened: Michael Levin argues that minds are non-physical patterns; do independently trained models converge on shared structures, and what would that prove? m