Taonga in the Latent Space

A community-scale world model for Indigenous data sovereignty

John G. Stroh · My Digital Sovereignty Ltd. · 13 slides with presenter notes

Taonga in the Latent Space · June 2026 · MDSL

The bet against the trench

  • Nov 2025: Yann LeCun leaves Meta to found AMI Labs — seed round ~$1.03B at a $3.5B valuation, reported as Europe's largest
  • “The AI industry is completely LLM-pilled.” “Everybody is … digging the same trench.”
  • His charge: LLMs “can't truly reason or plan, because they lack a model of the world. They can't predict the consequences of their actions.”
  • “…they're not going to be built on LLMs.”
Taonga in the Latent Space · June 2026 · MDSL

What a world model is

  • Perception → world model → cost → actor (LeCun, 2022)
  • World model = “a kind of ‘simulator’ of the relevant aspects of [the] world” — predicts future states as a function of actions
  • Mode-1: reactive, no model consulted. Mode-2: “reasoning and planning … akin to model-predictive control”
  • A text-only LLM is stuck in Mode-1 — no model to run a consequence through
Taonga in the Latent Space · June 2026 · MDSL

JEPA — predict meaning, not pixels

  • Joint Embedding Predictive Architecture: predict future representations in latent space, not raw pixels or tokens
  • “It is not generative.” “We advocate against the use of generative architectures.”
  • Learn to ignore what isn't worth predicting; predict what matters
  • V-JEPA 2 (2025): a world model doing zero-shot robot planning — proof it ships
Taonga in the Latent Space · June 2026 · MDSL

The world worth modelling first is not the planet

  • Nothing in LeCun's definition requires the world to be pixels — only a state, actions, and dynamics
  • An institution is a discrete world whose state is already a record — the expensive half (learning state from sensors) doesn't arise
  • Community-scale is tractable; a planet-model that dethrones the frontier is not
  • Rome was not built in a day — lay one true stone
Taonga in the Latent Space · June 2026 · MDSL

The Village already governs that world

  • Sovereign records = the community's state, signed and replayable — the ground truth a world model needs
  • The gate already exists: checks “before action execution … the process cannot bypass the check” — allow / deny / escalate-to-human
  • The Situated Language Model made language cognition sovereign
  • Our Sovereign-Record Architecture paper (§15) names the gap: sovereignty over data — “not by itself … sovereignty over cognition”
Taonga in the Latent Space · June 2026 · MDSL

The missing middle: a community-state model

  • Predict what a proposed action would do to the community's state — consent, taonga, rights, mana, relationships
  • Maps onto LeCun: perception=encoder, world model=community-state model, cost=kaitiaki critic stack, actor=the agent
  • Turns an institutional agent from Mode-1 to Mode-2 — foresee the effect on the people, then act
  • Steered, not retrained; built inside tenant isolation (Adversary A7's refusal property)
Taonga in the Latent Space · June 2026 · MDSL

Start filling it with what the community sees

  • Past text, like adding documents to a situated model: a farmers' market clip, footage of the breached stop banks, a graduation
  • LeCun's V-JEPA video models are open and runnable today — they turn video into representations, not pixels
  • And live feeds — river gauges, tide and swell, rainfall, water-quality probes: perception that updates continuously and runs the next hours forward
  • A community needs only its OWN world — a small corpus it already holds — not a million hours of internet video
Taonga in the Latent Space · June 2026 · MDSL

Who is kaitiaki of the world model?

  • Plural Values named the gap: “no kaitiaki on the artifact”
  • LeCun's cost has an intrinsic part that is “immutable (not trainable)” — in the Village that is the constitutional floor (Te Tiriti, tapu, human-in-the-loop)
  • The trainable critic sits under the community's signed steering pack — “the benchmark is downstream of the kaitiaki”
  • The unit of deliberation is the latent diff: a legible, signed, revertible change to the model of your world
Taonga in the Latent Space · June 2026 · MDSL

Federation: how it takes root

  • Federate, Don't Align — taken one layer down, to world models
  • Each Village holds the model of its own world; none holds a model of everyone's
  • Cross-community questions federate bilaterally — scope-limited, revocable, provenance intact
  • H-JEPA turned on its side: a polycentric mesh, no centre to capture
Taonga in the Latent Space · June 2026 · MDSL

The one word, used once

  • I'll call where this leads the first step toward “sentient AI” — my coinage, not LeCun's
  • Not consciousness, not feeling — accountability: a machine that has a world and can be governed in it
  • Tractatus 12.6: “Agency cannot be simulated, only respected”
  • Model the world the community lives in — never the community's agency
Taonga in the Latent Space · June 2026 · MDSL

The limits, stated plainly

  • Not shipped — proposed. No community-state model runs today; no aspirational building, no aspirational writing
  • Closes governance sovereignty, not substrate sovereignty (open weights, foreign hardware remain)
  • Federation mechanism is live and tested; no inter-community network yet
  • Written from a limited, non-Māori standpoint; appropriate use in te ao Māori is for tangata whenua to determine
Taonga in the Latent Space · June 2026 · MDSL

A model can answer. An agent can act.

  • But only an agent that can imagine what its act will do to a community — and let the community decide — has earned the right to act at all
  • The world model is how it imagines. The kaitiaki is who decides.
  • The machine models the world; the people remain its authors
Taonga in the Latent Space — John G. Stroh, My Digital Sovereignty Ltd., June 2026. agenticgovernance.digital · CC BY 4.0
Taonga in the Latent Space · June 2026 · MDSL

LeCun says today's AI lacks a model of the world. This talk: the first world worth modelling is not the planet but the community — and the Village already holds the hard parts around the gap.

The field's most decorated figure stakes a company against the consensus. The hinge quote is verbatim (MIT Tech Review, Jan 2026).

Reasoning, for LeCun, is imagining what would happen and rejecting what violates your constraints, before you move. A world model makes Mode-2 possible.

The heterodox half of LeCun's program — and the part that matters for governance: a latent state you can structure is a latent state you can govern.

The pivot of the whole essay. The achievable first step into world-grounded AI is a model of one community's own world, held by that community.

Two of three pieces are built: the substrate that records state, and the gate that checks an action. What's missing is the thing to check against.

The build, named precisely. The gate finally has something to check against: not the category of the action, but its predicted effect.

The on-ramp: you can begin grounding the latent space today, with the community's own footage and live sensor feeds. Guardrails: every input is a sovereign record (consent, taonga, kaitiaki); it stays in Aotearoa under its kaitiaki; and it is grounding, not surveillance — no biometrics, never a way of watching the people.

Here the Village goes beyond LeCun: the immutable values are community-authored law, not hard-wired drives; the community is kaitiaki of the model of its own world.

Capability at world-model scale by composition, not concentration. The pragmatic pathway for the approach to take root — buildable now, by the actors the capacity race wrote off.

Sentience here is not an interior; it is an answerability. The line between modelling the world and modelling the people's agency is the constitutional floor, in code. Taiuru's shadow: worth taking seriously, not worth a throne.

Same discipline as the rest of the corpus: mark proposed vs shipped, name the dependency, don't claim endorsement. Drafted on foreign substrate — the dependency is the argument.

Land the anchor. The first place to build world-grounded AI is the community; the community must be kaitiaki of the model of its world; federated, it takes root without a centre.