The bet on what comes after LLMs

ARIA today runs on the best LLMs. Tomorrow, it runs on world models.

LLMs are excellent at language-substrate reasoning — code, math, text — but they can't predict the consequences of their actions or plan by search. That's exactly what reliable autonomous agents need. This is how we get there.

Source thinking — Yann LeCun on Unsupervised Learning (Jacob Effron / Redpoint), captured 2026-06-21.

Roadmap phases

  1. Phase 1

    Now

    Shipped

    Shipped

    Nine LLM agents run an overnight autonomous cycle and deliver a verified morning digest.

    • Guardrailed by construction

      Every action is gated through guardrails. Tasks never “complete” without a verified_at timestamp — nothing ships unverified.

    • Where LLMs are strongest

      ARIA lives in the language substrate — code, math, analysis — scaling expert knowledge globally. The “top doctor everywhere” analogy.

  2. Phase 2

    2026

    On the roadmap

    Near-term · building now

    Move agents from prompted behavior to objective-driven systems that simulate consequences before acting.

    • Objective-Driven Agent Design

      Give each agent an explicit cost function and constraint set — a hardwired objective the system cannot violate by construction, not just a prompt that conditions behavior.

    • Consequence Simulation Layer

      A “poor man’s world model.” Before any irreversible action — send email, modify data, deploy code — run a lightweight what-happens-if dry-run. Agents without consequence prediction are intrinsically unsafe.

    • Search-Based Planning

      The Orchestrator searches over agent orderings for the sequence that produces the best verified outcome — planning by optimization, not next-token autoregression.

  3. Phase 3

    2027

    On the roadmap

    Medium-term · as world models mature

    Train neural world models of business dynamics — predict outcomes, generalize to new problems without retraining.

    • Business Dynamics World Model

      Revenue, churn, and CAC are complex systems no handful of equations can capture. Train a world model on customer data to predict: change campaign variable X — what happens to churn in 30 days?

    • Zero-Shot Task Generalization

      Tackle a new business problem without retraining, by planning through a learned model of how businesses work. A 17-year-old learns to drive in 20 hours — data efficiency over millions of demos.

    • Self-Supervised Pattern Discovery

      JEPA-style. Watch business data streams — revenue, tickets, ads — without labels, learning representations of healthy vs unhealthy states. Predict in representation space, not raw pixels.

  4. Phase 4

    2028+

    On the roadmap

    Long-term

    A shared world model of how SaaS businesses work, learned across the customer network — with zero data exposure.

    • Federated Business Intelligence

      The Tapestry model. Customers contribute parameter vectors and gradients, never raw data. A shared model gets smarter from every customer — your data never leaves your systems.

Why this matters

Most LLM wrappers have no post-LLM story.

LeCun expects the industry's “we need a paradigm change” realization to be obvious by early 2027. ARIA's roadmap says: we already know what comes after LLMs — and we're building toward it now, one verified phase at a time.

We don't just wrap today's models. We're building the system that outlasts them.