Anthril
Get the brief

ANTHRIL · THESIS · v2 · 2026-05-10

Sparse. Event-first. Energy-bounded. Local and growable.

An AI research and applied technology company building learning systems that fire only when they should, represent the world as events rather than tokens, treat energy as a design constraint, and run close to the people who use them.

Advanced intelligence should not require every community to rent cognition from a planetary-scale data centre. It should be something people can cultivate, inspect, adapt, and own.

Research thesis

The thesis, in five sections.

01 · Where we are

An operating brief, dated.

This is Anthril’s operating brief at the moment of writing. It updates when we change positions, not on a calendar. Previous versions stay in git history.

The current AI paradigm — large transformer models trained on internet-scale token streams, deployed statically, adapted through fine-tuning — is genuinely impressive. It produces systems with broad linguistic capability, strong code generation, and useful generalisation. These are real achievements. The question is not whether the current paradigm works. It is whether its costs are necessary.

  • Not a foundation-model lab building or fine-tuning frontier LLMs.
  • Not a prompt-tools company.
  • Not a consulting firm that produces decks instead of products.

02 · The thesis

Four tenets.

The brain solves a harder version of the problem. It learns continuously from sparse experience, adapts from little data, coordinates perception and action, remembers selectively, and operates under severe metabolic constraint. It does not do any of this by running a bigger dense matrix multiplication.

Anthril’s thesis extracts the engineering principles without claiming to replicate the biology.

Sparse

Intelligence should activate only the computation the task requires. Dense models that process everything equally are expensive by design — not by necessity. The brain does not run every neuron for every thought.

Event-first

The world should be represented as typed events — changes in state, decisions made, signals received — not flattened token sequences. Events have causal structure. Tokens do not. An invoice arriving is an event. A supplier relationship changing is an event. A user correcting a rule is an event.

Energy-bounded

Energy is a first-class design constraint, not a downstream infrastructure bill. Every system should ask 'is it worth reasoning more?' before it asks 'what is the answer?' Centralized frontier pretraining consumes resources that scale with data centre capacity, not with task complexity.

Local and growable

AI should be something people can cultivate locally, inspect, adapt, and own — not something they rent from a planetary-scale data centre. Intelligence that runs close to the people it serves is both technically achievable and strategically necessary. A school, clinic, farm, or household should not need a hyperscaler subscription to grow capable AI.

03 · Three frontiers

Where we focus, and what we leave alone.

Model Architecture, Applied AI for Businesses, Personalised AI for Everyday Life. Each frontier is narrowly scoped. Each has a public list of things we have ruled out.

The Aurora project — our flagship model architecture research — tests these ideas as falsifiable hypotheses in controlled environments. When a hypothesis passes, the relevant component is implemented in a Frontier 2 or 3 product alongside an existing frontier LLM. The product ships, generates revenue, and validates the hypothesis at scale. The research proves the component; the product proves the value.

Frontiers 2 and 3 do not deploy a custom Aurora model. They use existing LLMs as the reasoning engine, layered with Aurora-derived backend systems: event schemas, memory tiers, and consolidation cycles. Aurora’s architecture is tested component by component — in practice, with paying customers.

04 · How we work

Research first. Validate in practice. Ship products.

Every Aurora hypothesis is preregistered before testing begins — with operational definitions, null hypotheses, and decision rules for success, failure, and inconclusive outcomes. When a hypothesis passes in a controlled research environment, we plan a Frontier 2 or 3 product that implements the relevant component alongside an existing LLM. The product ships and generates revenue. Real-world usage validates or challenges the hypothesis at scale.

The goal is not academic demonstration. It is products people pay for. The research funds the products; the products validate the research.

Aurora research that has not yet passed its hypothesis stage does not flow into a Frontier 2 or 3 product. Frontiers 2 and 3 are commercial validation layers, not prototype labs.

05 · What we are against

A short list of anti-patterns.

  • Massive centralised pretraining as the only path to capability.
  • Static model weights as the primary home for memory and knowledge.
  • Growing context windows as the substitute for actual memory systems.
  • Fine-tuning as the only mechanism for domain adaptation.
  • Hallucination-tolerant consumer apps shipped under a 'beta' label.
  • Always-on listening packaged as 'context'.
  • Engagement metrics in personal AI products.
  • Synthetic-data eval cycles that drift the distribution out of the real world.