Anywhere Intelligence Research

Built from first principles.

Before we built Haven, we spent 18 months designing its operational AI layer, AIOS (Anywhere Intelligence Operating System), around two key questions: what makes AI genuinely personal, and what makes it trustworthy. The answers required fundamentally new architecture. Not a local wrapper around an existing cloud model, but a different kind of system designed from the ground up for personal intelligence.

These papers are our contribution to that conversation. We share them to advance research on personal, trustworthy AI and to explain the architecture behind Haven.

Personal

AI that accumulates what things mean to you — your relationships, your context, your language — not statistical averages calibrated across millions of users. Every interaction, remembered. Every meaning, yours to authorize.

Trust

AI that can show you, at any moment, exactly how it reached an answer — what it consulted, what it ruled out, and why. No hallucination without a trace. No action without authorization. Trust earned through architecture, not marketing.

Our Papers

Personal Intelligence

Personal World Models: Toward Human-Centric Personal Intelligence Systems

Justin Holbrook, Swaroop Kallakuri, Joshua Holbrook, Lucas Burgess · Anywhere Intelligence · 2026

The dominant paradigm in AI optimizes for machine-side intelligence: scale, benchmark performance, agentic capability. This framing is correct for many objectives and incorrect for one — systems designed to live with a specific human over years. For such systems, the central missing architecture is not more intelligence, but a different kind of structure: one that encodes not how the world generally behaves, but what the world specifically means to this user, in their specific history, with their specific relationships.

Formally distinguishes the personal world model from user profiling, persistent LLM memory, and graph RAG — structural differences, not cosmetic ones.
Introduces three-layer knowledge separation: ontology primitives, common-sense priors, and personal meaning commitments — user-authorized, provenance-tagged, reversible.
Derives continuous engagement (no session boundaries) as a necessary architectural consequence of coherence constraints, not a product feature.
Introduces constitutional continuity: an alignment approach for systems that accompany a human life across years, during which values genuinely evolve.

Read paper Download PDF

Trust Architecture

The Intelligence Processing Unit: Architectural Explainability as a Foundation for Trust in AI Systems

Justin Holbrook, Swaroop Kallakuri, Joshua Holbrook, Lucas Burgess · Anywhere Intelligence · 2026

Trust between humans rests on a simple mechanism: the ability to ask “why did you do that?” and receive an answer that is coherent, inspectable, and open to challenge. Current AI systems built on large language models lack this architecture — they act first, only later permitting post-process scrutability. The IPU addresses the trust deficit at the architectural level.

Introduces the IPU as the atomic unit of AI execution: a bounded, append-only, immutable record from query to response — every subsystem consulted, every routing decision made.
No durable system effect — memory creation, knowledge modification, scheduled action — can occur without a complete, unfailed IPU trace.
Argues that process-based explainability (why did the system take this path?) is more fundamental than prediction-based explainability (why did the model predict this token?).
Presents the IPU as implemented in AIOS, the Anywhere Intelligence Operating System running on Haven.

Read paper Download PDF

Research in Preparation

Forthcoming

Intelligence Squared (I²)

Why compact models with stronger geometry and routing can match or outperform much larger systems in targeted workloads — and why this makes Haven possible without cloud-scale infrastructure.

Forthcoming

Geometric Superposition Intent Classifier (GSIC)

A lightweight, ultra-low latency intent classifier built on a single-pass multi-head MLP employing orthogonal vector layers to simultaneously enable query classification and routing.

Additional research.

The papers above argue from first principles. The work below bears on the same questions: what people actually ask AI to do, how much of that a small model can handle, what it costs to run intelligence at all, and the benefits of buildling thinking traces into core AI architecture.

What people ask AI to do

How People Use ChatGPT
Chatterji, Cunningham, Deming, et al. · NBER Working Paper 34255 · 2025
A privacy-preserving analysis of approximately 1.1 million conversations. Messages unrelated to work grew from 53% to more than 70% of all usage, and practical guidance, seeking information and writing collectively account for nearly 80% of conversations. Programming and self-expression are small shares.

What small models are enough for

Small Language Models are the Future of Agentic AI
Belcak, Heinrich, Diao, et al. · NVIDIA Research · 2025
Models small enough to run on consumer hardware are already capable enough for most of what an agentic system actually does, at ten to thirty times lower cost — reserving a larger model for the open-ended conversation that genuinely needs one.

Whether local inference is viable

Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
Saad-Falcon, Narayan, Akengin, et al. · Stanford · 2025
Measured against a million real queries, local models of twenty billion active parameters or fewer answered 88.7% of single-turn chat and reasoning queries correctly. The share a local machine can service rose from 23.2% to 71.3% in two years.
Apple Intelligence Foundation Language Models
Apple · Tech Report · 2025
A three-billion-parameter model shipping in production on Apple silicon, using two-bit quantization-aware training and KV-cache sharing to fit the hardware people already own.

The problem of non-determinism

The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
Shojaee, Mirzadeh, Alizadeh, Horton, Bengio and Farajtabar · Apple Machine Learning Research · June 2025
Frontier reasoning models asked to execute algorithms they were given — Tower of Hanoi among them — collapsed past a complexity threshold and did not use the explicit procedure. Haven routes those queries to a deterministic solver: the model understands the request; the algorithm moves the discs.

What intelligence costs to run

LLM inference prices have fallen rapidly but unequally across tasks
Epoch AI · 2025
Measured against benchmark milestones, the price of a fixed level of capability falls between nine and nine hundred times per year depending on which milestone you pick, with a median of fifty. The steepest declines are the most recent ones.
The Price of Progress: Price Performance and the Future of AI
Gundlach, Lynch, Mertens and Thompson · MIT FutureTech · 2025
A larger price dataset puts the same decline more conservatively, at five to ten times a year — and finds the cost of running a frontier model rising three to eighteen times a year, because the last increments of benchmark performance are bought with ever more inference.

Stay in the loop.

Check back to see our research and product progress.

See how Haven works Back to home