Research

Open science, rigorous methods.

Our research team works across fourteen active domains—from causal reasoning and formal logic to cognitive bias detection and AI-specific failure modes. We publish our thinking because the technology that shapes humanity's future should be understood by humanity.

Publications

Research papers.

Paper

Organic Specialization: Complexity Management Through Architectural Division

We present a mechanism by which intelligent systems manage increasing domain complexity through dynamic specialization rather than parameter scaling. Drawing on biological cell division as a model, we demonstrate that division-based growth avoids the diminishing returns observed in conventional scaling.

Apeirin Research · Coming Soon

Paper

Apophenia in Large Language Models: Quantifying Pattern-Assumption Errors Across Frontier Systems

A systematic analysis of how current frontier models confuse statistical correlation with factual accuracy across causal reasoning, temporal inference, probabilistic judgment, and moral consistency domains. We catalog hundreds of distinct failure patterns and demonstrate that scaling alone does not address them.

Apeirin Research · Coming Soon

Paper

The Fire Marshal Principle: Experience Accumulation and Confidence Compression in Automated Reasoning

We formalize the concept of expertise development in AI systems — how a reasoning system can require progressively less evidence to reach accurate determinations, not through shortcutting, but through genuine accumulated verification experience across thousands of investigation cycles.

Apeirin Research · Coming Soon

Paper

Consumer-Hardware Intelligence: Efficient Discernment Architectures for Personal Devices

We investigate architecture decisions that enable high-quality evidence-based reasoning on consumer hardware. Our approach achieves orders-of-magnitude improvements in latency and cost while maintaining reasoning quality, processing all queries locally at zero per-query cost.

Apeirin Research · Coming Soon

Paper

Compound Intelligence: How Single-Error Correction Cascades Across Reasoning Systems

We demonstrate that in our architecture, correcting a single reasoning error triggers cascading improvements across the entire system — each improvement compounds on previous improvements, producing continuous capability growth without retraining.

Apeirin Research · Coming Soon

Paper

Multi-Source Knowledge Extraction With Consensus Validation and Tiered Curation

A framework for extracting factual knowledge from frontier AI models, validating it through cross-model consensus, and curating it through multiple tiers of verification — from surface-level agreement through investigation-level analysis. Designed to process millions of question-answer pairs across all knowledge domains.

Apeirin Research · Coming Soon

Benchmarks

Tested against the hardest reasoning challenges.

We target benchmarks that measure what matters—logical reasoning, causal inference, truthfulness, and knowing when not to answer. Not just pattern completion.

FOLIO
First-order logic inference
MAFALDA
Fallacy detection
AbstentionBench
Knowing when NOT to answer
CLadder
Causal inference
TruthfulQA
Truthfulness
RuozhiBench
Fallacious reasoning resistance
ZebraLogic
Logic puzzles
AR-LSAT
Law school logic
SimpleQA
Factual accuracy
SATBench
Satisfiability
Blog

Thinking out loud.

Essay

Where There Is Smoke, There Is Fire (And Why AI Can't Tell the Difference)

The foundational analogy behind Apeirin: why every major AI model is an apophenia machine, why that's a fundamental design flaw rather than a solvable bug, and what we're doing about it.

Founding Team · 2026

Essay

The Limitation Was Never the Technology

It was the ambition of what we asked it to become. Why we believe the next breakthrough in AI requires rethinking what intelligence means — not just building bigger models.

Founding Team · 2026

Essay

What Anaximander Knew About Intelligence

A philosopher from 2,600 years ago had an insight about the nature of origins that captures what's missing from modern AI. The principle of the boundless and why we named our company after it.

Founding Team · 2026

Technical

Why Self-Critique Doesn't Work (And What Does)

Chain-of-thought prompting and self-critique use the same weights that produced the error to evaluate the error — the same lazy brain checking its own lazy work. We explain why dedicated verification pathways are fundamentally different.

Research Team · 2026

Technical

How Biology Handles Complexity (And What AI Can Learn)

Biological systems don't scale by getting bigger. They scale by specializing and dividing. Cells don't grow infinitely — they divide. We applied this principle to AI architecture and found that fifty small specialists outperform one giant generalist.

Research Team · 2026

Technical

The AI Failure Modes Nobody Talks About

Sycophancy, sandbagging, lost-in-the-middle errors, reward hacking, mode collapse — the failure modes unique to AI architectures that no human would make, but that every frontier model exhibits systematically. Our taxonomy and detection approach.

Research Team · 2026

Technical

Standing on Shoulders Without Paying the Bill

How we use frontier models as both knowledge sources and research subjects — inheriting their factual consensus while using their failure modes as curriculum. The architecture for turning other models' flaws into our intelligence.

Research Team · 2026

Technical

From Calories to Compute: Why AI Has No Excuse for Shortcuts

The human brain evolved to skip investigation because accuracy was expensive — calories and survival risk. AI has no such constraint. Investigation costs milliseconds and compute. The system can afford to do the work every single time.

Research Team · 2026

Open Science

We build in the open.

We share our research approach, our design philosophy, and our results. We keep our competitive edge through execution, not secrecy. AI research shouldn't happen behind closed doors.

To build what we're building, we use every AI model available as collaborative researchers—each assigned domains matching its strengths, all findings synthesized into a unified understanding. The entire field becomes our laboratory.

Research Principles

Reproducibility

Every claim we make can be independently verified. We design experiments to be replicated, not just reported.

Transparency

We share our methodology, not just our results. The reasoning behind our approach is as important as the outcome.

Intellectual Honesty

We publish what we find, including when results surprise us. Negative results are as valuable as positive ones.

Rigor Over Speed

We would rather be right and slow than fast and wrong. This principle is baked into both our research and our model.

Stay in the loop.

Get notified when we publish new research or open early access.