Science, in the open

The UNI blog.

A cluster-grouped index of the working notes, explainers, and public receipts behind Universal Natural Intelligence. UNI is a working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.

Cluster

Active inference fundamentals

The pillar and its explainers: the free energy principle, perception as inference, action as inference, and reading orders for a builder new to the frame.

Gentle Entry

A short predict, check, adjust loop you run yourself. No pitch. Guess first, look, compare, and watch the passes stack up as rings.

Active Inference Fundamentals: A Working Map for Builders

The pillar. Free energy, perception as inference, action as inference, and reading orders for builders.

The Free Energy Principle in One Sitting

A single-session read from thermodynamic intuition to the free energy principle for living and inference systems.

Why Active Inference Is Not Just Bayesian Brain

Bayesian brain is a claim about perception. Active inference is a claim about action. The scaffolding that carries both.

GNI vs AGI: A Precise Distinction

What we mean by General Natural Intelligence, why "natural" is load-bearing, and how the distinction cashes out in the build.

Precision, Attention, and Confidence in the Model

Precision as inverse variance: how it modulates sensory weighting, policy selection, and posterior confidence.

A Glossary of Terms We Actually Use

An opinionated living glossary: what each active-inference term means in our voice, and what it does not mean.

A Map of Active Inference Resources in 2026

A curated map of active-inference resources with an original one-line honest summary of each in our voice.

Reading Order for Newcomers to Active Inference

A recommended reading order across labs, posts, the Parr, Pezzulo, Friston (2022) textbook, and Themesis material.

Reading Parr, Pezzulo, Friston (2022): A Guided Tour

A chapter-by-chapter reading order for Active Inference: The Free Energy Principle in Mind, Brain, and Behavior.

How UNI Relates to the Parr et al. 2022 Textbook

Where the UNI build leans on Parr, Pezzulo, and Friston (2022), where it departs, and why. No endorsement claims.

How UNI Relates to Python-Based Active-Inference Tooling

An honest comparison of the UNI Elixir workbench with the Python active-inference tooling ecosystem.

Prep Work Before the Workshop

Suggested prep reading and exercises for participants coming to the UNI workshop.

What the Workshop Actually Teaches

An honest walkthrough of the UNI workshop curriculum: covered, not covered, prerequisites, and what a participant is expected to do afterward.

The AGI Landscape in 2026: A UNI Reading

A field-level reading of how the intelligence conversation has shifted, and why framing UNI as an attainable path toward GNI is more useful.

Why We Refuse the AGI Word

A public position statement: what the wording forbids, what it commits us to, and what we build instead.

Cluster

Generative models

Generative models as the mathematical backbone of active inference: hidden states, observations, factorization across time, and how the workbench encodes a working model.

Generative Models: The Organism's Model of Its World

Pillar. Hidden states, observations, priors, likelihoods, and factorization across time.

Hidden States vs Observations: A Builder's View

An engineer-facing explainer on the split between what the agent sees and what the agent infers.

Priors and Likelihoods in Plain Language

A gentler pass through priors and likelihoods for a technical reader not yet at home in Bayesian vocabulary.

Prior Preferences and Goal-Directed Behavior

How prior preferences over observations shape goal-directed action in active inference.

Factorization, Time, and Hierarchy in Generative Models

How a generative model factorizes across time steps and hierarchical levels, and where the workbench lands.

Discrete vs Continuous Time: A Modeler's Choice

When a discrete-time POMDP is the right frame, and when a continuous-time formulation earns its keep.

Why Generative Models Are Not Neural Networks

A statistical generative model is a joint distribution. A neural network is a class of function approximators. The two are constantly confused.

Encoding a Generative Model in Elixir: A Sketch

A code-level sketch of how UNI encodes generative-model state and priors in the Elixir workbench.

Our Elixir Workbench: Why That Stack

Why UNI's active-inference workbench runs on Elixir and the BEAM: what it makes easy, what it makes hard.

Cluster

KL divergence and Bayesian inference

The workhorse metric of variational inference and the epistemic engine of active inference. What the quantity measures, why the direction matters, and how the bound bites.

KL Divergence and Bayesian Inference in Active Inference

Pillar. Variational free energy as an upper bound on surprise, with KL divergence as the workhorse metric.

KL Divergence: What It Actually Measures

A careful explainer on Kullback-Leibler divergence, why it is asymmetric, and why the direction matters.

Variational Inference: A Conceptual Walkthrough

The variational approximation, the evidence lower bound, and why minimizing VFE is equivalent to approximate Bayesian inference.

The Evidence Lower Bound (ELBO) in Active Inference

How the ELBO shows up in the active-inference literature, what it bounds, and why that bound matters.

Bayesian Model Comparison for Practitioners

How Bayesian model comparison guides model selection when you have several candidate generative models.

Conjugate Priors: When They Help, When They Hide Work

Why conjugacy makes updates tractable, and when the closed form quietly hides the modeling work.

Why Minimizing Surprise Is Not Avoiding Novelty

Minimizing variational free energy does not mean the agent becomes a rock. Expected free energy adds an epistemic drive.

Cluster

Expected free energy

The quantity active inference minimizes over policies. The two-part decomposition, planning as inference, and how it is wired in the workbench.

Expected Free Energy and Goal-Directed Action

Pillar. The pragmatic-plus-epistemic decomposition, and why the split matters for building learning agents.

Epistemic vs Pragmatic Value

The two-part decomposition and how it drives exploration versus exploitation.

Planning as Inference: What That Phrase Means

The same variational machinery that infers hidden states can be repurposed to infer good actions.

Policy Selection: A Conceptual Walkthrough

How active inference scores and selects policies under expected free energy, with a small worked example.

Action Selection in the UNI Workbench

Code-level walkthrough of how UNI's action selection loop is wired: enumeration, EFE scoring, softmax over G.

EFE vs Reward: A Careful Comparison

What each objective captures, where they overlap, and where they diverge.

Why We Do Not Use Value Functions

Why the UNI active-inference build does not lean on classical value functions, and what we use instead.

Cluster

Markov blankets

The statistical boundary that defines a system apart from its environment. Sensory and active states, nested blankets, and the modeling choice at the heart of the frame.

Markov Blankets: What They Are, and What They Are Not

Pillar. Sensory and active states, internal and external states, and why the blanket is a modeling choice.

Sensory and Active States: A Working Decomposition

The four-way state decomposition explained for engineers, with a note on where the split is a modeling choice.

Nested Blankets and Hierarchical Agents

How Markov blankets can nest to model hierarchical agents, what the abstraction buys, and where it leaks.

The Blanket Debate: What We Take as Working Assumption

An honest tour of the ongoing debate around Markov blankets, and where UNI takes a working assumption.

Why the Boundary Is a Modeling Choice

Where you draw the boundary depends on what you want to explain, and different choices give different agents.

Boundaries, Agency, and What a System Is

Why boundary drawing is central to modeling agents, and how UNI treats agency as a statistical claim.

Cluster

The benchmark and the paper

The Stratified Palimpsest benchmark and the Zenodo preprint. What the benchmark actually tests, what it does not, and how to reproduce a run from public artifacts.

The Benchmark and the Paper: The Stratified Palimpsest

Public receipts page for the Stratified Palimpsest benchmark and the Zenodo preprint.

The Benchmark: What the Stratified Palimpsest Actually Tests

A layered POMDP benchmark for active-inference controllers, and how the falsifiers are wired in advance.

What the Benchmark Does Not Measure

Honest limits of the Stratified Palimpsest and why we do not claim more than the benchmark supports.

Layered Temporal Tasks and Why They Matter

Why layered temporal structure is a hard test for any inference-based agent.

How to Replicate Our Benchmark Run

A step-by-step recipe for reproducing the UNI Cell Lab benchmark from public artifacts.

Reading the Zenodo Preprint: A Guided Tour

A section-by-section reader's guide to the Namjoshi 2026 Zenodo preprint on active inference and free-energy minimization.

SeedIQ and ARC-AGI 3: A Third-Party Datapoint

A third-party result as context, not a claim: active-inference-flavored systems on hard out-of-distribution tasks.

Cluster

Gates and falsifiers

Every claim in the UNI ledger carries an evidence class and a falsifier. How that discipline is written down, and how a reader can push back on it.

Gates and Falsifiers: How We Know When We Are Wrong

Pillar. How gate discipline works, and how a few gates look while they are in flight.

Evidence Classes A to E: What They Mean

A reference for the evidence-class tags used across the UNI family.

How We Log Honesty in the UNI Ledger

What a UNI ledger entry looks like, how each entry is evidence-classed, and how a public claim traces back to it.

Reading a UNI Claim: A User's Guide

How to parse a public claim about UNI, find its evidence class and falsifier, and its entry in the ledger.

How to Push Back on a UNI Claim

A practical guide for filing a challenge against a UNI claim: what to send, how we adjudicate it, and how the ledger records the outcome.

What Would Falsify UNI: A Standing List

An always-current list of the observations that would force us to revise or retract the UNI working hypothesis.

What UNI Does Not Claim

A standing list of the claims we deliberately do not make. Public red lines, kept in the open.

When We Were Wrong: A Running Log

A public log of positions the UNI project has had to revise. Dates, before, after, and the trigger for each revision.

Cluster

Science, in the open

The operating stance behind UNI: publish claims with evidence classes, publish the gates, ship code where we can, and invite challenge.

Gentle Entry: A Quiet Way Into the Science

Low-pressure, one idea at a time. Watch a prediction meet the world, then choose your own next step into the science, a lab, or the transparency page.

Science in the Open: What We Mean

The operating stance behind UNI: what it costs, and what it buys.

A note on how these posts talk.

Every post carries evidence-class tags, cites its sources, and points at the gates and falsifiers that would force a revision. UNI is a working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails. Nothing here is medical advice.