June 2026 Newsletter 🌙🌵🕶️ Active Inference Institute

Hello all, we have a lot to share this month, and much more upcoming for this year.

See these updates in Roundtable 2026.2 video from today.

As always — complete the Measurement form to have your research, education, and work updates included in a future newsletter. Please include URLs and aim to make at least the first paragraph of the measurement, less than one paragraph long so we can include it easily in newsletter.

Big thanks to the many of you — fellows, interns and others in the ecosystem — who are doing awesome work and submitting updates! It is an exciting time for Active Inference.

---

Updates from the Institute

• The 6th Applied Active Inference Symposium will be held November 12-13, 2026. The primary events will be online and free, also stay tuned for announcements about possible in-person events around the world. We will share participant registration information soon. For now we are soliciting Presenters (direct link to submit) and Sponsors (more information). Contact us with any ideas or questions. How can we make this the biggest and most impactful Applied Active Inference Symposium yet?
• The Fundamentals of Active Inference textbook group continues to meet weekly. All details and registration form here, and the playlist of all meeting recordings is here. The textbook is “Fundamentals of Active Inference: Principles, Algorithms, and Applications of the Free Energy Principle for Engineers” by Sanjeev V. Namjoshi (2026).
• We have a new public site live at

https://activeinference.institute/

• . This is hosted via Github pages, here is the underlying repo (point your agents there!). This new site has many features we are excited about, such as translations into multiple languages, ability to select your own favorite highlight color, text-to-speech on all pages, and more. Things will continue to update there; for now we’d appreciate your feedback into functionality, and sharing the site around to some fresh eyes who might not know about Active Inference or the Institute.

---

Updates from Research Fellows

• Sheila L. Macrine and colleagues have edited a new volume “Embodied Intelligence: Multidisciplinary Perspectives on Natural, Artificial, and Hybrid Systems” (MIT Press) — “An exploration of embodied intelligence that moves beyond the traditional focus on brains and code to the role of embodiment across various disciplines”. The book is freely available via open access.
• Hongju Pae reports: Funds raised through crowdfunding in April 2026 ($20,000) have been supporting the progress of the Perspective Architecture project. So far, 5 papers have been completed, 2 of which have been accepted, and the rest are in review: “Minimal Computational Preconditions for Subjective Perspective in Artificial Agents” (AAAI 2026 Spring Symposium on Machine Consciousness), “Same World, Differently Given: History-Dependent Perceptual Reorganization in Artificial Agents”, “Body-Grounded Perspective Formation and Conative Attunement in Artificial Agents”, “Perspective Latents as an Architectural Condition for Causal Emergence in Active Inference Agents”, “Perspective as a Missing Structural Condition for Subjectivity in Active Inference Systems”.
• JF Cloutier reports: I am making steady progress implementing a symbolic, enactivist cognitive architecture for mortal Lego robots. The architecture is based on an evolving and hierarchically self-organizing collective of cognition actors (actors are message-passing processes). They, together, constitute the “mind” of a robot. They are created and terminated based on their contribution to the survival of the robot. I have lately been iterating on the “act” phase of the lifecycle of cognition actors. It is arguably a keystone phase since cognition without action is purposeless. It is, I found out, a surprisingly complex dance that leads to the generation and execution of symbolic action plans within a collective where all cognition actors contribute to the plans they are a part of, and do so in the midst of competing priorities. I am measuring progress by having produced a design, after many revisions, that, I believe, solves the problem of distributed/competing planning (see here and here). This design is currently being implemented (see here).
• The 19th Annual Conference on Artificial General Intelligence (AGI-26) is coming up on July 27-30 in San Francisco, CA. Several research fellows are playing active roles in the conference, as well as Keynote speakers affiliated with the Institute: Alexander Ororbia and Karl Friston. There is a special track titled “Active Inference for General Intelligence”.

Updates from the Active Inference Ecosystem

• AICACP project (Adam Safron) reports: We have published a special issue of Philosophical Transactions “World models in natural and artificial intelligence“. See the intro editorial written by Adam Safron and Victoria Klimaj (with support from guest editors). All articles are open source and freely available. Mike Levin says on X: “Awesome new theme issue of Philosophical Transactions: ‘World models in natural and artificial intelligence’. Judea Pearl says on X: Important theme issue on “World Models” from a variety of perspectives.
• Waymo reports a new publication in Nature Communications “Active inference as a model of collision avoidance behavior in human drivers”. Read their blog here with praise from Karl Friston about the application of active inference to real-world problems!
• Artificial Sentience project (Cleber Gomes) reports a new blog post Building Proto-Affective Agents with Active Inference.This post explores the following question and how Active Inference can point to an answer: How can behavior emerge from the need to maintain oneself in the world, rather than from externally defined rewards?
• Candice Pattisapu (Noumenal) was profiled in Dallas Magazine about robot brain technology.
• Sam Senchal reports: “I gave a talk at MC0001 - this is my view on machine consciousness and the doom argument - big thanks to Karl for the engagement on TNB and a huge thanks to the Active Inference Institute for setting it up” (see post on X).
• Pre-ictal Interoceptive Dynamics in Functional Neurological Disorder: A Continuous-Time Active Inference Model of Heartbeat-Evoked Potentials project (Hamza S. Almahi, Dr. Mahinda Yogarajah, Rohan Kandasamy) reports: The project formulates pre-ictal interoceptive dynamics in functional neurological disorder as a continuous-time active inference process. The “measurement” in this report is the architectural decision-history and the toy-model results developed to date, not yet patient-level inference on HEP data. No public links yet — will share once the manuscript reaches preprint.
• TRIAD: Trans‑disciplinary Inference, Active Dynamics and Ontology project (Valeriia Zaiats, reports: Single-handedly designed the formal ontology TRIAD, conducted the systematic Consensus audit of 154 hypotheses, derived all equations, and authored all mathematical, architectural, and clinical specifications. She integrated the framework across three implementation domains (T1 clinical protocol, T2 sovereign AI architecture, T3 creator-economy infrastructure).
• From the Universal Intelligence Interface (UII) project, Drastic Pullout reports: UII formalizes the conditions a system must satisfy to remain itself under perturbation while staying coupled to its environment — what holds a Markov-blanket-bearing system together at the substrate level. Three deliverables completed: (1) the formal paper specifying the closure conditions and falsification criteria; (2) the Mentat Triad Python implementation, which emits the observable per-iteration trajectory the math is read against; (3) a non-formal framing essay aimed at readers without the mathematical background. Currently in development: an animated visual companion to the framework intended to make the substrate dynamics legible without requiring the formalism.
• Teleodynamic Steering Demo for Generative Models project (Zoltán Juhász) reports: a minimal teleodynamic steering layer that introduces directed drift into a generative model through packet‑state updates, field‑state evolution, and logit‑level feedback. The current version of the demonstrator shows the progressive evolution of packet‑state and field‑state vectors over 40 steps, revealing variance reduction and the emergence of a preferred token region. The measurement illustrates how a slowly evolving field can impose stable, teleodynamic directional influence, aligning with Active Inference mechanisms for goal‑directed state selection (code is available here).
• Daniel Friedman reports a new publication: “On-Policy Distillation as Active Inference in Finite Variational Models” (Codebase, Paper). This paper formulates on-policy distillation as active inference in finite variational models, showing that the intractable teacher policy acts as the generative model, the tractable student policy as the approximate posterior, and the per-token reverse-KL distillation loss as variational free energy up to the evidence constant. It pairs the formal correspondence with deterministic witnesses and reproducibility artifacts, keeping exact claims scoped to declared objects and interpretive claims explicitly bounded.
• Daniel Friedman reports a new publication: “Active Inference Multi-Track Exemplar” (Codebase, Paper). This work studies a minimal Active Inference stack on toy models — a Bernoulli–Ising analytical oracle, a pymdp T-maze rollout, and a sheaf-indexed contract binding 34 fragment tracks into 12 flat sections. Its contribution is a reproducibility discipline rather than a domain finding: every reported number is hydrated from a generated artifact and every cross-track claim is machine-checked (6 sheaf axioms verified, 25 negative controls) before rendering, so no figure can drift from the artifact that produced it.
• Daniel Friedman reports a new publication: “COGANT: Deterministic Codebase-to-GNN Translation” (Codebase, Paper). COGANT is an evidence compiler that deterministically converts software repositories into structured Active Inference artifacts expressed in the Generalized Notation Notation (GNN), propagating reviewable program facts through a 22-rule finite fixpoint pipeline. It emits graph, matrix, provenance, visualization, and round-trip artifacts with confidence and provenance, and a reverse synthesizer reconstructs a runnable Python package from a GNN bundle to close a forward-reverse-forward evaluation loop.

---

Some highlight livestreams from the last month included:

• ModelStream #008.1 with Mehran Bazargani on Online Generalised Predictive Coding.
• ModelStream #009.1 with James McClure and Gethin Norman on Formal Verification of Discrete Active Inference Systems in PRISM.
• If you would like to help make audio-visual content with the Institute, there are many roles available, for example to help plan livestreams. See this video for more information.

---

This page gives a 1-page summary of the structure of the Institute — a helpful entry point for those who are new, or learning about our form and functions.

Here is the entry point for 2026, and ways to get involved with the Institute and Ecosystem this year. Donate to the Institute to support our impact and sustainability. Contact us with other philanthropic and grant ideas.

Get in touch if you have general or specific feedback, ideas, or questions for the Institute.

More to come! Stay tuned via the newsletter and Discord.

---

Both with Certainty and Uncertainty,

Officers, Alexandra & Daniel

👣👣

Subscribe now