Why Now?
Something is stirring in the active inference world — and if you’ve been paying attention to AI news lately, you may have felt it.
[This blogpost accompanies our YouTube: “Your Options for Learning Active Inference.]
Over the past several months, AIX Global Innovations introduced Seed IQ™ — a technology built on active inference rather than transformers — and demonstrated remarkable performance on the ARC-AGI3 benchmark. Running on a MacBook Pro, Seed IQ performed at or near top human levels while LLM-based solutions struggled to reach 1%.
The ARC-AGI3 challenge organizers do not show Seed IQ on their public leaderboard (shown in Figure 1), as AIX has not disclosed their proprietary method — but AIX has widely publicized their results.
We described these results in our previous two April 17 and April 23 blogposts. That performance result stands and deserves serious attention.
In our prior YouTube, we unveiled how Seed IQ represents a scaling breakthrough, comparable to the deep learning breakthrough of the 2012-2016 era, and transformers in 2017.
AIX has also made broader claims about fault-tolerant quantum computing (FTQC) on commodity IBM hardware — claims that attracted detailed technical scrutiny from Marin Ivezic, Founder of Applied Quantum and a quantum computing expert with decades of experience. Ivezic’s analysis raised specific technical objections, centered on the distinction between the d=3 and d=5 QEC demonstrations and the d=1 register used for the primary computation workloads. We are not in a position to adjudicate these claims — FTQC is outside our area of expertise — and we note them here because they are part of the public record.
What is more interesting to us — and where we believe the real conversation now sits — is how Denis Ovseyenko, AIX Co-Founder and Chief Innovation Officer, has since reframed the work. Rather than engaging point-by-point with the critique, Denis has provided important additional technical context. In a recent LinkedIn post, he describes how AIX initially pursued conventional surface code QEC, running genuine d=3 and d=5 experiments on IBM Heron quantum hardware. The results were real: at d=3, Seed IQ governance reduced logical error rate by 88.5% relative to standard industry decoders; at d=5, by 93.1%.
Figure 2 shows Denis’s telemetry dashboard from the d=5 run — standard decoders produced a 75.6% logical error rate while Seed IQ governance reduced this to 5.2%, and entanglement fidelity improved from 0.31 (below classical separability) to 0.83 (well into the quantum entangled regime). These results were obtained on real IBM quantum hardware.
What Denis observed, however — and what has been reported across the industry by Google, IBM, Quantinuum, and QuEra — is that increasing code distance improves error suppression per round while simultaneously degrading overall logical fidelity due to hardware overhead. This observation led AIX to a different architectural direction entirely: the Governed Encoded Register (GER), which seeks fault tolerance through Seed IQ governance rather than through expanding code distance. Whether this constitutes fault tolerance in the standard definition remains an open question that the quantum computing community is actively assessing.
For broader context: the U.S. Department of Energy announced its Quantum Genesis initiative on June 22, 2026, targeting fault-tolerant quantum computing by 2028. That same week, IQM Quantum Computers published peer-reviewed research showing significant advances in quantum error correction — while targeting practical fault tolerance by 2030
We will follow this story and report back as the picture becomes clearer.
This isn’t only an AIX story. The broader AI landscape is moving in directions that point toward active inference.
Active inference as a framework is gaining serious attention — from researchers, from engineers, from corporate directors who are suddenly asking what it is and where to start.
Pourya Kordi discusses DeepMind’s directions in his recent YouTube, ‘DeepMind Was Two Steps Ahead, AGAIN!.’ The problems that DeepMind are now publicly prioritizing — world models, continual learning, physics-grounded intelligence — are problems that active inference has been addressing mathematically for years.
The questions that Demis Hassabis is now publicly asking are ones that active inference researchers have been working on for over a decade.
The question now isn’t whether active inference matters. It’s where to start.
The Oregon Trail of AI
Most people who study AI start in what I think of as Elm Grove, Missouri — the figurative trailhead of the Oregon Trail. We start by learning backpropagation. We learn the rudiments of reinforcement learning. We get comfortable with the basics. It’s open prairie, relatively flat, reasonably well-mapped.
Then comes the generative AI territory — and here’s where most people either fall asleep on the tour bus or skip it entirely, flying over the mountains and landing in LAX. You look around, see palm trees, a lot of smog, and realize you’re not far from Silicon Valley — home of the LLMs, the transformer-based systems, the agents built on top of them. It’s busy, it’s loud, and it’s where most of the AI industry is right now.
But that stack is already showing its limitations.
The problems that active inference addresses — the ones showing up in DeepMind’s world model research and in the SeedIQ story — require exactly the mathematical foundations that most people flew over.
The real gold — and I mean this — is back in the mountains. Back in the generative AI mountain range, where three distinct disciplines converge into something genuinely new. Most people haven’t gone there yet. Not because they’re not smart enough. But because nobody gave them a trail guide.
That’s what this video — and this blogpost — are about.
Into the Mountains
The generative AI mountain range is not one mountain — it’s three distinct disciplines that converge into a single territory. Most people who have tried to learn active inference on their own have encountered at least one of them and found themselves disoriented, not because the material is beyond them, but because the learning resources were written for a different audience — physicists, statisticians, mathematicians — and the translation into AI terms is rarely done carefully or completely. Figure 3 shows how these three disciplines interlock.
The three mountain ranges you need to navigate are:
First: KL Divergence — specifically, the reverse KL Divergence, where you measure a model against data and keep adapting the model until it does a good job of representing that data.
Second: Bayesian inference with latent variables — hidden variables that your model has to infer rather than observe directly. This is the key conceptual move that makes active inference possible.
Third: Statistical mechanics — or “stat mech” as we call it. After the right mathematical origami, the equations of active inference look a lot like the equations of statistical physics. This isn’t coincidence — the mathematical structure is genuinely shared, even though active inference operates in an abstract space of beliefs rather than physical thermodynamics
Understanding just enough of that connection — enough to read the papers and understand why it matters — is the necessary ingredient most people are missing.
These aren’t easy mountains. But they’re navigable. And the view from the top is worth it.
Five Pathways Into Active Inference
The good news: you don’t have to navigate those mountains alone. There are now several serious resources available — each suited to a different kind of learner, a different starting point, a different level of commitment.
Here are five pathways worth knowing about.
Pathway 1 — AIX Global Innovations / Learning Lab Central
Denise Holt, co-founder of AIX Global Innovations, has been one of the most active educators in the active inference space, Her materials (built as Learning Lab Central and available at deniseholt.us) are specifically designed for tech CEOs, investors, and people encountering active inference for the first time. If you want a conceptual overview without heavy mathematics — a way to understand what active inference is and why it matters commercially — this is a solid entry point.
Pathway 2 – SolutionWright
Michael Polzin, founder of SolutionWright, has developed an exploratory environment for working hands-on with active inference components under the framework he calls Universal Natural Intelligence (UNI). If you’re technically inclined and want to experiment directly with active inference implementations, SolutionWright is worth exploring.
Pathway 3 — The Books
Two books deserve your attention:
Active Inference: The Free Energy Principle in Mind, Brain, and Behavior by Thomas Parr, Giovanni Pezzulo, and Karl Friston (MIT Press) — conceptually rich and relatively light on equations. A strong entry point if heavy mathematics is not your starting point. [Link to MIT Press or Amazon]
Fundamentals of Active Inference: Principles, Algorithms, and Applications of the Free Energy Principle for Engineers by Sanjeev Namjoshi (MIT Press, 2026) — more rigorous, more mathematical, the most recent comprehensive treatment of the field. A study group at the Active Inference Institute is currently working through this text. This is the book that takes you deep.
For serious students who want to work through the Namjoshi text, the Active Inference Institute has published a YouTube series that records the study group discussions. These are assembled in a full YouTube playlist.
Pathway 4 – The Active Inference Institute
The Active Inference Institute (AII) is the global nonprofit community hub for active inference — bringing together researchers, educators, engineers, and curious learners from around the world.
The AII offers a rich and growing collection of free resources:
- YouTube Channel — hundreds of videos including research presentations, guest streams, and tutorial sessions
- Fundamentals Study Group — online study group meeting weekly – currently working through Namjoshi’s Fundamentals of Active Inference chapter by chapter
- Parr et al. Active Inference prior book study groups — study notes and multiple recorded cohorts (e.g. this playlist for Cohort 9) working through Active Inference: The Free Energy Principle by Parr, Pezzulo, and Friston (2022, 2025).
- Applied Active Inference Symposium — annual gathering, recorded and archived – each year’s Symposium is recorded and archived, with a full YouTube recording for each year – see, e.g., Symposium 2025.
- Model Streams — researchers presenting current work
- Discord Community — active discussion forum
Sign up for AII communications and stay dialed in. Things are moving fast and the AII is at the center of much of it.
Pathway 5 – Themesis, Inc.
Full disclosure: this is us.
Themesis has several years of YouTube content spanning statistical mechanics, generative AI, and active inference. Our videos typically pair with blogposts — the YouTube provides the lecture, the blogpost provides supporting references, links, and resources. Together they form a coupled learning structure that goes deeper than either alone.
We currently offer two structured courses, both on Thinkific:
T3 — The Themesis Active Inference Conceptual Course (originally titled “Top Ten Terms (T3) in Statistical Mechanics“) covers the mathematical foundations: KL Divergence, Bayesian inference with latent variables, statistical mechanics, and active inference. Currently under revision to strengthen the Week 3 Bayesian content.
Building Active Inference in Python is our new hands-on lab course — six weeks of structured Python labs building an active inference system from the ground up. This is a pre-alpha release, built incrementally with an active cohort. The Python code is free, hosted on GitHub under the MIT license.
You can sign up as a Learner at themesis.thinkific.com, and get both overview course content and access Lessons marked as “free” for the two courses.
Where to Start — Right Now
If this blogpost has piqued your interest, the single most useful thing you can do (if you haven’t done so already) is join the Themesis community. We send regular updates on active inference developments, new educational resources, and course offerings — not daily, but when something genuinely worth your attention emerges.
Go to: https://themesis.com/themesis/ (From here on the Themesis site, go to the About page.)
Scroll down to find our Opt-In form, and do the opt-in – and please move our email to your preferred inbox.
And if you’re ready to go hands-on: Building Active Inference in Python is open for enrollment now.
https://themesis.thinkific.com
References and Resources
Seed IQ(TM) on FTQC – Initial Publications and Rebuttal
AIX Technical Report on the Seed IQ(TM) performance on FTQC.
- Holt, Denise and Denis Ovseyenko. 2026. “Governed Fault-Tolerant Quantum Computing on Commodity NISQ Hardware Surface-Code QEC, Universal FTQC Primitives, FTQC Composition, and Chemical-Accuracy FTQC across H2, LiH, H2O, BeH2 Ground States, and the BeH2 Strongly-Multireference Transition State on the IBM Heron r2 and r3 Families of QPUs.” Zenodo Preprint. doi:10.5281/zenodo.20585365. (Zenodo Preprint.)
AIX press release on Seed IQ’s performance on FTQC.
- AIX Global Innovations Press Release. 2026. “AIX Global Innovations Announces FTQC Breakthrough Quietly Achieved in April 2026, Accelerating the Quantum Compute Timeline.” BusinessWire (Jun 15, 2026) (BusinessWire.)
Marin’s refutation of the AIX claims for Seed IQ performance on FTQC.
- Ivezic, Marin. 2026. “AIX Global Innovations Claims Fault-Tolerant Quantum Computing on Rented Hardware. Their Own Paper Says Otherwise.” Postquantum. com (June 20, 2026).(Ivezic on Seed IQ’s performance on FTQC.)
Denis O.’s conversation on how AIX is moving in a different architectural direction entirely: the Governed Encoded Register (GER), which seeks fault tolerance through Seed IQ governance.
- Ovseyenko, Denis. 2026. Untitled LinkedIn Post.
Seed IQ(TM) on the ARC AGI-3 Challenge
One of Denis O.’s LinkedIn posts on their Seed IQ performance on the ARC-AGI 3 challenge.
- Ovseyenko, Denis. 2026. LinkedIn post: https://www.linkedin.com/feed/update/urn:li:activity:7456627497258921984/
More AIX publications regarding Seed IQ performance on the ARC-AGI 3 challenge, summarized in these two Themesis blogposts.
- Maren, Alianna J. 2026. “The AGI Landscape Just Changed: SeedIQ(TM), Active Inference, and What ARC-AGI 3 Actually Tells Us.” Themesis, Inc. Blogpost (Apr. 23, 2026). (Themesis on Seed IQ performance on ARC-AGI 3 – post #1.)
- Maren, Alianna J. 2026. “SeedIQ™ Just Stomped ARC-AGI 3 on a MacBook Pro. Here’s Why That Changes Everything.” Themesis, Inc. Blogpost (Apr. 20, 2026). (Themesis on Seed IQ performance on ARC-AGI 3 – post #2.)
This prior Themesis YouTube discusses Seed IQ(TM) as a scaling breakthrough, comparable to the deep learning breakthrough (2012 – 2016) and transformers (2017).
Recent Quantum Computing News
- IQM Quantum Computing. 2026. “IQM Achieves Milestone in Quantum Error Correction, Enabling Fault-Tolerant Computing in the Near-Term.” Business Wire (23 June 2026). ( IQM Quantum Computers )
- U.S. Dept. of Energy. 2026. “Energy Department Announces Initiative to Create and Deploy the World’s First Scientifically Relevant, Fault-Tolerant Quantum Computers.” U.S. DoE (23 June 2026). (US DoE Quantum Computing Initiative)