Bio

<aside> 💡 Dr Sanjeev Namjoshi is a machine learning engineer at Noumenal Labs and a prominent researcher in the free energy principle and active inference epistemic community. He has a PhD in neuroscience from The University of Texas at Austin and is the textbook author of ‘Fundamentals of Active Inference: Principles, Algorithms, and Applications of the Free Energy Principle for Engineers’ and is writing a second book on Bayesian mechanics.

</aside>

https://youtu.be/JsZVS937GHk

General questions

• How would you explain your phenomenology of writing a textbook through active inference (AIF)?
• How would you articulate the difference between FEP and AIF technically, conceptually and perhaps philosophically?

• What is Bayesian mechanics, and how did FEP/AIF spawn this field?
  • Along these lines, can you reflect on Schrödinger’s (1992) famous quote, “How can the events in space and time which take place within the spatial boundary of a living organism be accounted for by physics and chemistry?” that Friston (2013) start his landmark paper with?
• As you point out in the introduction, the FEP world has many moving parts and evolving theories/concepts, be it deep active inference, POMDP, multi-agent AIF, etc., but what are the core mathematical concepts of FEP and AIF? What type of “first-principles math” should one know to even engage with the field?
• What is the relationship between physics and FEP? Why, for instance, does statistical physics allow us, as it were, to have a physics of intelligence? (Maxwell et al., 2023)
• Even for those of us without a deep mathematical training, it’s well-known that information theory (InfT) is foundational for AI research: Can you take a moment to explicate InfT’s—and information geometry’s (Namjoshi, pp. 394)—pertinence to AIF and broadly Bayesian mechanics?
• Fundamentally, how does AIF differ from the Reinforcement learning paradigm? (Namjoshi, pp. 394 - 396)
  • Along those lines, what differences between AIF and deep learning (DL)? For instance, do you see a type of distributed computing approach where systems built on fundamentally different principles work together? Here I’m thinking of the notion in Agüera y Arcas (2025) that intelligence is inherently social and network-based.
• In many parts of the book, e.g., Namjoshi (2026, p. xxiv, 89, 124), you mention FEP is a priori - this is a term that comes to the academic nomenclature mainly through Kantian transcendental philosophy. What do you mean by characterising FEP as being an a priori principle?

• Why is FEP usually compared to Hamilton’s principle of least action?
  • How empirically verifiable is AIF? For instance, is AIF falsifiable as a scientific theory? Here I am thinking of papers such as Andrews (2021).

Fundamentals of Active Inference

• You may have already answered this question in other interviews, but how does this book differ from the Parr (2024) one?

• Do you think the fact that AIF integrates from many fields gives it its explanatory power/coherence?

  

• Apropos of action and perception, what is statistical inference? (Namjoshi, 2026, p. xxii)

Generative Models