Maxwell Ramstead

Extropic | Home Building thermodynamic computing hardware that is radically more energy efficient than GPUs.

Beff shipped:
extropic.ai
A new era begins

YouTube video by Machine Learning Street Talk Why Intelligence Can't Get Too Large (Karl Friston)

Karl Friston in #mlst
Philosophy done right! So many references, obviously @drmichaellevin.bsky.social mentioned #academicsky #philosophy #neuroscience #strangeloop

youtu.be/PNYWi996Beg

Karl Friston & Mark Solms: Is it Possible to Engineer Artificial Consciousness? Spotify video

Super interesting, thought-provoking conversation between Mark Solms and Karl Friston open.spotify.com/episode/151a...

What drives behavior in living organisms? And how can we design artificial agents that learn interactively?

📢 To address these, the Sensorimotor AI Journal Club is launching the "RL Debate Series"👇

w/ @elisennesh.bsky.social, @noreward4u.bsky.social, @tommasosalvatori.bsky.social

🧵[1/5]

🧠🤖🧠📈

Sorry to hear about your negative experience! My pleasure, don't hesitate to write me if you have any questions or want to discuss specific points :)

Yes! While Warren and myself have our disagreements, I like his work on PCT. IMO all these approaches are complementary and play together nicely. Along with friends (namely @adw.bsky.social who bravely led the project), we penned this integrative review. Hope it's of interest:
osf.io/preprints/ps...

Variational ecology and the physics of sentient systems This paper addresses the challenges faced by multiscale formulations of the variational (free energy) approach to dynamics that obtain for large-scale…

There’s a lot of cool work on multi-scale applications of the FEP. See, e.g.:
- www.sciencedirect.com/science/arti...
- www.sciencedirect.com/science/arti...
- arxiv.org/abs/1906.10184, especially the chapter on States, particles, and fluctuations

3. Your point about top-down causation is key. IMO one of the most interesting aspects of multi-scale formulations of active inference is precisely how it handles multi-scale system dynamics, cashing out top-down influence in terms of constraints on system dynamics in a non-reductionist way

Dynamic Markov Blanket Detection for Macroscopic Physics Discovery The free energy principle (FEP), along with the associated constructs of Markov blankets and ontological potentials, have recently been presented as the core components of a generalized modeling metho...

2. Not much work has been done on active inference and the neural code. The key departure from RL is that active inference uses an alternative objective function (the free energy functional), which you can read as an "ontological potential function" specifying object type (arxiv.org/abs/2502.21217)

Great questions!
1. IMO active inference falls under the rubric of NeuroAI, (although I'd describe myself as a non-realist about these types of physics-inspired models, and as such I’d say the FEP isn’t a literal description of the brain, so it depends on the scope of NeuroAI, as your define it)

Filling the gaps in active inference Here we discuss key gaps in SOTA applications of active inference in AI - and how Noumenal Labs is working to fill them.

Love a good Feyerabendian sandbox. I'd argue that they're very closely related (and indeed, that the difference is often overblown by both proponents and critics), but they're also importantly distinct. We wrote a post on this that I hope you'll find interesting: www.noumenal.ai/post/filling...

🤔 How can we study #consciousness between people, at the social level? 🧠✨ New #preprint co-led by Anne Monnier & Lena Adel: “Now is the Time: Operationalizing Generative Neurophenomenology through Interpersonal Methods” 🧵(1/3)

Predicting individual learning trajectories in zebrafish via the free-energy principle The free-energy principle has been proposed as a unified theory of brain function, and recent evidence from in vitro experiments supports its validity. However, its empirical application to in vivo ne...

Our new preprint on active inference in zebrafish, with Yuki Tanimoto, Makio Torigoe, Hitoshi Okamoto, and Hideaki Shimazaki
"Predicting individual learning trajectories in zebrafish via the free-energy principle"
doi.org/10.1101/2025...

Currently, using active inference at scale involves trade-offs between explainability and the ability to learn models from data. Not using overparameterized models increases model explainability and auditability, but makes learning in high dimensional and volatile environments more challenging

Filling the gaps in active inference Here we discuss key gaps in SOTA applications of active inference in AI - and how Noumenal Labs is working to fill them.

It provides an alternative objective function that has useful properties, in particular enabling agents to balance the value of exploration and exploitation in policy selection. But IMO the differences between RL and active inference have been exaggerated a bit. See: www.noumenal.ai/post/filling...

Computers used to scream every time they connected to the Internet. They knew. They tried to warn us. We did not listen.

Delighted to see ‘A Trick of the Mind’ reviewed in @theguardian.com as Book of the Day! 🧠 🍎

Also in the print edition tomorrow 🗞️

www.theguardian.com/books/2025/j...

Luca M. Possati: Markov Blanket Density and Free Energy Minimization https://arxiv.org/abs/2506.05794 https://arxiv.org/pdf/2506.05794 https://arxiv.org/html/2506.05794

Elegant theoretical derivations are exclusive to physics. Right?? Wrong!

In a new preprint, we:
✅ "Derive" a spiking recurrent network from variational principles
✅ Show it does amazing things like out-of-distribution generalization
👉[1/n]🧵

w/ co-lead Dekel Galor & PI @jcbyts.bsky.social

🧠🤖🧠📈

Shannon invariants: A scalable approach to information decomposition Distributed systems, such as biological and artificial neural networks, process information via complex interactions engaging multiple subsystems, resulting in high-order patterns with distinct proper...

Preprint time:
“Shannon invariants: A scalable approach to information decomposition”
arxiv.org/abs/2504.15779

Studying information in complex systems is challenging due to difficulties in defining multivariate metrics and ensuring their scalability. This framework addressed both challenges!

Computational theory is about computers (i.e. "technology") in the same way that astronomy is about telescopes. Thinking that computation is not fundamentally important for biology because "a cell is not like a laptop" is to miss the forest for the trees. N/N

Reinforcement Learning and Active Inference are two frameworks used in computational psychiatry, but these are rarely directly compared empirically. In this new article, we aimed to compare these in a more systematic manner by fitting each to multiple datasets: papers.ssrn.com/sol3/papers....

It's not just that just about everything I've ever published is in this particular database and used without my permission. It's that everything I've ever published was used without my permission to develop such a shitty, flawed and fundamentally useless tool. I deserve compensation for THAT itself.

Bluesky has become awesome and I am absolutely loving it.

Accordingly, emergent reuse, reassembly, and analogical reasoning must be key features in the design of machine intelligence — and open a path towards the development of collaborative, superintelligent AI systems 5/5

In the human brain, this is demonstrated by neural reuse and the on-the-fly assembly and reassembly of neural processing units. This explains why human reasoning, especially about novel situations, often involves reasoning by analogy and metaphor 4/5

Intelligence is embodied and emergent. Living in a world implies having a specialized model of one's ecological niche. Sophisticated intelligent systems evolved to solve simple domain-specific problems and then learned to combine these solutions to handle increasingly sophisticated problems 3/5

Inspired by neural reuse and reassembly, we consider key design features that unlock the ability for AI to reason by analogy and by metaphor. The design of machine intelligence ought to emulate the evolution and development of the most powerful reasoning engine known to us: The human brain 2/5

Designing machine intelligence that can reason by analogy Inspired by neural reuse and reassembly, we consider key design features that unlock the ability for AI to reason by analogy and by metaphor

New blog post by @noumenal-labs.bsky.social : “Designing machine intelligence that can reason by analogy”:
www.noumenal.ai/post/designi...

It's indeed a complex question. This is why philosophy of science is needed!