Pinned Tweet
It is the deepest honor to have been joined by Michael Levin (@drmichaellevin), Victoria Klimaj, Zahra Sheikhbahaee (@zah_bah), Dalton Sakthivadivel (@DaltonSakthi), Adeel Razi (@adeelrazi), David Ha (@hardmaru), Nick Hay, Kevin Schmidt, Irina Rish (@irinarish), David Krakauer (@sfiscience), Melanie Mitchell (@MelMitchell1), Samuel Gershman (@gershbrain), and Joshua Tenenbaum in organizing this special issue of the Royal Society’s (@RSocPublishing) Philosophical Transactions A:
“World models, A(G)I, and the Hard problems of life-mind continuity: Toward a unified understanding of natural and artificial intelligence”
royalsocietypublishing.org/r…
This collection was motivated by a question with far reaching implications, ranging from the fundamental nature(s) of mind to choices that may determine the future of our civilization/species: what kinds of world modeling capabilities are likely to be realized by which kinds of minds and what world might we be in with respect to increasingly advanced artificial intelligences?
Will the scaling and refinement of present approaches result in AI with human-like (and beyond) cognitive abilities, or do we need radically different paradigms that more closely follow the principles of natural intelligence? Learning “world models” to predict/compress information may be how biological learners so efficiently learn (to learn) to achieve goals and generalize that knowledge across a broad range of task environments. World models may also be useful for reverse-engineering forms of “System 2” cognition, or the self-reflexive, deliberate, multi-step reasoning associated with cognitive capabilities that may be unique to humans. Predictive models that reflect how the world may be causally modified by actions allow agents to adaptively control their behavior with flexibility and context-sensitivity. Spatiotemporally and causally coherent models of the physical world may not only be the key for creating AIs that we can rely on for real-world deployment, but may even be the (dynamic) core of conscious cognition.
The contributions to this special issue consider the varieties of world models worth modeling from diverse points of view:
Douglas Hofstadter explores whether sufficiently coherent self-referential world modeling could ground meaning, consciousness, and a genuine “I” in future AI systems.
David Krakauer (@sfiscience), Melanie Mitchell (@MelMitchell1), and John Krakauer (@blamlab) examine the principles of emergent intelligence from a complex systems perspective.
Alexander Ku (@alex_y_ku), Declan Campbell, Xuechunzi Bai (@baixuechunzi), Jiayi Geng (@JiayiiGeng), Ryan Liu (@theryanliu), Raja Marjieh (@RajaMarjieh), R. Thomas McCoy (@RTomMcCoy), Andrew Nam, Ilia Sucholutsky (@sucholutsky), Liyi Zhang (@LiyiZhang_Leo), Jian-Qiao Zhu (@JQ_Zhu), and Thomas Griffiths (@cocosci_lab) argue for using the tools of cognitive science to understand and evaluate LLMs across multiple levels of analysis.
Evelina Leivada (@EvelinaLeivada), Gary Marcus (@GaryMarcus), Fritz Günther, and Elliot Murphy (@ElliotMurphy91) test whether LLMs deeply understand language and the “world behind words,” or primarily learn surface statistical regularities.
Pedro Tsividis (@ptsividis), João Loula, Jake Burga, Juan Pablo Rodriguez, Sergio Arnaud, Nate Foss (@_npfoss), Andres Campero, Ajay Subramanian (@ajaysub110), Thomas Pouncy, Samuel Gershman (@gershbrain), and Joshua Tenenbaum introduce a theory-based meta-learning architecture inspired by the remarkable flexibility and efficiency of human cognition.
Eunice Yiu (@eunice_yiu_), Kelsey Allen, Shiry Ginosar (@shiryginosar), and Alison Gopnik (@AlisonGopnik) explore empowerment, controllability, and causal reasoning as means of understanding the remarkable learning abilities of both child and adult minds.
Nadav Amir, Stas Tiomkin, and Angela Langdon investigate how goals shape the structure of experience and how the world modeling abilities of natural intelligences may be inseparable from values.
Vickram Premakumar, Michael Vaiana, Florin Pop (@FlorinPop17), Judd Rosenblatt (@juddrosenblatt), Diogo Schwerz de Lucena, Kirsten Ziman, and Michael Graziano show unexpected benefits of self-modeling as an inductive bias and regularizer for training artificial agents.
Hanlin Zhu, Baihe Huang, and Stuart Russell analyze why model-based reinforcement learning may fundamentally outperform model-free approaches in representational efficiency.
Bradly Alicea (@balicea1), Morgan Hough (@mhough), Amanda Nelson, and Jesse Parent (@JesParent) revisit fundamental cybernetic principles of regulation, adaptation, and world modeling across a wide assortment of complex adaptive systems.
Francesco Sacco (@FrancescoSacco1), Dalton Sakthivadivel (@DaltonSakthi), and Michael Levin explore topological constraints on self-organization and suggest that biological systems maintain long-range coherence in ways that are fundamentally different from current transformer architectures.
Georg Northoff (@NorthoffL), Yasir Catal, and Samira Abbasi examine how biological intelligence may depend on capabilities for flexible “inner time” to ensure adaptive alignment between the dynamics of system and world.
Nicolas Rouleau (@DrNRouleau) and Michael Levin explore whether theories of consciousness generalize beyond brains to unconventional embodiments and living systems more broadly.
Benjamin Lyons and Michael Levin investigate economies and collective intelligence as systems coordinated by “cognitive glues” in the form of shared models of scarcity and value.
Katherine Collins (@katie_m_collins), Umang Bhatt (@umangsbhatt), and Ilia Sucholutsky (@sucholutsky) consider “Rogers’ paradox” to demonstrate ways in which collective learning is impacted by different kinds of human-AI interactions.
Ruairidh Battleday (@RMBattleday) and Samuel Gershman (@gershbrain) distinguish between the “easy” and “hard” problems of science, and describe how while current AI systems demonstrate powerful narrow forms of optimization with respect to well-defined inference-spaces, further developments are needed for achieving capabilities for novel scientific discovery.
Fritz Breithaupt (@FritzBreithaupt) explores narrative world models and the roles of uncertainty and transformative experiences in natural intelligences, suggesting that coherent agency may depend on better understanding human-like meaning-making.
Taken together, these diverse perspectives suggest that while LLMs can clearly learn powerful generative models of language, they likely do so without having world models of sufficient spatiotemporal and causal coherence to achieve human-like reasoning abilities, capacities for generating subjective conscious experiences, or pathways to realizing artificial general superintelligence. However, by further developing world modeling architectures, we may eventually be able to create forms of intelligence that recapitulate the remarkable flexibility and generality of human intelligence. Finally, enhanced (e.g. more coherent/integrated) world models may not only afford expanded capabilities, but could potentially help ensure that increasingly powerful AI systems achieve both inner and outer alignment with human(e) values.
🤠💗🤠
In here is a video of The Bitter Lesson as a pretty-good country-music song. Enjoy.
oneusefulthing.org/p/the-dot…
2
230
Looking forward to reading this.
Brain asymmetry operates as a multivariate organizing principle rather than a simple left-right dichotomy, a view reinforced by advances in neuroimaging and genomics.
doi.org/10.1038/s41583-026-0…
8
413
"The do-operator... It's the difference between association and intervention. ..the do-operator itself, connects the operation of doing to something that we can see. So as opposed to the purely observing, you're making the choice to change a variable. That's what it expresses. And then, the way that we interpret it, the mechanism by which we take your query, and we translate it into something that we can work with, is by giving it semantics, saying that you have a model of the world, and you cut off all the incoming arrows into x, and you're looking now in the modified, mutilated model, you ask for the probability of y. That is interpretation of doing x, because by doing things, you've liberated them from all influences that acted upon them earlier, and you subject them to the tyranny of your muscles. So you remove all the questions about causality by doing them."
nitter.cf/adamsafron/status/1329…
This is a good slide, nicely summarizing what causal models enable us to do, and how little we need to assume to get them to work. Thanks for posting. @ConjectureInst
Excerpts (with some slight paraphrases) from a fascinating presentation by @yudapearl on the question of which kind of machines can (and ought to) have free will:
If free will is an illusion, why did evolution labor to endow us with it?
Should we program the next generation of computers to communicate with ‘illusions’ such as credit, blame, guilt, regret, and attribution?
Would robots play better soccer if equipped with credit and blame vocabulary?
The phenomenology of free will/choice is generated by explicating world knowledge into a Belief-Desire-Intention model.
The survival value of this intentional model/stance lies in efficient communication, lean performance, and moral responsibility.
The computational benefits of moral responsibility lie in
1) the setting of priority
2) efficient communication using world models
3) purposeful conversion of implicit to explicit knowledge
nitter.cf/yudapearl/status/21039…
See: "Can Computers Have Free Will?" A video of the talk I gave at Cal Tech: youtu.be/rnW5KZmuH0Q?si=y5I_… youtube.com/watch?v=rnW5KZmu…
Looking forward to reading this.
Finally published:
“Symmetry as a common origin of hierarchical emergence and abstraction”
science.org/doi/10.1126/scia…
Symmetry generate hierarchical macroscales and shape the structure of Bayesian beliefs, making inference of high-dimensional systems tractable
1
7
464
Looking forward to re-reading this:
nitter.cf/drmichaellevin/status/…
Replying to @drmichaellevin
Part of a very cool theme issue ‘Making and breaking symmetries in mind and life’ organized by @adamsafron, @adeelrazi, Zara Sheikhbahaee, @DaltonSakthi and Magnus Bein
royalsocietypublishing.org/t…
with our (actually, MidJourney's) cover:
1
120
Beautiful paper by @RubenLaukkonen, @shamilch, @_fernando_rosas, and others, and which has extensive overlap with ideas I have been publishing on since 2020 with Integrated World Modeling Theory (IWMT).
frontiersin.org/journals/art…
frontiersin.org/journals/com…
[I also presented on IWMT to Morten Kringelbach when I visited his lab at @UniofOxfordin 2018 to discuss points of overlap between my proposed integration/binding mechanism of "self-organizing harmonic modes" and @DrSelenAtasoy's seminal work on "connectome harmonics."]
Kringelbach et al. (2026) write:
"Most theories of consciousness isolate one of the operations that the Entangled Loop treats as inseparable... Each theory captures an important property of conscious processing, but none begins with the biological problem that precedes them all, namely how a living system continually decides what matters, binds the distributed information bearing on that decision and gives the resulting content control of one body."
Grounding consciousness as a cybernetic control process for embodied-agents/living-organisms is precisely what I attempted to do with IWMT, whose core claims are as follows:
1. Consciousness is what it feels like when generative models for embodied agents infer likely (body-)world states (i.e., qualia), given past sensory experience/data.
2. The stream of experience (for control and learning) is generated as a series of iteratively estimated system-world relations (i.e., chained qualia), calculated with sufficient rapidity to both adaptively inform and be informed by action-perception cycles on the timescales of their formation.
3. Qualia are valenced by virtue of being associatively linked to histories of agentic inferences/experiences about bodily states whose value is grounded in life-management (e.g. interoception and homeostasis/allostasis).
4. Phenomenal-experience-entailing world models require coherence with respect to space (organized locality), time (proportional changes in space), and cause (predictable changes in space and time, particularly with respect to self-generated actions).
5. Various forms of conscious access/awareness and meta-cognitive reflection require multiple kinds of self-models with counterfactual (or imaginative) processing abilities.
For a very brief synopsis of IWMT, please see this entry on Landscape of Consciousness (@CloserToTruth):
loc.closertotruth.com/theory…
For a brief but slightly more detailed summary of IWMT and the elaboration of the theory over the past several years, please see this conference proceedings paper cowritten with @VictoriaKlimaj and @zah_bah, presented at a fascinating workshop at @RealAAAIorganized by the California Institute for Machine Consciousness (@CIMCAI):
ojs.aaai.org/index.php/AAAI-…
The major thing that is lacking from previous presentations of IWMT that is present in Entangled Loop is the emphasis on feelings/emotions and pleasure-desire cycles as primary. While I have an extensive discussion of these issues in a manuscript published in Entropy that draws on IWMT and is referenced in subsequent papers, I should have done a better job of clearly foregrounding the primacy of affect in other publications discussing the theory.
mdpi.com/1099-4300/23/6/783
Safron (2021) wrote:
"Consciously-experienced feelings may provide unified attractors for coordinating global organismic states. While emotional shaping may occur without consciousness, these affects may be more likely to entrain the overall system when integrated into coherent fields of experience. Even if these feelings take the form of “isolated qualia” without conscious access (as described above), these self-stabilizing cores may still provide sources of greatly elevated control energy. However, this entraining power would be even greater when made consciously accessible in ways that afford planning and continuous adjustments of actions based on organismic value. This mapping of hedonic states onto consciously introspectable models of enaction also provides a partial means of handling the credit assignment problem, via conjoining value and actions in both experience and memory..."
I also made a Gemini notebook to explore points of intersection between IWMT and Entangled Loop that may be of interest (and about which I would be very interested to know the thoughts of the authors):
notebook.google.com/notebook…
New paper!
In the beautiful loop, we proposed three features that a good theory of consciousness should explain:
1) Selection: what becomes conscious and why
2) Binding: how we sustain a coherent world model
3) Awareness: how we know what we know
The main critique of that paper is that we were substrate-agnostic, and didn’t make a strong case for where these mechanisms are instantiated in human brains and bodies.
In this paper—which is a fruition of decades of bleeding edge neuroscience by Prof. Kringelbach and Prof. Deco—we fill that gap and provide implementations for the three criteria in the dynamics of the human brain.
In terms of our three criteria:
1. Selection: the affective valuation system encodes the preferred states. Or put simply, emotion defines what matters to a living organism enough to breach consciousness.
2. Binding: the process that binds distributed inferences in the brain is quantum-like (i.e., non-quantum entanglement) and emerges naturally out of a classical system modelled using coupled Staurt-Landau oscillators with connectomic topology. This regime enables the brain to exploit interference effects that produce integration at low thermodynamic cost.
3. Awareness: Finally, the meta-loop afforded by functional rich-clubs in the brain evince the secondary, higher-order processing reminiscent of epistemic depth and the sensation of knowing what we know.
The resulting Entangled Loop satisfies the computational criteria provided by the Beautiful Loop while expanding it to specific functional substrates in the mammalian brain. It also provides a nice unification of several existing ToCs.
Global Workspace Theory and the affective traditions of Panksepp, Damasio, Solms and others capture the ignition and broadcast of conscious content (i.e., selection). Integrated Information Theory emphasises the unity and integration of experience (i.e., binding). Higher-Order theories foreground the recursive representation that enables us to know that we know.
The Entangled Loop proposes that each tradition has identified part of the total architecture needed for consciousness, while somewhat missing an integrated picture.
Perhaps the deepest implication concerns the hard problem. The usual framing begins with neutral information processing and then tries to figure out how feeling could possibly emerge. But biological inference was never neutral. From its evolutionary beginnings, it had to distinguish what was good or bad for the organism, what should be approached or avoided, and ultimately what mattered to it.
On this view, consciousness is an entangled loop of mattering, binding and knowing.
Open Access:
sciencedirect.com/science/ar…
Congrats especially to Morten who really brought this to life, but also all the coauthors including @shamilch and @_fernando_rosas.
Developing advanced intelligence and solving alignment problems/opportunities together with games?
Bootstrapping the infinite game off of finite games?
I believe @colosseumsai/@tarinziyaee is a project/person very worth watching/supporting.
🚨 Personal news!
2000 years ago, Rome figured out a key idea Silicon Valley still hasn't.
Arenas can train - and monetize - the best embodied intelligences.
We're building embodied agents the same way.
Introducing Colosseums AI.
2
262
Deeply enjoyed/loved having this conversation with John Krakauer (@blamlab) and @michaelgarfield on the quest to understand and reverse-engineer cognition (and consciousness!) in natural and artificial intelligences.
Best past attempt at explaining consciousness:
ojs.aaai.org/index.php/AAAI-…
Best past attempt at explaining humor:
nitter.cf/adamsafron/status/1363…
With AI on everyone’s mind, we have to ask: what does it mean to think?
When does the brain actually create predictive models and rely on mental representations — and when is it really just a matter of inputs and outputs with no simulation going on in between?
Why do we feel pain and find things funny? Not just avoid damage or understand the logic of a joke, but actually feel these things?
Agreeing on what we don’t know matters, not least because trillions of dollars are invested in the dream of artificial general intelligence — but if thinking emerges in AI, how would we even know where to look? Behavioral indicators aren’t enough. What is the internal structure that would let us say with confidence “There’s real thinking going on here?”
Because as the success of deep neural networks makes clear, most intelligence doesn’t need this extra layer. Most intelligence isn’t thinking — at least, at the level of the whole organism or AI system. And, as much as we want to believe otherwise, maybe none of it has to be.
We need to remind ourselves what we don’t know. This is why guest co-host @adamsafron and I called on neurologist and neuroscientist John W. Krakauer, Director of the Johns Hopkins @blamlab and External Professor at the @sfiscience, who has spent decades studying the modularity of the brain and the relationship between sensorimotor functions and deliberative thought through research on stroke patients.
Our conversation with John gets at the very heart of the biggest taboo in cognitive science: that there is a difference between intelligence and thinking…that we’re conscious, and we don’t know why, and these distinctions matter to intelligence research. As AI demonstrates more and more capability to do things we thought it took conscious thought to achieve but shows no indisputable concrete evidence of being conscious, it turns out science still doesn’t understand what consciousness is.
Humans On The Loop Episode 41 is now everywhere you go for podcasts!
#cognitivescience #ArtificialIntelligence #neuroscience #sciencepodcast #consciousness
Wonderful interview with one of the broadest/deepest thinkers I know.
@FritzBreithaupt
New episode on The Narrative #Brain:
@FritzBreithaupt and his lab looked at how stories evolve as they are retold between over 20,000 human subjects. And then compared it to the same stories changing over time as they are retold by ChatGPT.
Depending on who you are, what they discovered either clarifies the real difference between humans and AI, unsettles our ideas about selfhood by anchoring it in the rupture of experiential learning, or suggests a path forward for the future of technology.
Fascinating conversation with Fritz and cognitive neuroscientist @adamsafron for our world models mini-series!
Full episode and interactive discourse map for the whole Phil Trans A special issue anywhere you go for podcasts (but here)
Chapters
00:00 Intro
06:56 Meet Fritz Breithaupt
08:08 A Childhood Rupture
11:41 Narrative as Therapy
13:30 Telephone Games
18:26 Emotions as Story Anchors
21:04 Prediction and Multiverses
23:58 Unnamed Narrative Emotions
31:28 AI Retellings vs Humans
34:07 Perspective-Taking and Experience
43:58 Limited World Models
48:41 Curiosity and Scientific Creativity
54:03 Suspense and Surprise
55:26 Three Modes of Experience
01:02:40 Self Modeling Tradeoffs
01:07:56 Reinventing the Self
01:14:21 Genre Collapse and Play
01:19:53 Sharing Experience by Story
01:27:47 Continual Learning and Agency
01:32:42 Mortality and Selfhood
01:38:44 Self as Process Not Object
01:45:09 Plural Consciousness and Politics
01:48:08 Outro
#AI #cognitivescience #neuroscience #storytelling
I find Georg @NorthoffL's perspective(s) to be extremely compelling, and we explored a few points of intersection with IWMT (nitter.cf/adamsafron/status/2106…) in this semi-recent paper:
royalsocietypublishing.org/r…
Also extremely interested in @MillerLabMIT's theories of consciousness, and in particular his emphasis on traveling waves. His suggestions regarding analog computing seem consonant with this work on "complex" (or (quasi-)quantum) harmonics by Deco et al., as well as work from @wellingmax's group on "flow equivariant" networks for modeling continuous phenomena:
journals.aps.org/pre/abstrac…
arxiv.org/abs/2409.13669
Also compelled by the idea that ephaphtic coupling may be important, although recent evidence seems to suggest it is not significantly implicated with cortical traveling waves in general:
cell.com/neuron/abstract/S08…
Although perhaps electromagnetic fields evoked by population activity could be important for some parts of the brain if my speculations are correct regarding the importance of concentrated message passing in geometrically-structured cortical hubs as a source of workspace dynamics and phenomenal binding of body-world estimates (i.e., it seems that electromagnetic fields may not be able to effectively enslave neural dynamics if they involve spatially distributed functional networks).
Replying to @adamsafron
@adamsafron Are you familiar with Northoff (TTC) and Earl K. Millers (Analog brain waves) theories? This experiment seems to support pieces of your construction as well as theirs.
2
4
587
"(i.e., it seems that electromagnetic fields may not be able to effectively enslave neural dynamics if they involve spatially distributed functional networks)."
Correction: "spatially distributed functional networks" should read "highly spatially distributed functional networks (of effective connectivity)"
Although something that's unclear to me is what constitutes too "highly spatially distributed" for electromagnetic fields to be ineffective at constraining the synchronization manifolds of the neuronal activity from which they emerge with circular causality?
1
1
87
Looking forward to reading this.
Theta rhythms may support social interactions.
(We are finding that theta provides a broader temporal scaffold, organizing local processing at higher frequencies, gamma vs alpha/beta).
nature.com/articles/s41467-0…
Very well said.
And that's precisely what I've been doing for the better part of a decade:
loc.closertotruth.com/theory…
Slide 8 of this presentation has an attempted fit-on-a-t-shirt encapsulation of the (synthetic) theory:
1drv.ms/p/c/bfa212aa225a3f04…
Sam Harris' confidently asserted category error 'argument' regarding free will makes no sense. It is literally nonsense in terms of being "not even wrong."
Causation doesn't cross ontological levels, and for causal analyses that do involve level-crossing, macro-level coarse grainings can have greater predictive power and channel capacity for communicating information about a system.
@yudapearl
@erikphoel
What if heat didn't simply spread out, but travelled in focused beams like light?
Researchers have now observed exactly that at room temperature.
Inside boron arsenide, long-lived phonons carry heat along preferred crystal directions, producing a ray-like patterns instead of ordinary circular diffusion.
1
2
330
Looking forward to reading this.
Can we build world models purely in the latent space, without pixel prediction?
We present Contrastive World Models - we train the latent states to maximize mutual information with future observations, without any decoders / pixel reconstruction.
Contrastive World Models enables
⚡ Substantially more robust representations
🤖 More efficient training by removing the pixel decoder entirely
🌍 A general approach with minimal assumptions [1/n]
283




















