Can large language models *introspect*?
In a new paper, @kmahowald.bsky.social and I study the MECHANISM of introspection in big open-source models.
tldr: Models detect internal anomalies through DIRECT ACCESS, but don't know what the anomalies are.
And they love to guess “apple” 🍎
Happy to announce this new preprint!
In it, we use info decomp (ΦID) on fMRI in Alzheimer's and MCI to explore how info-dynamic representations change.
AD saw big decreases in synergy ('deductive' information) and increases in redundancy.
Check it out here:
👇
www.biorxiv.org/content/10.6...
"The Self-Evidencing Agent" - my new book - is out now with @mitpress.bsky.social
Can be purchased, or just download the whole thing for free, via the 'Open Access' option.
I'm grateful to @anilseth.bsky.social and Karl Friston for the generous endorsements.
mitpress.mit.edu/978026255389...
Uh oh .... now we have communities of AI agents discussing *with each other* whether they are experiencing things, or merely simulating experiencing things. Humans can only observe, not interact www.moltbook.com/post/6fe6491...
From Tai himself
mathstodon.xyz/@tao/1158558...
Classical billiards can compute
Classical billiards can compute.
With Isaac Ramos, we show that 2D billiard systems are Turing complete, implying the existence of undecidable trajectories in physically natural models from hard-sphere gases to celestial mechanics.
Determinism ≠ predictability. 🎱🧠 @upc.edu @ricardsole.bsky.social
Preprint time:
“Symmetries at the origin of hierarchical emergence”
arxiv.org/abs/2512.00984
On how symmetries generate hierarchical macroscales and shape the structure of our beliefs, making high-dimensional inference tractable
How do brain areas control each other? 🧠🎛️
✨In our NeurIPS 2025 Spotlight paper, we introduce a data-driven framework to answer this question using deep learning, nonlinear control, and differential geometry.🧵⬇️
What if there were a drug that allowed us to study consciousness in its most basic form?
Christopher Timmermann describes the potential of 5-MeO-DMT, a psychedelic that strips away everything but awareness.
Read the full article here: bigthink.com/neuropsych/5...
George Deane and Daphne Demekas win the € 20.000,- Computational Phenomenology of Pure Awareness Prize
mpe-project.info/wp-content/u...
for 2025 with this contribution:
osf.io/preprints/ps...
Christopher Timmermann
profiles.imperial.ac.uk/c.timmermann...
wins the €20.000,- "2025 Neuroscience of Pure Awareness Prize" with this contribution:
osf.io/preprints/ps...
The two main hubs in the brain for the processing of human time perception have been identified: SMA and Insula. Here Alice Teghil from Sapienza Università di Roma and I provide the conceptual background in our review on 'How the body and brain process time'. www.sciencedirect.com/science/arti...
Love the “all is fog” framing
The deadline for submissions for the 2025 Neuroscience of Pure Awareness Prize (€20k) is soon: Sept 30th. For the best contribution to neuroscience that substantially advances our understanding of the neural mechanisms underlying the experience of pure awareness. For details, see screenshot!
📜🪇[PUBLISHED]: NEW PREPRINT!🪇📜
I am incredibly excited to announce that we have published our paper on how "Dynamical independence reveals anaesthetic specific fragmentation of emergent structure in neural dynamics"
w@thomasandrillon.bsky.social @anilseth.bsky.social Barnett, Carter
Strap in!
1/n
New paper: "Large Language Models and Emergence: A Complex Systems Perspective" (D. Krakauer, J. Krakauer, M. Mitchell).
We look at claims of "emergent capabilities" & "emergent intelligence" in LLMs from the perspective of what emergence means in complexity science.
arxiv.org/pdf/2506.11135
Photo of a happy-looking Rich Sutton
Figure 1a from Lappalainen et al 2024, shows schematic diagram of their strategy to train a recurrent ANN with connectivity constrained by fly optic lobe connectome, then compare it to actual fly brain activities.
“The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.” “The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.”
The goal of neuroscience is to understand the brain. But what if we humans are just not smart enough? Neuroscience is entering the big-data era: Connectomics (brain wiring) Cell types census (brain parts catalogue) Large-scale neural activity recording methods Spatial transcriptomics (molecular expression) Should we just combine all this data with machine learning, do a big brain simulation, then go home?
today we had a nice discussion on q: 'does Sutton's Bitter Lesson apply to neuroscience?'
Sutton argued that big compute and data lead to AI systems better than any human-crafted alternatives.
Neuroscience has big data now, is it going the same way?
Eg www.nature.com/articles/s41...
Today marks a big milestone for me. I'm launching @law-zero.bsky.social, a nonprofit focusing on a new safe-by-design approach to AI that could both accelerate scientific discovery and provide a safeguard against the dangers of agentic AI.
Coming in September 2025: 16 chapters in the @springernature.com book on "The Bodily Self, Emotion, and Subjective Time: Exploring Interoception through the Contributions of A.D. (Bud) Craig". Subsequently I present all online chapters here. Today 1/16: link.springer.com/chapter/10.1...
1/2 And its out! The @arc-cogitate.bsky.social project publishes its first study today in @nature.com. Many congrats to the entire team, & esp. to Lucia Melloni, @liadmudrik.bsky.social, & @michaelpitts.bsky.social for shepherding the project so well - & over 7 years! www.nature.com/articles/s41...
𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗱𝗲𝘀𝗰𝗿𝗶𝗯𝗲 𝗰𝗮𝘂𝘀𝗮𝘁𝗶𝗼𝗻-𝗹𝗶𝗸𝗲 𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝘀𝗵𝗶𝗽𝘀 𝗶𝗻 𝗻𝗲𝘂𝗿𝗼𝘀𝗰𝗶𝗲𝗻𝗰𝗲?
As argued by John Krakauer et al. most of the time we use "filler" verbs, promissory notes that we hope to "fill with substance" at some later time.
Lemma 2 Suppose B is a CIS and h a surjective CIS homomorphism from P(Σ∗) → B. Then let G be a context-free grammar. Let φ be the cfg- morphism given by φ(N) = h(L(G,N)). Then for all nonterminals N in G, L(φ(G),φ(N)) ⊆h∗(φ(N)) Proof We will assume for clarity of exposition that the grammar is in CNF – the proof of the more general case is straightforward but less legible. Suppose we have a rule N →PQin G. Therefore L(G,N) ⊇L(G,P)L(G,Q). So h(L(G,N)) ≥h(L(G,P))◦h(L(G,Q)). Then h∗(h(L(G,N))) ≥h∗(h(L(G,P))◦h(L(G,Q))) by monotonicity of h∗. By Proposition 39 h∗(h(L(G,N))) ≥ h∗(h(L(G,P)))◦h∗(h(L(G,Q))). Similarly suppose we have a rule N →a in G. So h(L(G,N)) ≥h({a}), and h∗(h((L(G,N))) ≥h∗(h({a})). Now suppose that N′ ∗ ⇒φ(G) w; we want to show that w∈h∗(N′). We pro- ceed by induction on the length of the derivation. Base case: the production is of length 1. Suppose N′→ais a production in φ(G). Then for any production N →a in G such that N′ = φ(N), h∗(h((L(G,N))) ≥h∗(h({a})) ≥{a}. Suppose it is true for all derivations of length at most k, and let N′ ∗ ⇒φ(G) w be a derivation of length k+ 1 that starts with N′→P′Q′→uv= w, where P′ ∗ ⇒φ(G) u and Q′ ∗ ⇒φ(G) v. these last two derivations must be of length at most k. Then consider any production N → PQ in G such that φ(N) = N′ , φ(P) = P′ and φ(Q) = Q′ (of which there must be at least one). Now by the inductive hypothesis, u∈h∗(P′) and v∈h∗(Q′). So uv∈h∗(P′)h∗(Q′) = h∗(h(L(G,P)))◦h∗(h(L(G,Q))). Now since h∗(h(L(G,N))) ≥h∗(h(L(G,P)))◦ h∗(h(L(G,Q))) this means that uv ∈h∗(h(L(G,N))). So w ∈h∗(φ(N). By induction, it is therefore true for derivations of any length.
I spent so many hours working through papers like this only for the field to switch to the 'neural network go brrr' paradigm
Had the pleasure to give a lecture at Uppsala University last week 😊.
I presented our recent study on ketamine and the self www.nature.com/articles/s41... and my upcoming clinical trial on psilocybin for complicated grief
#neuroskyence
#PsychSciSky
#philsci
#psychedelics
Nature Reviews Physics
Network renormalization
www.nature.com/articles/s42...
The 2025 Neuroscience of Pure Awareness Prize has been announced.
The 2025 Computational Phenomenology of Pure Awareness Prize has been announced.
mpe-project.info/the-mpe-proj...
What's the right way to think about modularity in the brain? This devilish 😈 question is a big part of my research now, and it started with this paper with @solarpunkgabs.bsky.social, finally published after the first preprint in 2021! 🤖🧠🧪
www.nature.com/articles/s41...
New paper! Now in press at Cognition:
Experimental evidence that exerting effort increases meaning
Check out @aidanvcampbell.bsky.social's new paper. This was a real effort...and boy was it meaningful (especially now that it was accepted!). Check out the preprint here: osf.io/preprints/ps...
