This paper was heavily influenced by "Generating meaning: active inference and the scope and limits of passive AI" by Pezzulo, , Parr, Cisek, Friston
Highly reccomended to frame our work! doi.org/10.1016/j.ti...
Written with Caleb Weinreb. Thanks to Felix Binder and Mo Osman for their contributions.
The purpose of our paper is to extend questions about AGI beyond “how smart is the model?” into “where does control live?”
That choice shapes capability, robustness, oversight, and safety. Our paper tries to name that choice, and explain why it matters now.
That may be the biggest shift underway in AI.
Prediction-trained systems inherit human traces. Action-trained systems can discover strategies humans never wrote down. AlphaGo did this in Go.
In open-ended domains, this dynamic will be much harder to audit, predict, & align
LLMs used to be trained predominantly as autoregressive models with some RLHF to fine tune them.
Now it’s switched: prediction is used to bootstrap linguistic competence, but the main player is reinforcement learning on action.
Frontier systems are already moving along this spectrum: they are building up Cartesian agents which are more deeply integrating LLMs with action.
In the paper, we use this to distinguish 3 paths forward:
Boxed cognition: keep systems useful, but inside human control loops
Cartesian agents: today’s dominant design
Integrated agents: learn more of memory, arbitration, and control end-to-end
(imagine if your motor cortex could only communicate with your basal ganglia or hippocampus through words!)
This is also where artificial and human agency diverge.
Brains do not separate perception, action, memory, and decision making into a predictor + runtime stitched together by words. Those systems are integrated and coevolved with continuous, high-dimensional, feedback loops
But the cut is also a bottleneck.
The same model can look reliable or flaky depending on prompt format, memory serialization, tool schemas, routing logic, or retry policies.
That boundary, where memory, action, and control get pushed outside the model, is what we call **the Cartesian cut**.
It’s a brilliant hack: you can bootstrap from pretrained LLMs, plug in modular tools, monitor trajectories, and add guardrails at the interface.
The agent is not a unified thing, like people are.
Core pieces of control are pushed outside the model: memory is offloaded, actions are mediated by tools, stopping and retrying are handled by a runtime, success is defined externally
The model only ever sees this through text
Why it works: LLMs learn by predicting words in human text, which reflects how people perform tasks.
Once you bolt that predictor onto tools, prediction becomes control: searching, executing code, editing files, taking actions in the world.
Figure contrasting biological intelligence with cartesian agents, where cartestian agents form an inverted pyramid of control in biological cognition.
Today’s AI agents share the same basic architecture: a thinking part (the LLM) is coupled to an acting part (an engineered runtime + tools) through a narrow symbolic interface (e.g. json & text).
In our new paper, we call this split *Cartesian agency*.
arxiv.org/abs/2604.07745
Very cool mason! Thanks for sharing
🧠🌟🐭 Excited to share some of my postdoc work on the evolution of dexterity!
We compared deer mice evolved in forest vs prairie habitats. We found that forest mice have:
(1) more corticospinal neurons (CSNs)
(2) better hand dexterity
(3) more dexterous climbing, which is linked to CSN number🧵
Reposting this open postdoc position with more detailed specifications. Couldn't find the right fit in the first round—not due to lack of strong candidates, but because we're looking for someone whose expertise matches a specific project. More details in the job ad. Apply!
tinyurl.com/2uskxyrp
Me with my poster on the greylag goose vocal repertoire and how data representation types influence the predictions of unsupervised methods
Thank you to everyone at #ibac2025 for this amazing conference. I met so many great people and learned a lot about so many interesting projects. Excited to see you all again some time! And proud that my poster got the student poster price ☺️ stay tuned for the publication!
The psych job market may not be dead... but it is gravely injured 😬 So far it's looking like the Trump administration's attacks on higher ed/research are going to have more than 2x the impact on the job market as the covid-19 pandemic. #psychjobs #neurojobs #academicjobs
⭐ Long-term support: Actively maintained for 6+ years, 1700+ stars on GitHub, hundreds of citations (even without a paper!)
🐦 We also release Birdsong NOIZEUS, a new benchmark for bioacoustic denoising
✅ Domain-general: strong baseline when ML models/data aren’t available
🎛️ Stationary & non-stationary variants
⚡ GPU-accelerated for real-time and high-throughput use
🧪 Validated across many domains
New paper out today with Asaf Zorea : "Domain-general noise reduction for time-series signals with Noisereduce" (open access)!
We present Noisereduce, a lightweight Python library for denoising signals.
Read the paper: doi.org/10.1038/s415...
Code (library): github.com/timsainb/noi...
🚨Very happy that my PhD work is now out in @nature.com!
We discovered that evolution, by acting in the midbrain, shifted the threshold to escape in Peromyscus mice, to fine-tune defensive strategies in different environments
www.nature.com/articles/s41...
This was a truly collaborative effort! 🧵⬇️
Graphical abstract for "Vocal communication is seasonal in social groups of wild, free-living house mice." The abstract has, from top to bottom, a title, four middle image panels, and two bottom text panels. Image title: "Vocal communication in social groups of wild-free living house mice" Middle image panels from left to right: (1) An aerial snap shot of the region where the study site is located, an agricultural landscape in rural Switzerland. (2) An image of the study site, a small barn in the forest inhabited by mice. (3) An image of a radio frequency identification (RFID) box used to track mouse social interactions. A mouse is entering the box from the left while another sits outside. (4) A spectrogram showing example vocalizations - one low frequency squeak and one ultrasonic call - recorded from an RFID box. Bottom panels: Left: Data Collection - 10 years of RFID-based tracking data (from 6,946 mice) - 15 months of acoustic monitoring (totaling 6,594 hours) - Machine learning for vocal detection and labeling (CNN) Right: Key Findings - Vocalization is seasonal (most in spring and summer) - Vocalization is associated with the presence of pups - Vocalization is correlated with social group dynamics
Very happy to share the latest from my postdoc!
10 yrs of mouse social networks + 1.25 yrs of acoustic data ➡️ insight into vocalization & sociality in a wild population of your favorite lab model 🐁
paper: bit.ly/4n93yyD
data: bit.ly/4lfFBEk
code: bit.ly/4kNnMwx
#bioacoustics #neuroskyence
1/8
https://www.sarasotamagazine.com/travel-and-outdoors/2020/08/dolphin-watching-tour-sarasota
1/8 Decoding Dolphin Communication
After studying 313 dolphins (in Sarasota, Florida) for over 40 years and across six generations, a catalog of their vocalizations has been produced.
These vocalizations are more complex than expected.
(preprint) www.biorxiv.org/content/10.1...
Our latest superb starling work in @nature.com. We observe long-term reciprocal helping relationships, and suggest reciprocity is an underappreciated mechanism promoting the stability of cooperatively breeding societies. Led by Alexis Earl and @gerrycarter.bsky.social
www.nature.com/articles/s41...
Very cool @danielpollak.bsky.social @healeylab.bsky.social - this aligns with what I found in the starling paper we just published using a different context paradigm (sequence integration). NCM (among others) also does not show context modulation but sharpens acuity/attn with expectation
