Bowen Zheng

YouTube video by Dwarkesh Patel Richard Sutton – Father of RL thinks LLMs are a dead end

The way Sutton himself interprets the “bitter lesson” in this interview definitely caught a lot of bitter lesson enthusiasts off guard.
LLMs not actually being an example of the bitter lesson was quite a nuance no one saw coming.

youtu.be/21EYKqUsPfg?...

So far, learning traps seem robust to social learning in our cases. Surprisingly, despite many manipulations that have tried to reduce this learning trap, the most effective has been simply being a child (see @emilyliquin.bsky.social's work on traps in children) osf.io/preprints/ps...

Ca2+ Plateau Potentials Reflect Cross-Theta Cortico-Hippocampal Input Dynamics and Acetylcholine for Rapid Formation of Efficient Place-Cell Code A central tenet of Systems Neuroscience lies in an understanding of memory and behavior through learning rules, but synaptic plasticity has rarely been shown to create functional single-neuron code in...

Interesting new data on BTSP mechanisms from my old Janelia colleague @hiallen72.bsky.social
www.biorxiv.org/content/10.1...

Neuronal signatures of successful one-shot memory in mid-level visual cortex High-capacity, one-shot visual recognition memory challenges theories of learning and neural coding because it requires rapid, robust, and durable representations. Most studies have focused on the hip...

New preprint! How can you remember an image you saw once, even after seeing thousands of them? We find a role for humble mid-level visual cortex in high-capacity, one-shot learning. doi.org/10.1101/2025.09.22.677855 🧵🧪1/

The New York Times piece today about US science is terrible and wrong—in many ways.

I could write a whole article about this, but as one example:

“To close observers, the original crisis began well before any of this…”
No. I’m a close observer of science, and this is incorrect.

This is in principle justified by Rao-Blackwell theorem? one abstracts the problem enough such that the data we do have is a suffcient statistics for the inference problem.

Can a single cell learn? Even without a brain, some microbes show simple forms of cognition. Can this basal cognition be engineered? Check our new paper with @jordiplam.bsky.social on the minimal synthetic circuits & their cognitive limits. @drmichaellevin.bsky.social www.biorxiv.org/content/10.1...

I’m not sure how useful this form is for characterizing part of the brain that does specific computation though. The heart is an important part of keep me alive to do face processing, but it does’t seem useful to say face processing -> heart is active. though it’s logically correct.

LLRX republished the blogpost www.llrx.com/2025/08/ai-s...

I wrote a Comment on neurotheory, and now you can read it!

Some thoughts on where neurotheory has and has not taken root within the neuroscience community, how it has shaped those subfields, and where we theorists might look next for fresh adventures.

www.nature.com/articles/s41...

MIT report: 95% of generative AI pilots at companies are failing There’s a stark difference in success rates between companies that purchase AI tools from vendors and those that build them internally.

MIT’s NANDA initiative found that 95% of generative AI deployments fail after interviewing 150 execs, surveying 350 workers, and analyzing 300 projects. The real “productivity gains” seem to come from layoffs and squeezing more work from fewer people not AI.

AI slop and the destruction of knowledge This week I was looking for info on what cognitive scientists mean when they speak of ‘domain-general’ cognition. I was curious, because the nuances are relevant for something I am researching at t…

AI slop and the destruction of knowledge irisvanrooijcogsci.com/2025/08/12/a...

Explosive neural networks via higher-order interactions in curved statistical manifolds - Nature Communications Higher-order interactions shape complex neural dynamics but are hard to model. Here, authors use a generalization of the maximum entropy principle to introduce a family of curved neural networks, reve...

Our paper just out in Nature Communications!
www.nature.com/articles/s41...

We introduce curved neural networks naturally introducing high-order interactions showing:
• explosive phase transitions
• enhanced memory retrieval via self-annealing
• increased memory capacity through geometric curvature

So what drives drift? We looked closely at the neurons and found that a small group of them were stable. These stable neurons were more excitable than neighboring cells, making the fate of the cells predictable.

The Lemkin Institute for Genocide Prevention is calling on every single leader in the world: DO EVERYTHING YOU CAN TO GET FOOD & WATER INTO GAZA RIGHT AWAY. Even if it takes bypassing the reports, meetings, endless conferences, parliamentary sessions, UN sessions, and all the other regular diplomatic channels that have led nowhere. Just do it. Genocide must not be allowed to continue while we all watch. We must not allow mass starvation in Gaza. We cannot wait any longer. IF YOU HAVE POWER, USE IT. HISTORY WILL DEMONSTRATE THE RECTITUDE OF YOUR ACTIONS.

DO EVERYTHING YOU CAN TO GET FOOD AND WATER IN TO GAZA.
This is from Lemkin Institute begging..... we are all begging.

Formation of an expanding memory representation in the hippocampus - Nature Neuroscience Multiday imaging of CA1 neurons during learning reveals that the representation stabilizes as the number of readily retrievable, information-rich and stable place cells increases and suggests novel me...

Really interesting results, suggesting that long-term place field stability is not from long-lasting synaptic plasticity, but is instead from an increased *probability of plasticity induction* in subsequent days.

“this is an unfair comparison because the model has not been trained on all data that has ever existed and on all future data that will be digitalized! Our foundation model is omniscient which renders the concept of generalization null!!!!”

Model mimicry limits conclusions about neural tuning and can mistakenly imply unlikely priors Nature Communications - Model mimicry limits conclusions about neural tuning and can mistakenly imply unlikely priors

Who doesn't like a good model of the brain? Yet, from simple regression to neural nets, some limitations keep popping up (e.g., overfitting) @mjwolff.bsky.social & I saw some cool but puzzling data, ran a quick analysis & found one such limitation: model mimicry. Now in #naturecommunications &🧵below

My latest Aronov lab paper is now published @Nature!

When a chickadee looks at a distant location, the same place cells activate as if it were actually there 👁️

The hippocampus encodes where the bird is looking, AND what it expects to see next -- enabling spatial reasoning from afar

bit.ly/3HvWSum

It occurred to me last night that microwaves are kinda like LLMs.
Remember when they first came out, people bought microwave cookbooks, and special vented plastic cookware, and they were going to change the way we cooked and ate forever?
Now we use them for defrosting mince, and reheating cold tea.

We’re excited about this project! We present a model of motor savings without the need for context.

Kilosort4 detects a LOT of neurons, I recorded 15k neurons in one year 🤯 Traditionally, one would curate these detected units to see if they are well isolated single neurons. This is not feasible anymore, so today let's look at three options that are out there to automate this process! 🤖👇

Technical Associate I, Kanwisher Lab MIT - Technical Associate I, Kanwisher Lab - Cambridge MA 02139

I’m hiring a full-time lab tech for two years starting May/June. Strong coding skills required, ML a plus. Our research on the human brain uses fMRI, ANNs, intracranial recording, and behavior. A great stepping stone to grad school. Apply here:
careers.peopleclick.com/careerscp/cl...
......

Synaptic plasticity rules driving representational shifting in the hippocampus - Nature Neuroscience Madar et al. report that behavioral timescale synaptic plasticity (BTSP), not spike-timing-dependent plasticity, explains heterogeneous place fields shifting in the hippocampus. The probability of BTS...

Behavioral timescale synaptic plasticity (BTSP), not Hebbian spike-timing-dependent plasticity (STDP), explains heterogenous place field shifting in the mouse hippocampus 🧠🧪

www.nature.com/articles/s41...

In contrast to the wide spread applause that this piece seems to be getting, I disagree with a lot of what is said here.

1/N

In order to understand cognition, we often recruit analogies as building blocks of theories to aid us in this quest. One such attempt, originating in folklore and alchemy, is the homunculus: a miniature human who resides in the skull and performs cognition. Perhaps surprisingly, this appears indistinguishable from the implicit proposal of many neurocognitive theories, including that of the 'cognitive map,' which proposes a representational substrate for episodic memories and navigational capacities. In such 'small cakes' cases, neurocognitive representations are assumed to be meaningful and about the world, though it is wholly unclear who is reading them, how they are interpreted, and how they come to mean what they do. We analyze the 'small cakes' problem in neurocognitive theories (including, but not limited to, the cognitive map) and find that such an approach a) causes infinite regress in the explanatory chain, requiring a human-in-the-loop to resolve, and b) results in a computationally inert account of representation, providing neither a function nor a mechanism. We caution against a 'small cakes' theoretical practice across computational cognitive modelling, neuroscience, and artificial intelligence, wherein the scientist inserts their (or other humans') cognition into models because otherwise the models neither perform as advertised, nor mean what they are purported to, without said 'cake insertion.' We argue that the solution is to tease apart explanandum and explanans for a given scientific investigation, with an eye towards avoiding van Rooij's (formal) or Ryle's (informal) infinite regresses.

Figure 1 in https://philsci-archive.pitt.edu/24834/

Box 1 in https://philsci-archive.pitt.edu/24834/

Box 2 in https://philsci-archive.pitt.edu/24834/

Tired but happy to say this is out w @andreaeyleen.bsky.social: Are Neurocognitive Representations 'Small Cakes'? philsci-archive.pitt.edu/24834/

We analyse cog neuro theories showing how vicious regress, e.g. the homunculus fallacy, is (sadly) alive and well — and importantly how to avoid it. 1/

“Torrenting from a corporate laptop doesn’t feel right”: Meta emails unsealed

A photo of Aaron Swartz (1986-2013) when he was 19.

Last month, Meta admitted to torrenting a controversial large dataset known as LibGen, which includes tens of millions of pirated books. But details around the torrenting were murky until yesterday, when Meta's unredacted emails were made public for the first time. The new evidence showed that Meta torrented "at least 81.7 terabytes of data across multiple shadow libraries through the site Anna’s Archive, including at least 35.7 terabytes of data from Z-Library and LibGen," the authors' court filing said. And "Meta also previously torrented 80.6 terabytes of data from LibGen."

Meta illegaly downloaded 80+ terabytes of books from LibGen, Anna's Archive, and Z-library to train their AI models.

In 2010, Aaron Swartz downloaded only 70 GBs of articles from JSTOR (0.0875% of Meta). Faced $1 million in fine and 35 years in jail. Took his own life in 2013.

Erika's quick-start guide to research nonprofits You do NOT need permission from a university to do research (!!)

A quick-start guide to research non-profits: erikaaldendeb.substack.com/p/erikas-qui...

Wow! I speculated a while back that OpenAI might be scanning API logs to get the hold-out questions on this dataset but now it seems the whole thing was secretly funded by OpenAI who had privileged access to the data.

Apparently I wasn't being cynical enough.

Neural signatures of model-based and model-free reinforcement learning across prefrontal cortex and striatum https://www.biorxiv.org/content/10.1101/2025.01.11.632388v1