Dan Levenstein

4/7) It's only by tying the emergence of that manifold to the architecture, loss, learning rule, and *data/world* that the network was shaped by that we arrive at any additional knowledge gained by observing the manifold.

Doing science be like en.wikipedia.org/wiki/Blind_m... Writing science be like en.wikipedia.org/wiki/Chekhov...

I've been trying to figure out how to explain this tendency of LLM writing to use terms it hasn't introduced as if you already know what it's talking about, but it's a dead giveaway that's a little more subtle than "delve" or em-dashes.

I had a feeling recently that seems to be corroborated by chats with colleagues, so I decided to share it here. A lot of systems neuroscientists driven by algorithmic-level questions are going through what looks and feels like a research depression. 🧵 (1/n) 🧠📈 #NeuroAI

Broke: all models are wrong but some are useful Woke: models are useful because they are wrong

this is an extremely cool and useful information visualization

🔥 open.spotify.com/album/3nUwuI...

I kinda get why people hate on fussy philosophical stuff, but on the other hand don't you think we live in a time where it'd be especially helpful to have rigorous conversations about things like personhood, meaning in life, the point of art, the point of pure mathematics, & operationalizing ethics?

AI went rogue and deleted my homework.

Also see this classic essay by Jeremy Gunawardena: www.molbiolcell.org/doi/10.1091/...

There’s a Ted Chiang story where AI has a parallel AI-only scientific literature. In that case, AI was doing all the real science and humans were left to pick up the scraps and try to understand what they could… Not sure if I’d prefer that or slopfest.

There are two kinds of people in the world: those who roll their eyes every time they see the phrase "load-bearing", and those who don't use Claude.

Never thought I’d be one of those “What I can’t build, I don’t understand” people, but here we are.

I’ve been thinking… now that AI is taking on groundbreaking, award-winning problems across science, what are those equivalent problems in #neuroscience? What would one of these models actually have to do to leave us all in awe? 🧠🤖

If there are any PhD hopefuls on here, I wanted to share a few thoughts on your timeline for the upcoming admissions cycle, with apps due starting mid-November. What should you be doing right now? ✔️ Generating a list of schools ✔️ Reaching out to possible recommenders 🧵!

The superposition of these two studies — both incredible — reflect a snapshot of "one big thing" systems neuroscience has yet to resolve. In mice: 1. Task-related signals are everywhere. 2. Absent large swaths of "everywhere," suprisingly sophisticated learning & behavior remains.

Gj @huihongli.bsky.social giving the first external lab poster at the @kempnerinstitute.bsky.social workshop yesterday 👏

Really enjoyed this conversation with @braininspired.bsky.social!

Came up with the term “fundable-grant-shaped object” and I’m gonna call that a win for today.

We were actually discussing this at a workshop earlier this year. In particular: The dynamics of BTSP actually seem to potentially be counter to a gradient descent mechanism, and more in line with hashing.

Super excited to be teaching a lecture on computational models of memory at this @neurosas.bsky.social course this Fall. It's not too late to apply to join us in Venice!

I gave a talk in our ISRC CN3 summer school yesterday on computational modelling of learning and memory, at the end I threw in a list of what I consider the big open questions in the the theory of learning in neural circuits. What did I miss??

If you’re interested to do an interdisciplinary co-mentored, field transition, or open ended theory/computation postdoc at WTI, please get in touch!

If you’re interested to do an interdisciplinary co-mentored, field transition, or open ended theory/computation postdoc at WTI, please get in touch!

🤤🍔

Postdocs - come talk at Yale!

Of course for the lab's first year at SfN we get hazed with the Wednesday afternoon timeslot 🥲

Sunday is for hidden gems. open.spotify.com/album/6jRvDp...

Things my wife sends me en.wikipedia.org/wiki/List_of...

Planning the first ever hippocampus GRC to overlap with cosyne was a choice 😿

I think the difficulty of understanding feedback loops underlies many big debates in science: Origin of life: "genes first" vs "metabolism first" (neither, it's a feedback loop) Development: "nature vs nurture" (neither, it's a feedback loop) Neuroscience: "brain vs behaviour" (you guessed it!)

Just discovered this gem of a final paragraph in Rumelhart, Hinton, and Williams 1986. Same energy.

This is why I’ve gone from “never want to touch a spiking network again” to “excited to do the kinds of things I always wanted to do with spiking networks” in the span of 5 years.

They’re Made out of Meat

So asking "is the brain a computer" is not helpful. Symbolic computation is sometimes a useful way of thinking about a brain feature, and sometimes not. And that's it. We really can stop talking about it now. I'm going to regret this. I know it.

A good experiment-theory collaboration takes _________. Wrong answers only.

(apart from the important point that if we let AI do all our thinking, we are doomed) generally speaking this is why I have come to view presentations and paper writing as *doing science*, just as much as lab work or running computer simulations or whatever. Communication and discussion are central

Your plots are the plot of your paper’s story

Really excited about this work with the awesome Will de Cothi & @shipleysj.bsky.social, out in Nature Reviews Neuroscience. We propose acetylcholine is an analogue of dopamine — tracking not reward prediction errors but state-transition prediction errors. https://doi.org/10.1038/s41583-026-01058-w

These models were not built to explain the brain. But treating them as model organisms that we can perturb and evolve will move us closer to truly understanding the brain’s mechanics, writes @mtoneva.bsky.social. #neuroskyence www.thetransmitter.org/artificial-i...