Brian DePasquale
@briandepasquale.bsky.social
science. https://briandepasquale.github.io
Our review on superposition in AI & brains is published in @nature.com #Machine #Intelligence 🌐 Neural networks encode concepts in superposition. We can recover them in 3 steps: identifiability, compressed sensing, interpretability. W/ @david-klindt.bsky.social O'Neil Reizinger & Maurer 🌟
AI may soon enable researchers to generate high-quality science at a previously unimaginable speed. The effects for researchers will be as disruptive as industrial mass production was for artisan manufacturers, writes @kenneth-harris.bsky.social. #neuroskyence www.thetransmitter.org/artificial-i...
Mass-produced science is coming. What happens to scientists?
Artificial intelligence may soon enable researchers to generate high-quality science at a previously unimaginable speed.
thetransmitter.org
I'm so disillusioned, I almost don't even care who gets the FIFA Peace Prize next year
Friday: FIFA announces, after further consideration, the game was a 4-6 win by USMNT, knocking Belgium out of the tournament and paving the way for the U.S. team to face Spain instead. 🙃
At Cosyne, Chris Olah from Anthropic told us that neuroscientists had “3-4 years left to make a contribution” Why was not clear. Perhaps he was hinting at the below graph accompanying today’s Claude Science launch. Apparently we’ll all be obsolete soon.
New paper: We recast automated scientific model discovery with LLMs as Bayesian inference! LLMs write code and carry domain knowledge, great for proposing models. The key idea: discovery is inference, not just generation. What distribution of models explains the data? 🧵 arxiv.org/abs/2602.18266
Thanks to everyone for contributing their comments! I really enjoyed seeing the diversity of takes — an important conversation at a critical moment!
The Transmitter @thetransmitter.bsky.social · 2mo agoHear from Larry Abbott, @neurograce.bsky.social, @gjorjulijana.bsky.social, @neural-reckoning.org, @talmo.bsky.social, @neurokim.bsky.social and @tonyzador.bsky.social on what that tectonic change will bring to the field.
Hear from Larry Abbott, @neurograce.bsky.social, @gjorjulijana.bsky.social, @neural-reckoning.org, @talmo.bsky.social, @neurokim.bsky.social and @tonyzador.bsky.social on what that tectonic change will bring to the field.
How can AI change theoretical neuroscience? Technologies separate us from the mechanics of computation. Are we liberated to focus on the big picture or abandoning the substrate of thought? Born from an ongoing experiment in my lab about balancing these technologies for research and training
The Transmitter @thetransmitter.bsky.social · 2mo agoAgentic coding makes it possible to specify a neuroscience model in hours instead of months, writes @briandepasquale.bsky.social. The field risks becoming prolific but shallow—generating models faster than we can generate insights. #neuroskyence www.thetransmitter.org/the-big-pict...
I had no idea men played soccer, too! Good for them! Go boy soccer!! ⚽️
🚨 New preprint from the lab! 🚨 We introduce BiXformer, a bidirectional cross-attention transformer for disentangling inter-regional neural dynamics. Recordings from multiple regions superimpose feedforward & feedback signals 🔁, offset in time, making it difficult to tease these signals apart. 🧠 🔁
bioRxiv Neuroscience@biorxiv-neursci.bsky.social · 2mo agoBiXformer: A Bidirectional Cross Attention Transformer for Disentangling Inter-Regional Neural Dynamics https://www.biorxiv.org/content/10.64898/2026.06.05.730511v1
Computational models are a key part of science but discovering new ones is hard! DataDIVER discovers concise models from data, which surface new mechanistic ideas and clear predictions for future experiments From Google Deepmind Neuroscience Lab + collaborators www.biorxiv.org/content/10.6...
FY27 Focused Research Programs Spotlight: "AI for Characterizing and Designing Biomolecular Interactions" Learn more about this FRP: spr.ly/63323B8Y3sz
🚨New Preprint🚨 from amazing PhD student @ryguy.io! We augmented the classic DDM to account for state-dependent changes in decision strategy and found a clear improvement. Applied to a novel 24 hour dataset we found support for circadian influence on decisions and more! Check out thread and paper!
Ryan Senne (he/him)@ryguy.io · 2mo agoMove over, GLM-HMM—there’s a new hidden Markov model in town. In our new preprint, we introduce the DDM-HMM: a model that jointly infers latent decision states from reaction times and choices. Using 24-hour behavioral data from rats, we found that evidence accumulation changes with time of day!
National Academy of Sciences experts denounce Trump’s NSF board purge www.scientificamerican.com/article/nati...
National Academy of Sciences experts denounce Trump’s NSF board purge
In an open letter, thousands of researchers criticized the White House’s firing of the research agency’s board as ‘an alarming attack’ on U.S. science
scientificamerican.com
1/7 New paper accepted as ICML spotlight arxiv.org/abs/2605.03517! We unify self-supervised learning (SSL) algorithms (e.g., contrastive, VICReg, stopgrad) via latent distribution matching (LDM), which matches an induced latent distribution to an explicit latent model.
Excited to see our paper describing a stereotyped receptor map for smell (led by @davidhbrann.bsky.social ) out in the world! See below for short thread with open access link to paper www.nytimes.com/2026/04/29/s...
Scientists Unveil ‘Long Lost’ Map for Smell
nytimes.com
maybe besides the point but didnt Lebos et al show that good models can recover interpretable models of orbital mechanics like 5 years ago? arxiv.org/abs/2202.02306
Rediscovering orbital mechanics with machine learning
We present an approach for using machine learning to automatically discover the governing equations and hidden properties of real physical systems from observations. We train a "graph neural network" ...
arxiv.org
Scrap all of this wasted time and effort and just award grants randomly. The article briefly suggests and dismisses this but I think it would solve a lot of problems with science funding. If you're overwhelmed with good submissions, you're already choosing randomly and just not admitting it.
Elizabeth Gibney@lizziegibney.bsky.social · 3mo agoHave grants become more competitive lately? 🧪 At least some of the uptick in submissions (that we've reported on @nature.com) seems driven by AI-assisted proposal writing -- funders, like journals, are flooded Great comment piece by @jameswilsdon.bsky.social & co: www.nature.com/articles/d41...
Film you've watched more than six times with a gif. Hard mode: no Star Trek, Star Wars, LOTR, DC, Marvel, Disney, Pixar, Spielberg, Cameron, Wachowskis, or Ghibli:
a close up of a man with the word rocky on the bottom right
ALT: a close up of a man with the word rocky on the bottom right
media.tenor.com
BWJones@bwjones.bsky.social · 3mo agoFilm you've watched more than six times with a gif. Hard mode: no Star Trek, Star Wars, LOTR, DC, Marvel, Disney, Pixar, Spielberg, Cameron, Wachowskis, or Ghibli:
One of the hardest + most important things to learn as a creative. Stop waiting to make your “opus” — keep going, get better
MJ O’Neill 🏳️⚧️@mjoneillchw.bsky.social · 4mo agoI think the greatest trick for getting good at creative work i know is - finish making as many mediocre things as possible. Just finish stuff, again and again. Share it or don’t, but put a bow on it, put it aside, make the next thing, repeat.
New OpenFold3 preview out! (OF3p2) It closes the gap to AlphaFold3 for most modalities. Most critically, we're releasing everything, including training sets & configs, making OF3p2 the only current AF3-based model that is functionally trainable & reproducible from scratch🧵1/9
Johnny Knoxville says Ed Witten should be more careful when talking about physics.
Delighted to share our discoveries about one of the brain's neurotransmitter systems: www.biorxiv.org/content/10.6... Together with colleagues at the @alleninstitute.org, we have learned a lot about a tiny cluster of neurons in the brainstem locus coeruleus (LC) that releases norepinephrine (NE). 1
biorxiv.org
Interesting piece claiming that the asymmetry between writing acceleration and review acceleration is breaking science. Essentially Brandolini’s law (The amount of energy needed to refute bs is an order of magnitude bigger than that needed to produce it) applied to science publishing >
Publish and Perish: How AI-Accelerated Writing Without Proportional Verification Investment Degrades Scientific Knowledge
Artificial intelligence tools are accelerating manuscript production far faster than peer review capacity can expand. Applying the theory of constraints from manufacturing science, we formalize this a...
arxiv.org
Fix for next 11th hour grant push — Dr Robby tells me to Dr the f**k up 🔥
This is satisfyingly niche. Shots fired by the chemical reaction networks community. arxiv.org/abs/2603.12060
Chemical Reaction Networks Learn Better than Spiking Neural Networks
We mathematically prove that chemical reaction networks without hidden layers can solve tasks for which spiking neural networks require hidden layers. Our proof uses the deterministic mass-action kine...
arxiv.org
Among the frustrating aspects of the funding situation in the U.S. is that the strategic response is to apply for more grants, with the result that the equilibrium success rate is pushed lower (up to a point). Scientists get less while spending more time writing grants rather than doing science.
