iiinova — where your digital self keeps evolving

Right, I agree that the Pandey transformer work is convincing.

020

Yes, this is a legitimate question, but it can’t be answered by existence proofs. You need models that directly reproduce the relevant data, including the developmental milestones. Most of the “digital twins” don’t do this.

130

If the empirical observation is that a competence is present at birth, and the neural networks (or any kind of learning system) require experience to achieve the same competence, then I don’t see how the data are consistent with such a mechanism.

031

You don’t need that assumption to make a nativist argument on the basis of the chick data. All that matters is that certain competences are present at birth for the claim that they’re innate. The neural network results aren’t relevant to this claim.

120

I don’t see how. The neural network results show only that something is possible; they can only disconfirm necessity claims. The claims about the chick data are not necessity claims.

110

I also disagree that results from training artificial models can tell us anything positive about how real minds work, beyond sufficiency for certain competences. You can't go from existence proof to confirmation that it actually works that way in real minds.

2171

Sam Gershman @gershbrain.bsky.social · 12/09/2026

This is interesting, but I'm having trouble with parts of the argument. It says that the same phenomena support both radical positions, but it doesn't seem to be the same phenomena. The data from newborn chicks supports nativism, but the "data" from ML support empiricism. These are different.

1122

It depends... sometimes dramatic reductions in cortex early during development (e.g., from hydrocephalus) can be compensated. We still don't understand why. www.science.org/doi/pdf/10.1...

science.org

Is Your Brain Really Necessary?

181

Reposted by Sam Gershman

Ryan Truong @heyodogo.bsky.social · 10/09/2026

We present PlayTrain, a new Reinforcement Learning (RL) framework that reimagines how we build video-game environments for RL by harnessing LLMs’ ability to write JavaScript games. PlayTrain turns single-file JS browser games into RL environments ready for efficient training.

1164

Sam Gershman @gershbrain.bsky.social · 09/09/2026

Pay attention to Arthur: he’s breathing fresh life into psychophysics.

0271

Reposted by Sam Gershman

Arthur Prat @arthurpr4t.bsky.social · 09/09/2026

We derived six laws of psychophysics from a single efficient-coding equation. The laws include Weber's law; scaling laws in visual working memory (which had been noted before, but with unclear theoretical justification); Wei & Stocker's law of human perception; arthurprat.com/pdfs/Prat-Ca...

arthurprat.com

49628

Reposted by Sam Gershman

Momchil Tomov @momchiltomov.bsky.social · 02/09/2026

Excited to share that our paper on explainable AI for autonomous driving is out today in @nature.com! Super proud to have been part of this amazing collaboration between Motional and MIT! Details in the 🧵

2259

Reposted by Sam Gershman

Naomi Saphra @nsaphra.bsky.social · 01/09/2026

Yes, some people are worse at writing than an LLM. Those people are also usually incoherent thinkers. If they learned to write better, they would get better at thinking.

2726

Sam Gershman @gershbrain.bsky.social · 01/09/2026

If this makes you mad, submit your cognitive science research to @openmindjournal.bsky.social No APCs, no subscription fees. It's funded by university libraries, at low cost.

18934

Sam Gershman @gershbrain.bsky.social · 23/08/2026

I think what you're asking about is what we looked at in an earlier paper on numerosity estimation: gershmanlab.com/pubs/PratCar... The story is somewhat different there because we frame it in terms of Bayesian inference rather than efficient coding, but the manipulations are very similar.

120

Sam Gershman @gershbrain.bsky.social · 23/08/2026

Nathaniel, did you get a concussion? I never thought I'd hear you utter the words "replacing bacon".

141

Sam Gershman @gershbrain.bsky.social · 22/08/2026

Yes, actually @rbhui.bsky.social and I have an earlier paper showing how efficient coding is related to decision by sampling: gershmanlab.com/pubs/BhuiGer...

020

Sam Gershman @gershbrain.bsky.social · 18/08/2026

@arthurpr4t.bsky.social has a new preprint with important results on a famous psychophysical law (Weber's law). It isn't, in fact, a law, because it can be broken. A more fundamental principle (efficient coding) shows when and why Weber's law holds true. www.biorxiv.org/content/10.6...

biorxiv.org

Efficient coding makes and breaks Weber's law

Weber's law is a rare quantitative regularity in psychology, yet its origins remain debated. Here we provide causal evidence that it arises from the more fundamental principle of efficient coding. Thi...

413041

Reposted by Sam Gershman

hombrecerebro.bsky.social @hombrecerebro.bsky.social · 10/08/2026

Harvard Neurobiology + Kempner are hiring a tenure-track Assistant Professor in computational neuroscience. The search is broad; researchers combining computational and experimental work (including nonhuman-primate research) should apply! academicpositions.harvard.edu/postings/16615

academicpositions.harvard.edu

Assistant Professor of Neurobiology

The Department of Neurobiology at Harvard Medical School (HMS) and the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University invite applications for a tenure tra...

04735

This is a good time to remind everybody that if you communicate with university PR or journalists about correlational research, tell them that language like "linked to", "associated with" etc. is *not* neutral; readers will draw causal implicatures even if the work does not establish causation.

0121

There was also interesting related work at CogSci from @meanwhileina.bsky.social and collaborators on the downstream consequences of causal implicatures (memory, communication, belief updating): escholarship.org/content/qt8q...

160

The key idea is that (in English) predicates which take the subject as cause are more frequent and syntactically simpler, thus favored by a cost-sensitive pragmatic reasoner.

120

arXiv cs.HC Human-Computer Interaction @cshc-bot.bsky.social · 30/07/2026

Cool work by @amirzur.bsky.social, @rdhawkins.bsky.social and their collaborators replicating our finding on causal implicatures from correlational language, and showing that this can arise from rational pragmatic inference: escholarship.org/content/qt60...

1113

Reposted by Sam Gershman

Lance Ying, William Van Uitert, Tan Zhi-Xuan, Joshua B. Tenenbaum, Samuel J. Gershman: Pragmatic Reasoning in Design arxiv.org/abs/2607.26322 arxiv.org/pdf/2607.26322 arxiv.org/html/2607.26322

022

Reposted by Sam Gershman

arXiv cs.MA Multiagent Systems @csma-bot.bsky.social · 31/07/2026

Lance Ying, Ryan Truong, Joshua B. Tenenbaum, Samuel J. Gershman: Using Theory of Mind to Arbitrate between Social and Non-social Learning arxiv.org/abs/2607.28601 arxiv.org/pdf/2607.28601 arxiv.org/html/2607.28601

052

Sam Gershman @gershbrain.bsky.social · 28/07/2026

Sometimes my kids are so out to lunch they literally forget to eat the lunch I packed them.

080

Reposted by Sam Gershman

David Colaço @davidcolaco.bsky.social · 27/07/2026

For those interested in memory transfer (the topic of my book): #16mmfilmography (avgeeks.bsky.social) hosts The Way Out Men, a documentary that features James V. McConnell studying his cannibal worms (featured at 9:45): 16mmfilmography.org/film/the-way... #hps #philsky

16mmfilmography.org

The Way Out Men (Part One) — 16mm Filmography

192

Sam Gershman @gershbrain.bsky.social · 23/07/2026

I'm really excited about this work. In addition to a new genome assembly for Stentor, we show through transcriptomics that a pathway for long-term depression (in neurons) is involved in mechanosensory habituation... despite the cell having no synapses!

0281

There is a long history of carefully designing experiments with many controls to make strong claims about specific forms of selectivity in the brain. There are trade-offs (low coverage in the space of variables, so you miss things), which is why you need many experiments to make this work.

1192

As a matter of experimental design, if you want to draw very broad conclusions, your stimuli and task need to be correspondingly broad. This is not the case for the IBL task, and I doubt any single task would be adequate. I worry that the IBL approach is not suitable for such broad questions.

2252

The authors acknowledge this concern at the end of the Discussion, raising the point that maybe clustering only arises in more complex tasks. I think the more important point is not about task-specific functional organization but rather about the questions that can be answered with a single task.

1151