Lightnews — Scholar-powered news

Sam Gershman

@gershbrain.bsky.social

10K followers 66 following 670 posts

Professor, Department of Psychology and Center for Brain Science, Harvard University https://gershmanlab.com/

gershmanlab.com

Posts Media Videos Starter Packs

Sam Gershman @gershbrain.bsky.social · 32m

"War on unicorns"
or
"War by unicorns"
?

Reposted by Sam Gershman

Ashley Thomas @ashleyjthomas.bsky.social · 1d

Interested in understanding how young humans think about social relationships? I am reading PhD applications this year! **Please note**, that Harvard now requires the GRE. More information here: www.ashleyjthomas.com/workwithme

WANT TO WORK WITH ME? | Mysite

www.ashleyjthomas.com

Reposted by Sam Gershman

Merriam-Webster @merriam-webster.com · 7d

We are thrilled to announce that our NEW Large Language Model will be released on 11.18.25.

Reposted by Sam Gershman

Mike Frank @mcxfrank.bsky.social · 8d

Ever wonder how habituation works? Here's our attempt to understand:

A stimulus-computable rational model of visual habituation in infants and adults doi.org/10.7554/eLif...

This is the thesis of two wonderful students: @anjiecao.bsky.social @galraz.bsky.social, w/ @rebeccasaxe.bsky.social

infant data from experiment 1

conceptual schema for different habituation models

title page

results from experiment 2 with adults

Sam Gershman @gershbrain.bsky.social · 9d

I suggested that play is an example where the metric maximized by adults in the ARC games (task completion) is precisely the metric minimized by children during play: the ideal imaginative play is never completed. "Success" is failure. Learning "efficiency" in this setting is irrelevant.

Sam Gershman @gershbrain.bsky.social · 9d

In the same symposium, Junyi Chu and Laura Schulz gave wonderful talks showcasing exactly how children defy Chollet's definition and the ARC operationalization. This led to an interesting discussion of whether it's possible to devise benchmarks that capture true play.

Sam Gershman @gershbrain.bsky.social · 10d

I'm more sympathetic to a "diverse intelligences" viewpoint, where different umwelts and ecological niches require different inductive biases that lead to different patterns of skill acquisition efficiency. Each is efficient over a particular distribution of skills.

Sam Gershman @gershbrain.bsky.social · 10d

What this excludes is the possibility that you could have different intelligent agents that generalize over disjoint skill sets. You could imagine agents that are really good at generalizing over skills that humans are bad at. In fact, you don't need to look far to find examples.

Sam Gershman @gershbrain.bsky.social · 10d

Chollet responded that this was "universal" intelligence, not "general" intelligence. To summarize (if I've understood correctly): general intelligence = generalizing like humans; universal intelligence [mathematically impossible] = generalizing over all skills.

Sam Gershman @gershbrain.bsky.social · 10d

I had always assumed that "general intelligence" refers to intelligent systems that show efficiency on essentially all tasks (that's the "general" part). This is why I also always assumed that (by the NFL theorems) general intelligence is impossible.

Sam Gershman @gershbrain.bsky.social · 10d

What surprised me is that Chollet defined the reference distribution of skills as precisely those skills humans are efficient at acquiring. By this definition, general intelligence is human intelligence. I find this problematic as a definition.

Sam Gershman @gershbrain.bsky.social · 10d

I asked Chollet for what skill distribution an intelligent system was expected to be efficient. To the extent that efficiency is derived from learning (i.e., it's a form of generalization), the no free lunch theorems say that no system can be equally efficient for all skills.

Sam Gershman @gershbrain.bsky.social · 10d

Chollet famously defined general intelligence as skill-acquisition efficiency. This is the idea behind the various ARC challenges: testing how efficiently artificial agents can solve novel problems. I should say that I agree with this as an important pillar of intelligence.

Sam Gershman @gershbrain.bsky.social · 10d

I was part of an interesting panel discussion yesterday at an ARC event. Maybe everybody knows this already, but I was quite surprised by how "general" intelligence was conceptualized in relation to human intelligence and the ARC benchmarks.

Sam Gershman @gershbrain.bsky.social · 10d

@arthurpr4t.bsky.social has written another tour de force, showing how efficient coding reshapes numerosity representations in parietal cortex.

bioRxiv Neuroscience @biorxiv-neursci.bsky.social · 12d

Distributed range adaptation in human parietal encoding of numbers https://www.biorxiv.org/content/10.1101/2025.09.25.675916v1

Sam Gershman @gershbrain.bsky.social · 11d

This doesn't surprise me at all, because it's basically nobody's job to defend "the highest ideals of the university." Every job optimizes some local constraint. This is why US universities are toast in the face of attacks on civil society.

Sam Gershman @gershbrain.bsky.social · 13d

Baccarelli was paid $700/hour for his work on a case against Tylenol. I will charge $0/hour to tell courts, FDA, and everyone else that correlations can arise from latent variables in the absence of a direct causal relationship.

www.thecrimson.com/article/2025...

Harvard’s Public Health Dean Was Paid $150,000 to Testify Tylenol Causes Autism | News | The Harvard Crimson

Harvard School of Public Health Dean Andrea A. Baccarelli received at least $150,000 to testify against Tylenol’s manufacturer in 2023 — two years before he published research used by the Trump admini...

www.thecrimson.com

Sam Gershman @gershbrain.bsky.social · 14d

And today in the @nytimes.com...

I suggest that papers stop "suggesting a link" so that reporters stop suggesting that papers are suggesting to readers that there might be a causal relationship. They could alternatively suggest that correlations are sometimes caused by latent variables.

Sam Gershman @gershbrain.bsky.social · 15d

My favorite ending of any paper I've written.

Sam Gershman @gershbrain.bsky.social · 15d

Waiting for the news report that latent variables exist and can explain correlations between two other variables.

Sam Gershman @gershbrain.bsky.social · 15d

Timely reminder that 'associative' language leads lay people to confuse correlation and causation, as @tomerullman.bsky.social and I showed a few years ago.
journals.plos.org/plosone/arti...

Snapshot from the BBC:
www.bbc.com/news/article...

Sam Gershman @gershbrain.bsky.social · 16d

Scientists have plenty of math envy, but at least we don't have to deal with discovering infinitely many counterexamples to our conjectures.

Quanta Magazine @quantamagazine.bsky.social · 16d

For nearly a century, mathematicians wondered whether a conjecture about knots was true. In a paper posted in June, researchers found infinitely many counterexamples, revealing a new layer of complexity in this area of math. Leila Sloman reports: www.quantamagazine.org/a-simple-way...

A Simple Way To Measure Knots Has Come Unraveled | Quanta Magazine

Two mathematicians have proved that a straightforward question — how hard is it to untie a knot? — has a complicated answer.

www.quantamagazine.org

Sam Gershman @gershbrain.bsky.social · 16d

"Get a job, neuron!"

Sam Gershman @gershbrain.bsky.social · 16d

I agree with you that labels in themselves don't matter that much. But I would strongly contend that our descriptions of data are not theory-neutral. It's precisely because people ignore this point that we get stuck with stubborn operational definitions.

Sam Gershman @gershbrain.bsky.social · 16d

Also, I'm always hearing from neuroscientists that various cognitive constructs are old-fashioned folk psychology, but I think it's important to understand the long theoretical literature behind these terms, which is very far from folk psychology.

Psychology != Folk Psychology