iiinova — where your digital self keeps evolving

Reposted by Roman Feiman

David Levari @davidlevari.bsky.social · 08/10/2026

I'm hoping to recruit a new PhD student to join my lab at Brown (@browncopsy.bsky.social) this cycle. Applications are due Nov 15. Please apply! To read about our whole deal: levarilab.com

levarilab.com

Levari Lab

The Levari Lab at Brown University uses social and cognitive psychology to study the behavioral science of performance evaluation, improvement, and learning.

11810

Reposted by Roman Feiman

Urvi Maheshwari ツ @urvi.bsky.social · 06/10/2026

New w/ @drbarner.bsky.social in Cognitive Science! We explored how kids learn the “later-greater” principle — the idea that numbers that come later in a count are greater in magnitude. Turns out this simple property governing the count system doesn't come easy! dx.doi.org/10.1111/cogs... 1/6

dx.doi.org

One, Two, Three, Four… What's More? Successor Relations, not Mappings to Approximate Magnitudes, Support Learning the Later‐Greater Principle

As we count up, number words that occur later in our count denote greater quantities. Here, we tested whether children's understanding of this “later-greater” principle originates from associating nu...

15919

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

That’s surprising and cool, and it’s some evidence against that capacity being innate in chicks (again, not conclusive, but good evidence!) But it doesn‘t make the two explanations compatible.

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I just wanna add one (hopefully new) point. I think the cool thing about these studies is they suggest the experience we thought was needed to merely trigger an innate capacity for object representation (enough exposure to an object to imprint on it) may be enough to learn that capacity NN-style

101

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I don’t think this is “Without learning” in Spelke’s sense — if that “experience-based tuning” is required, it’s just what she would call learning. And it’s not like it’s surprising the experience should be related -- it’s hours of different views of the same object!

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Either the prior exposure (limited view angles, whatever) is necessary for the subsequent capacity, in which case it’s definitionally not at-first-encounter. Or else it isn’t necessary, in which case the capacity may be innate in an NN architecture. There’s no case of compatibility

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

There’s no in-principle compatibility — either X arises from NN-style learning in chicks (so X is not innate) or it doesn’t (so maybe it is, or maybe it’s learned some other way). I don’t see a case that’s “compatible” with both.

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I think NN studies can show something is learnable by an NN and that there are (sometimes unexpected) generalizations from learning about one thing to another. They can also show that something in innate in an NN. All are interesting, none are a paradox or problem for a coherent notion of "innate"

110

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I think we might be recapitulating Samuels here (note he tackled these issues ~20 years before these twin studies)

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

1) I like Samuels's version more, for the reasons he gives. 2) This evidence is really not at first encounter! There's a lot more to say about imprinting vs. object representations, but I think I'm reaching the limit of what bsky discourse can resolve :(

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I think it shows imprinting might not be innate in chicks (if chicks' imprinting proves more robust to corrupted input than NNs' learning, that's counterevidence). Anyway, that would be interesting! What it wouldn't be is a paradox, or evidence "innate" is incoherent or polysemous.

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Again, Justin's own preprint suggests chicks' object representations are remarkably robust to disruptive evidence. I don't know of a digital twin comparison here, but I can't imagine NNs would be so insensitive arxiv.org/abs/2402.14641

arxiv.org

Object permanence in newborn chicks is robust against opposing evidence

Newborn animals have advanced perceptual skills at birth, but the nature of this initial knowledge is unknown. Is initial knowledge flexible, continuously adapting to the statistics of experience? Or ...

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Being shown an object after seeing 4 hours of other views of the same object is hardly a "first encounter". In NNs, this really suggests transfer learning from that input to other views of the same object. Maybe chicks are transfer-learning too! That'd be cool and surprising, but it's not a paradox.

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

A that point the argument is about what generalization we should expect, which input is relevant to which representation, or where we make type-token distinctions (if a chick generalizes from X1 to some minimally different X2, X2 isn't a test of a chick "at first encounter")

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Is what you have in mind that the NN was first exposed to Y, tuned its parameters on Y, and then applied what it learned (i.e. generalized) to its first encounter with X? That's cool evidence of (maybe unexpected) generalization from Y to X, but it wouldn't be "at first encounter" in Spelke's sense

200

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I really don't understand what you have in mind... There's no sense of not-learning here. "Tuning parameters" is just learning, which would presuppose that before it 'tuned its parameters' it didn't have the same capacities as after. At-first-encounter means no encounter-relevant parameter-tuning

200

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Could you point me to the stuff that tests NNs at first encounter?

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Moreover, if anyone did find X-at-first-encounter in NNs, there still wouldn't be any paradox. It would just show that an NN architecture innately processes X, with no learning needed.

210

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

I don't know of any digital twin study that tests either models or chicks "at first encounter", including all the ones you cite. They all include a training phase for the NNs (often with some kind of degraded/impossible input, which raises separate Qs about disruption as opposed to learning).

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

(There are tricky caveats about what curves maturation might predict, what counts as 'evidence', and whether some evidence can have an effect by disrupting an otherwise innate capacity rather than by learning... but I don't think any of these are at issue in your paper)

000

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Sorry, I'm not following. I'm not assuming any particular learning curve or classic associationism, just differentiating between there being any evidence-dependent learning curve, or not. Would like to understand what you mean though! Maybe email is better?

200

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

In those cases, an NN's learning trajectory is evidence of learnability, while the chick's lack of learning trajectory would be evidence against actual learning (i.e. evidence that X is innate). There's no paradox.

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

But in some cases there won't be a learning trajectory at all. That's what, e.g. Spelke has in mind when she says X is innate = "if it is present and functional on the infant’s first effective informational encounters with the entities to which it applies".

100

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Digital twin studies are not a gold standard. If there's an evidence-dependent learning trajectory of X in chicks, then that already shows X isn't innate, without the twin. If the NN twin matches that trajectory with the same input, that further shows learning plausibly happens via that kind of NN.

110

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Here’s older work I find useful on how to think about “innate”: www.cambridge.org/core/journal... sas-space.sas.ac.uk/942/1/R_Samu...

cambridge.org

Is nativism sufficient?[*] | Journal of Child Language | Cambridge Core

Is nativism sufficient?[*] - Volume 21 Issue 1

120

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

In linguistics the distinction between learnability and learning goes back to the 50s and is constantly made. Eg. Poverty-of-stimulus arguments are about what children can actually do with the input they actually get, not any possible learner with any possible input.

120

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Again, maybe different communities use it differently, but in my world it refers to denial in-fact, not in-principle. Since in-principle would entail in-fact, people sometimes talk about the former, but it’s as a route to the latter. I see @gershbrain.bsky.social made the same point in his thread.

230

Roman Feiman @romanfeiman.bsky.social · 28/09/2026

Maybe we're just in different worlds, but in mine "nativism" is an empirical question. It's not about in-principle unlearnability, it's about in-fact not being learned by a certain kind of mind. It's the mirror image of "learnable" vs. "learned".

140

Reposted by Roman Feiman

Gabor Brody @gaborbrody.bsky.social · 21/09/2026

I have a new preprint with Athulya Aravind. In it, we ask how can children learn the presuppositional meaning component of a word (e.g. "both" presupposing a |2| domain) Our claim... ...is that they can't. Presuppositional meanings are unlearnable. Preprint: lingbuzz.net/lingbuzz/010...

1206

Reposted by Roman Feiman

Mark Histed @markhisted.org · 20/09/2026

“The arrangement would give Vought, the director of the White House Office of Management and Budget, a direct role in determining how NIH distributes its $48 billion in congressionally-appropriated money.” Bypassing the many checks and balances at NIH on grant funding decisions.

8271193

Reposted by Roman Feiman

Stand Up for Science! @standupforscience.net · 17/09/2026

The Trump admin has decimated American science budgets, taking away life saving opportunities, and the potential for groundbreaking discoveries. This is how you destroy American innovation. apnews.com/article/trum...

apnews.com

Trump administration has cut or frozen $177 billion in grants across every state, analysis shows

An analysis finds that the Trump administration has cut or frozen up to $177 billion in federal grants since the president took office for his second term.

19621305

Reposted by Roman Feiman

Jeremy Berg @jeremymberg.bsky.social · 16/09/2026

Developing a plan that captures the best aspects of what NSF and NIH were while addressing some long-standing issues is what the effort at the @standupforscience.net policy shop is developing.

26318

Reposted by Roman Feiman

Harrison Ritz @hritz.bsky.social · 14/09/2026

Social Psychology: US: 23 CA: 2 Cognitive Psychology US: 36 CA: 1 (* no grad program) Same story for the past three years: ~1 Canadian job per year for cognitive psych/neuro. Amazing arbitrage opportunity for Canadian U15 schools to start hiring cog again!

0102

Reposted by Roman Feiman

Michael Clemens @mclem.org · 12/09/2026

The US President's new proposal to tax international students has been approved by the White House for release this week. It's an unprecedented $100,000 tax on int'l grads who stay to work after graduation on Optional Practical Training (most grads do), a crucial bridge to remaining by any channel.

21434715

Reposted by Roman Feiman

Michael Clemens @mclem.org · 13/09/2026

This will devastate our research+innovation capacity, bedrock of the 21st century economy. 35% of *all* employed people in the US with STEM PhD degrees are both foreign-born *and* graduated from a US university. —> nap.nationalacademies.org/resource/292... Shutting them out is extremist folly.

1829224

Reposted by Roman Feiman

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

Roman Feiman @romanfeiman.bsky.social · 30/08/2026

I don't think enough people know how hostile the legal US immigration system is (and already was, pre-Trump)

0121

Reposted by Roman Feiman

Sanjay Srivastava @sanjaysrivastava.com · 28/08/2026

The U.S. is a massive funder of science. Canada does not come anywhere close. Stories about a handful of scientists get attention, but rather than "brain drain," the likely outcome if the U.S. continues its path is that there will simply be fewer scientists and less science getting done in the world

1117156

Roman Feiman @romanfeiman.bsky.social · 26/08/2026

I just found this — I’ll delete my quote!

000

Roman Feiman @romanfeiman.bsky.social · 22/08/2026

You could say "W is being used to talk about X if the sense of W that's being used is the one that's relevant to X", but that'd be circular. The way you'd know W has senses or which sense is relevant to X is just whenever using W does in fact bear on X.

010

Roman Feiman @romanfeiman.bsky.social · 22/08/2026

Maybe to put Gabor's point another way: they also might not be if W is "monosemous", because a single sense of W might or might not apply to X in any context C' != C. So positing that W is polysemous doesn't help with where it applies to X.

120

Roman Feiman @romanfeiman.bsky.social · 20/08/2026

Heinous. there should be serious investigations and trials after the Trump administration is over. Every future administration has to get the message that it can’t do stuff like this.

062

Reposted by Roman Feiman

Dan Garisto @dangaristo.bsky.social · 20/08/2026

Scoop: For nearly a year, NSF has struggled to make new grants. Now we know that it will not catch up. In fact, NSF will award just 6,100 new grants—a 30% y/o/y decline, itself down from a previous avg. of 11,000+. In fact, the last time NSF awarded so few grants was in the early '80s.

nature.com

Exclusive: NSF set to issue lowest number of new grants in four decades

The US science-funding agency is withholding $1 billion of its budget so that the money can go to a special White House project.

14760478

Reposted by Roman Feiman

Dan Garisto @dangaristo.bsky.social · 20/08/2026

Finally, this has gone a bit under the radar, so I want to highlight it: The social, behavioral and economic sciences directorate (SBE) has simply ceased to deliver grants to many of its subareas. Any topic not specifically OK'd by the Trump administration is receiving zero new awards.

All of the NSF’s directorates are awarding fewer new grants, but the hardest hit is the social, behavioural and economic sciences (SBE) directorate, which the White House proposed eliminating in its 2027 budget request to Congress. SBE has so far awarded 84 new grants this year, about 11% of its average, mainly in areas of artificial intelligence, cognitive neuroscience and decision-making. Much of its portfolio, such as sociology, archaeology, political science, economics and geography, has received zero new awards. “I don’t think they’ve realized yet that there’s no awards,” one NSF staff member says of researchers in those fields. “That’s going to be clear soon.”

6423226

Reposted by Roman Feiman

Hakeem Jefferson @hakeemjefferson.bsky.social · 17/08/2026

“I was asked not to mention the Times at all. I’m choosing not to honor that norm, b/c silence abt how accountability journalism gets suppressed is itself a form of suppression.” Read the back story of how @nytimes.com and @thebulwark.com killed @adambonica.bsky.social’s piece. What a mf badass.

data4democracy.substack.com

With Fear and Favor

Institutions don’t have courage. They have incentives.

11197451

Reposted by Roman Feiman

vincentholst.bsky.social @vincentholst.bsky.social · 13/08/2026

In 2023, @nature.com published 'Papers and patents are becoming less disruptive over time', receiving world-wide media attention. Our Matters Arising, published after a 32 month delay (more on that soon), shows that the reported decline can largely be attributed to dataset artefacts. 🧵

The average CD_5 index per year for Web of Science. The original Park et al. decline (top curve) becomes essentially flat (bottom curve) when removing papers with CD_5=1. Those papers largely correspond to dataset artefacts.

9352173