Reposted by Jeffrey BowersThe Tennessee Holler @thetnholler.bsky.social · 06/10/2026JON STEWART: “So in the FUTURE we WILL use A.I. to cure diseases… but right now it’s mostly murder, surveillance, and surge pricing for the McRib — What a trade! I’m getting the sense you guys might be full of shit.” 17999733243
Reposted by Jeffrey BowersJeffrey Bowers @jeffreybowers.bsky.social · 21/09/2026For more evidence that object recognition processes do not drive brain predictions in the ventral visual stream, see our paper below. We found that Brain-Score predictions in IT and V4 largely driven by the backgrounds not the objects. bsky.app/profile/jeff... 111
Jeffrey Bowers @jeffreybowers.bsky.social · 21/09/2026For more evidence that object recognition processes do not drive brain predictions in the ventral visual stream, see our paper below. We found that Brain-Score predictions in IT and V4 largely driven by the backgrounds not the objects. bsky.app/profile/jeff... 111
Jeffrey Bowers @jeffreybowers.bsky.social · 16/09/2026www.facebook.com/share/r/19Sg...facebook.com 010
Jeffrey Bowers @jeffreybowers.bsky.social · 08/09/2026This is a great organisation I’ve got involved with. Please consider getting involved or donating. If you have any questions please feel free to get in touch. 032
Jeffrey Bowers @jeffreybowers.bsky.social · 22/07/2026Maybe I'm missing something, but isn't this the message from "tit for tat" many decades ago? neurosciencenews.com/social-coope.... e.g., see: cs.stanford.edu/people/erobe...neurosciencenews.comRethinking the Prisoner's Dilemma: How Cooperation Wins - Neuroscience NewsSimple opponent-specific responses allow cooperation to spontaneously evolve and thrive. 051
Jeffrey Bowers @jeffreybowers.bsky.social · 21/07/2026Sorry, should have tagged authors in the exchange: @andreadevarda.bsky.social @fepdelia.bsky.social @hopekean.bsky.social @lampinen.bsky.social @evfedorenko.bsky.social @fedeadolfi.bsky.social @rachelfloodheaton.bsky.social @guillermopuebla.bsky.social 020
Jeffrey Bowers @jeffreybowers.bsky.social · 21/07/2026Here is our latest response regarding the alignment between large reasoning models and human reasoning. The debate hinges around the methods NeuroAI researchers use to study natural intelligence–the focus of our upcoming GAC: doi.org/10.31234/osf... bsky.app/profile/cogc...doi.orgOSF 1101
Jeffrey Bowers @jeffreybowers.bsky.social · 20/07/2026Democrats are funded by big corporations that don’t want any of these things. 000
Reposted by Jeffrey BowersAnna (Anya) Ivanova @neuranna.bsky.social · 17/07/2026Excited to participate in a GAC discussion at #CCN2026 on Aug 4th! Is NeuroAI on the right track? What are the most glaring issues? Comment below and help us prepare for our panel discussion! sites.google.com/ccneuro.org/... 0148
Jeffrey Bowers @jeffreybowers.bsky.social · 13/07/2026On this view is some sort of "in-life" optimisation and evolution destined to produce similar brain solutions when successfully solving sufficiently challenging tasks? @dyamins.bsky.social 000
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026In sum, current LLMs learn just as Chomsky would predict: they learn far slower and do not share human-like learning biases. Look forward to any responses: @kmahowald.bsky.social @futrell.bsky.social @spiantado.bsky.social 000
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026It seems a stench to use the results of the *Reverse and Hop *languages as evidence that LLMs, like humans, find impossible languages harder to learn. Especially in light of some of the other findings we summarized, where there were impossible languages that were just as easy to learn. 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026And more importantly, here are the results of Kallini et al. The left-hand figure shows two impossible languages that were indeed harder (because they required learning multiple impossible languages), but the other figures show highly similar rates, as we wrote. 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026Not to be pedantic, but this is what we wrote regarding Kallini et al. (highlight added): 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026However, Futrell and Mahowald claim we have misrepresented the finding of Kallini et al., writing: 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026In the Bowers and Mitchell commentary we review findings from 6 labs and show that LLMs often learn impossible languages as easily, concluding: doi:10.1017/S0140525X26104579, e202 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026One finding that speaks to this is whether LLMs learn “impossible” languages as easily as humans. Unfortunately, there has been some confusion about the findings, and the recent BBS paper by Futrell and Mahowald contributes to this confusion. doi.org/10.1017/S014...doi.orgHow linguistics learned to stop worrying and love the language models | Behavioral and Brain Sciences | Cambridge CoreHow linguistics learned to stop worrying and love the language models - Volume 49 110
Jeffrey Bowers @jeffreybowers.bsky.social · 02/07/2026There is a great deal of interest in whether LLMs challenge Chomsky’s claim that human language acquisition requires some sort of innate inductive bias (e.g., some sort of universal grammar). aclanthology.org/2020.coling-... pemt.ru/wp-content/u... doi.org/10.1037/rev0...aclanthology.orgPriorless Recurrent Networks Learn CuriouslyJeff Mitchell, Jeffrey Bowers. Proceedings of the 28th International Conference on Computational Linguistics. 2020. 1104
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Organizers: please repost and start the debate! @jfeather.bsky.social @neuranna.bsky.social @neurograce.bsky.social @anayebi.bsky.social @mschrimpf.bsky.social 000
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026This is a community event, so great to have comments/feedback from a broader audience. For the sake of debate, here is one challenge I would pose: What have we learned about mind/brain from correlational studies with ANNs? 100
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Current plan is that I will introduce the GAC, and then there will be a debate that follows. Details of the format still to be determined. 110
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026This article outlines my position in upcoming GAC at CCN conference: 2026.ccneuro.org/gac-debates/2026.ccneuro.orgIntroducing the GAC Debates 110
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Basic claim is that the standard methods of NeuroAI are flawed, and that we need to stop focusing on prediction of out-of-sample variance in correlational studies and instead focus on effect sizes that test hypotheses. 100
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Check out my critique of prediction studies in NeuroAI (along with Guillermo Puebla, Sushrut Thorat, Konstantinos Tsetsos, and Casimir Ludwig) osf.io/preprints/ps...osf.ioOSF 193
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Current plan is that I will introduce the GAC, and then there will be a debate that follows. Details of the format still to be determined. 000
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026This article outlines my position in upcoming GAC at CCN conference: 2026.ccneuro.org/gac-debates/2026.ccneuro.orgIntroducing the GAC Debates 100
Jeffrey Bowers @jeffreybowers.bsky.social · 30/06/2026Basic claim is that the standard methods of NeuroAI are flawed, and that we need to stop focusing on prediction of out-of-sample variance in correlational studies and instead focus on effect sizes that test hypotheses. 100
Reposted by Jeffrey BowersMargarete Rosalind Weidenhammer @margareterosali.bsky.social · 24/04/2026🎯 "The Mad King having another tantrum about #Spain" #NoKings 26125
Reposted by Jeffrey BowersT.J. Stiles @tj-stiles.bsky.social · 11/02/2026This is another very important and revealing exchange. Rep. Scanlon is making a crucial point, one well worth listening to. And she does it calmly. But Trump has created a market for obnoxious, childish defiance of Congress. The Attorney General‘s response is outrageous, even impeachment-worthy. 532117574
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026That seems plausible to me. Yes, our visual experience will shape the visual representations we learn. Prediction is an interesting hypothesis, and may well be part of the story (more plausible to me than next word prediction in language). But getting better RSAs and Brain-Scores not good evidence. 120
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026So what do we learn about human vision? Trained on much more data and the objective is object classification. I would think the core of a theory of human vision would be the architecture 110
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026Another problem is that wide variety of architectures get the same brain score. And all do terribly on psychology experiments. 120
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026It is a deep neural network thought to be a model of the human ventral visual system that can predict brain activation on brain score. The better the prediction the better the model of core object recognition. But can’t recognise line drawings unless trained on them. Doubt Gibson would agree but 🤷♂️ 100
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026How would his “direct perception” view fit with a 50 layer RESnetx model? 100
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026Models in science are generally designed to understand phenomena, not be replicas of nature. In any case, showing an ANN can process & predict responses to natural images doesn't imply that the ANN is a good replica of biological vision. When tested in psych experiments, they are shown not to be. 000
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026Not too well though. So they were not as good models of human vision as alexnet? And a transformer trained on trillion images gives us better understanding still? And earlier models that don’t process natural images at all don’t give any understanding? I don’t get it. 110
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026So all models prior Alexnet did not improve our understanding of human vision? 110
Jeffrey Bowers @jeffreybowers.bsky.social · 10/02/2026It is a good model if it helps us understand vision. Engineering success is not the most important point as very different systems can solve the same problem. 110
Reposted by Jeffrey BowersChicago Todd @chicagotodd.bsky.social · 09/02/2026What professional US reporter has ever confronted a politician in this manner - just continuing to ask the question when the politician doesn't answer cogently. They just move on . . . 1432668
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026These models account for few findings in psychology (eg. see reference above). They often succeed using shortcuts. They are trained on orders of magnitude more data. There is almost no consideration of the fact that correlation do not imply mechanistic similarities. What have we learned so far? 010
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026Sure if incompatible with the rest of vision you have a problem. But seems to me jumping the gun - e.g. how does your model of motion perception explain consciousness? The starting point is not some unified model of neural predictions of naturalistic images. 130
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026You might build such a model in an Attempt to understand gestalt principles? 110
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026Yes, you should not train on the stimuli in mindset. A infant can recognise a line drawing the first time they see one. Training on line drawings to succeed on line drawings is not a way to evaluate a model. Falsification has an important role in science. But not so much in NeuroAI. 031
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026Yes, not everyone. But few researchers manipulate stimuli to test hypotheses. The vast majority of researchers are relying on some alignment metric (some version of correlation) to make causal conclusions. Or they seem to make mechanistic claims till pushed (motte-and-bailey). But not everyone. 020
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026Sure, and that is why relying on benchmark tests, where everyone competes on specific datasets, is misguided. We should be running a series of experiments on each phenomena, like done in psychology (and most everywhere else). Learn more, and reduce danger of specific confound in a given dataset. 110
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026I’m not sure how these experimental conditions from MindSet could be translated into naturalist stimuli, and researchers don’t try. But as we and paper cited above show, ANNs do fail on these stimuli; should change the interpretation of ANN prediction successes. openreview.net/forum?id=VkP...openreview.netOpenReviewPromoting openness in scientific communication and the peer-review process 000
Reposted by Jeffrey BowersAnthony L. Fisher @anthonylfisher.bsky.social · 09/02/2026Weird. I tried to warn libertarians about Trump 123399301
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026That is not the main reason. The key reason is that artificial stimuli are designed to test specific hypotheses, and they have the critical features to selectively assess the hypotheses. Complex naturalistic stimuli include many additional features that can contribute to uninteresting short-cuts. 170
Jeffrey Bowers @jeffreybowers.bsky.social · 09/02/2026Here is a paper just out using a wide range of artificial stimuli (many taken from our MindSet benchmark) showing how ANNs that identify objects do it in a non human-like way. Not sure how/why you would use naturalistic stimuli; their simplicity is by design. www.nature.com/articles/s42...nature.comVisual language models show widespread visual deficits on neuropsychological tests - Nature Machine IntelligenceTangtartharakul and Storrs use standardized neuropsychological tests to compare human visual abilities with those of visual language models (VLMs). They report that while VLMs excel in high-level obje... 1100
