Jeffrey Bowers

A screenshot of an email from the Bluesky Moderation Team with the subject line “Bluesky Account Behavior” sent at 22:15. The email states that a Bluesky account controlled by the recipient, with the handle @spacelawshitpost.me, posted a reply with an image where the alt text reads: Charlie Kirk sitting in a white T-shirt that says “Freedom.” A negative consequence follows. The email indicates this post violates Community Guidelines regarding violent or threatening speech, specifically listing threats of violence or physical harm against individuals, groups, institutions, or the general public, and incitement of self-harm or suicide.​​​​​​​​​​​​​​​​

A screenshot of an email continuation from the Bluesky Moderation Team displaying a bulleted list of Community Guidelines violations. The first bullet point describes threats of violence or physical harm as including threats or encouragement of violence against individuals, groups, institutions, or the general public. The second bullet point defines incitement of self-harm or suicide as content that encourages self-harm or suicide. The third bullet point explains wishes of harm as expressions of a desire for harm directed at a specific person or group. Below the bullets, a paragraph states that users may not use Bluesky to break the law or cause harm to others and that all users must be treated with respect. The next paragraph informs the recipient that as a result of these violations, your account has been taken down. A closing paragraph thanks the user for doing their part to keep Bluesky a welcoming and empowering place for all users and instructs them to email moderation@blueskyweb.xyz if they have questions or wish to appeal the decision. The email ends with the signature Bluesky Moderation Team.​​​​​​​​​​​​​​​​

A screenshot of an email reply from a user with the initials AJ and profile picture to the Bluesky Moderation Team, sent at 08:01 with the subject line Re: Bluesky Account Behavior. The email states the user wants to appeal the decision, asserting there was no threatening or violent connotation behind their use of the image. The user explains that the Charlie Kirk post is not a threat but rather a reaction meme expressing disappointment, with no threat made and no incitement of violence. The user mentions they have attached multiple examples and explanations, specifically 12 images in total, demonstrating the use of the image as a reaction meme. Below this text is a blue hyperlink to a Bluesky post URL. At the bottom of the screenshot is an embedded image showing a Bluesky post from a user with the handle @britculpsapp.bsky.social with several emoji reactions and a repost indicator showing Bluesky Elder. The post text reads: how to explain: this has been used as a reaction image to spending $20 on doordashing wendy’s & isn’t a threat. The timestamp shows October 5, 2025 at 3:34 PM.​​​​​​​​​​​​​​​​

A screenshot of an email from the Bluesky Moderation Team sent at 08:11 with the subject line Re: Bluesky Account Behavior addressed To: Hide My Email. The email reads: Hi there, Thank you for reaching out to appeal the takedown of your account on Bluesky. We have carefully reviewed your case, including the information you provided and the activity that led to the suspension of your account. After investigating, we have determined that your account activity was in violation of our community guidelines, specifically promoting hateful content or being disruptive to other users. We take these guidelines seriously to ensure a safe and respectful environment for all our users. As a result, we regret to inform you that your appeal has been denied, and your account will remain suspended. This decision is final and cannot be reversed. We understand that this may be disappointing news, but we must prioritize the well-being of our community and maintain a consistent enforcement of our policies. We encourage you to review our community guidelines to better understand the expectations we have for user behavior on our platform. Please do not create new accounts, since those will be taken down for ban evasion. Thank you for your understanding and cooperation.​​​​​​​​​​​​​​​​

Unfortunate update: Link reached out to BlueSky and got email back now on reason.

His appeal is denied. He will remain permanently banned on BlueSky’s side.

Here are the screenshots he just sent me as of 8:22 AM EST(my time)

Yes, the Home Secretary really did just say we need to hold off protesting against the British government's continuing complicity in Israel's daily slaughter in Gaza out of respect for Jews grieving over the Manchester synagogue attack.

Although we disagree on some points, @rtommccoy.bsky.social has been really helpful in responding to my comments.

@rplevy.bsky.social @rylanschaeffer.bsky.socia @mcxfrank.bsky.social @nancykanwisher.bsky.social @melaniemitchell.bsky.social @lampinen.bsky.social @jeanremiking.bsky.social and others, my list just a quick check through who I follow.

@spiantado.bsky.social @tallinzen.bsky.social @evfedorenko.bsky.social
@mschrimpf.bsky.social @brendenlake.bsky.social @gretatuckute.bsky.social @tonyzador.bsky.social @garymarcus.bsky.social @neuranna.bsky.social @bayesianboy.bsky.social @cocoscilab.bsky.social @mpshanahan.bsky.social

Curious what people think about what LMs are telling us about innate priors, and human language more generally.

The failure of LMs with language-priors does not undermine the importance of innate domain-specific priors for humans. But it does suggest that LMs are not going to be a good model of human language anytime soon.

Although this would appear to lend support to the importance of language-priors, their findings do not support this conclusion either. Most basically, their prior-trained LMs still need orders more magnitude of data to learn.

Modeling rapid language learning by distilling Bayesian priors into artificial neural networks - Nature Communications Children can learn language from very little experience, and explaining this ability has been a major challenge in cognitive science. Here, the authors combine the flexible representations of neural networks with the structured learning biases of Bayesian models to help explain rapid language learning.

McCoy and Griffiths (2025) argue that distilling LMs with a language prior through meta-learning allows LMs to learn natural languages (English) quickly and flexibly: www.nature.com/articles/s41....

In Bowers (in press; Psychological Review) I challenged the claim that LMs can learn a language when trained on human-scale diet of words; consistent with the need for language-specific priors: osf.io/preprints/ps...

Check out following article where I challenge some of the conclusions of McCoy and Griffiths (2025) regarding language learning in language models (LMs) doi.org/10.31234/osf...

Q: Why did the LLM cross the road?

A: We're not sure, but it achieved 94.7% on CHIKENBench-Large

Critical part of the President's new $100,000 charge for H1-B visas: The Administration can also offer a $100,000 discount to any person, company, or industry that it wants. Replacing rules with arbitrary discretion.

Want visas? You know who to call and who to flatter.

WhatsApp wouldn’t let me write the word genocide in a text this week. It kept changing it to ‘Genie I’ (?) I slowed right down and realised the word isn’t in its dictionary. I found that weird.

The papers you started with are not inconsistent with grandmother cells, for reasons detailed in many of the papers I've linked above. But please feel free to send papers you think are inconsistent. Again, my claim is simply the current evidence does not rule them out. bsky.app/profile/tyre...

You assert the data support distributed processing. I show through simulations that the data are consistent with grandmother cells (and that we should not rule out localist models based on current data). I guess we can leave it there.

You assert the data support distributed processing. I show through simulations that the data are consistent with grandmother cells (and that we should not rule out localist models based on current data). I guess we can leave it there.

You argued against grandmother cells based on: (a) many neurons always respond to a single input; (b) neurons always respond several inputs. My point is that these findings are consistent with grandmother cells - see article below where we quantify this. Findings also consistent with a threshold.

So when you present alexnet an image of a dog and an output unit for dog is activated 90%, and alexnet categorises the image as a dog it is not a grandmother cell because the output unit cat is activated at 5% and mouse 1%, etc? If I say cat is a bit like a dog it is a grandmother cell?

Does AlexNet and the rest use one-hot encoding for object ID? Does AlexNet use localist or distributed coding for object classification? If a model classifies an object based on a single unit being the most active, then a grandmother cell model, yes? (even if other units are partially activated).

I don’t think one-hot mean that. It means one unit more active than all others for classification. Units have confidence levels and soft-max applied. Neither leads to all other units being completely off.

And that is not how one-hot encoding works (even after soft-max) or localist representations in the interactive activation model (the canonical example of localist as opposed to distributed representation). I think you are rejecting a strawman grandmother cell theory. Discuss in detail in my papers

Is the category grandmother a continuous variable? It was always about categorising objects.

The following article is now in press at Psychological Review. Interested to hear what people think! "The successes and failures of Artificial Neural Networks (ANNs) highlight the importance of innate linguistic priors for human language acquisition".

osf.io/preprints/ps... via

Easy to miss this one, but here is another relevant article highlighting that predictions are often driven by the background of the images, not the objects themselves: 2024.ccneuro.org/poster/?id=156

Another question: how do you reconcile the following claim with the finding that humans identify objects on the basis of line drawings just fine? Indeed, humans identify line drawings the first time they see one, and the cavemen drew line drawings of objects without any training on line drawings.

Not sure the DCNNs do a lot of explaining in or out of lab. But let’s see how things develop.

The point of our paper is that you can capture an impressive amount of variance in a model that classifies objects in a qualitatively different way to humans. I’m not clear what we should take from these predictions. You think the texture predictions tell us something about early human vision?

Is the claim that ANNs trained to classify objects learn human-like texture representations even though humans learn to classify objects based on shape? Do you take the findings to suggest that ANNs are poor models of core object recognition?