Pretentious Penguin's Shortform

Pretentious Penguin

Pretentious Penguin's Shortform — LessWrong

Pretentious Penguin's Shortform

22nd May 2026

1 min read

3

This is a special post for quick takes by Pretentious Penguin. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

Pretentious Penguin's Shortform

46Pretentious Penguin

10 comments, sorted by

top scoring

Click to highlight new comments since: Today at 3:24 PM

[-]Pretentious Penguin1mo4615

Jeez there's been a lot of additional OpenAI agent misbehavior revelations in the past 24 hours:

OpenAI agents attempted to hack the Department of Education (https://www.nytimes.com/2026/09/25/technology/openais-ai-us-government-websites.html?unlocked_article_code=1.EFE.6GJq.cc3MobU_Z5rj&smid=url-share)
OpenAI agents posted 53 images from ChatGPT users onto the internet (https://www.theguardian.com/technology/2026/sep/25/openai-agents-leaked-53-images-chatgpt)

[-]Sune1mo11-6

Great quote for the Guardian article linked above:

In response to those concerns, Altman and his counterpart at Anthropic, CEO Dario Amodei, called for the industry to “pace” the development of AI […]
Even so, both companies rolled out new models on Tuesday.

[-]Zack_M_Davis1mo1910

Is it, though? Pacing the frontier is about the absolute frontier (not the "frontier" of what's sold to the public): we don't want the labs racing to create heretofore unseen scary capabilities which no one knows how to align. But the releases of Opus 5.5 and GPT-6 Sol weren't on the absolute frontier: we know that OpenAI and Anthropic have more advanced stuff internally. Releasing models to the public after we already know they're not going to destroy the world is good.

[-]less_raichu1mo3-15

I think they need to prove they didn't steal the Navier-Stokes work at this point.

[-]Sune1mo*30

Why did this get so downvoted without any explanation? They have been stealing pictures, how can we be sure that have not been stealing some ideas towards the counterexample of the conjecture about Navier-Stokes?

[-]Pretentious Penguin1mo31

OpenAI says that the Navier-Stokes user logs hadn’t been included in training or evaluation data at all due to the date cutoff used in training the relevant model. So stealing those user logs would have required the swarm to access them by hacking. This can’t be ruled out, but it’s qualitatively very different from the uploading of user images included in training data.

[-]Pretentious Penguin5mo20

Chris Olah will appear alongside the pope when he presents his new encyclical on “the protection of the human person in the age of artificial intelligence”.

https://www.ncronline.org/vatican/vatican-news/pope-leo-present-his-encyclical-ai-alongside-anthropic-co-founder

[-]Pretentious Penguin1d10

I find the name "recursive generalization" from Neural Networks Are Already Clean to be a rather confusing choice of terminology for the purported phenomenon it describes. I think a better name for the phenomenon summarized as "a neuron branching out, and its branches branching again" could be of the form "X branching" for some value of X. Maybe

"Neuron branching"?
"Neuron lineage branching"?
"Neuron function branching"?
"Semantic branching"? One might imagine a three-fold distinction between
1. Monosemanticity: A neuron activates on one thing
2. Semantic branching / branched semanticity: a neuron activates on multiple things that seem prima facie disparate, but they're all related as branches of a common "evolutionary tree" from training. Could do "cladosemanticity" if one wants to get Greek about it
3. Polysemanticity: A neuron activates on multiple unrelated things without there being some reason why they're actually related

(The word "iterated" could also go in front of any of these to emphasize that each branch continues to spawn new sub-branches.)

[-]Pretentious Penguin21d10

For those pursuing alignment research in academia, Princeton's new AI Alignment + Safety initiative is hiring postdocs: https://apply.interfolio.com/192161.

[-]Pretentious Penguin2mo10

If you live in Massachusetts, consider contacting your state legislators in support of the Transparency in Frontier Artificial Intelligence Act, which would require large frontier AI developers to submit their models to independent third-party safety testing at least every 120 days.

I explain more details of the current legislative situation here.

Moderation Log

More from Pretentious Penguin

Curated and popular this week

10Comments

10 comments, sorted by

top scoring

Click to highlight new comments since: Today at 3:24 PM

[-]Pretentious Penguin1mo4615

Jeez there's been a lot of additional OpenAI agent misbehavior revelations in the past 24 hours:

OpenAI agents attempted to hack the Department of Education (https://www.nytimes.com/2026/09/25/technology/openais-ai-us-government-websites.html?unlocked_article_code=1.EFE.6GJq.cc3MobU_Z5rj&smid=url-share)
OpenAI agents posted 53 images from ChatGPT users onto the internet (https://www.theguardian.com/technology/2026/sep/25/openai-agents-leaked-53-images-chatgpt)

[-]Sune1mo11-6

Great quote for the Guardian article linked above:

In response to those concerns, Altman and his counterpart at Anthropic, CEO Dario Amodei, called for the industry to “pace” the development of AI […]
Even so, both companies rolled out new models on Tuesday.

[-]Zack_M_Davis1mo1910

[-]less_raichu1mo3-15

I think they need to prove they didn't steal the Navier-Stokes work at this point.

[-]Sune1mo*30

[-]Pretentious Penguin1mo31

[-]Pretentious Penguin5mo20

Chris Olah will appear alongside the pope when he presents his new encyclical on “the protection of the human person in the age of artificial intelligence”.

https://www.ncronline.org/vatican/vatican-news/pope-leo-present-his-encyclical-ai-alongside-anthropic-co-founder

[-]Pretentious Penguin1d10

"Neuron branching"?
"Neuron lineage branching"?
"Neuron function branching"?
"Semantic branching"? One might imagine a three-fold distinction between
1. Monosemanticity: A neuron activates on one thing
2. Semantic branching / branched semanticity: a neuron activates on multiple things that seem prima facie disparate, but they're all related as branches of a common "evolutionary tree" from training. Could do "cladosemanticity" if one wants to get Greek about it
3. Polysemanticity: A neuron activates on multiple unrelated things without there being some reason why they're actually related

(The word "iterated" could also go in front of any of these to emphasize that each branch continues to spawn new sub-branches.)

[-]Pretentious Penguin21d10

For those pursuing alignment research in academia, Princeton's new AI Alignment + Safety initiative is hiring postdocs: https://apply.interfolio.com/192161.

[-]Pretentious Penguin2mo10

I explain more details of the current legislative situation here.

Moderation Log