Adverb

Interview from the Titan submarine investigation. The witness name is blacked out. The first line is "Well, I'm sure you're familiar with my film Titanic."

Great moments in redaction history.

The power of even just this frame lmao

A man's face is shaved by a robot

It's gonna be a long forever 🫩

Zitron diaper-posting will always be more interesting to normal people than serious, thoughtful critiques. This is bad news but it's the state of the world. Hopefully policy-makers know better; I doubt it.

Serious critiques of "AI" just will not exist in the mainstream discourse until the topic itself becomes very boring.

also: "ah yes, return_tensors='pt' can give you a numpy array"

baffling to me that huggingface still has error messages that are just straight-up incorrect after they were reported three years ago

We're looking for 2D visual artists (in US/Canada) to collaborate on a research project, co-designing a new art feedback tool.
It's a remote, compensated opportunity.
Details & apply: rise.csit.carleton.ca/studies/arti...
Approved by Carleton University Research Ethics Board (CUREB-B #124851).

If I were to just ask it to give me a completion on the actual site, it'd be so much better every time.

gpt5 code like uuuhhh gpt3

yeah they are so much noticeably worse at everything, esp coding on cursor. I don't understand it. Maybe something about how the tool-use is done? Baffling to me.

Being yelled at in-person is a bad experience even for people who are big strong secure self actualized whatever, and I think people can typically see that but putting it online somehow obscures it to them.

Having been in/under it a few times, I think another aspect is that people take social conflict & online conflict especially as "whatever, fighting is good/neutral" when the reality is that at any scale on social media it's usually an unproductive stressor!

oh we're gonna *lose* lose the next big war

Even though the models run many forward passes, each one is a single case of "glimpse this noisy image and do a forward pass of denoising" which I don't think is sufficiently connecting the timesteps to learn hierarchical integration.

It even points to diffusion models' failure cases as an example, which ofc I love.

Things like HRM & TRM almost feel like the answer for models actually learning to, through recursion that's reasonably connected by gradient, sort through hierarchical integration.
arxiv.org/abs/2508.15082 has imho some interesting thoughts on the symbolic side

I point out that his research ideas are compelling because I'm absolutely not belittling his contributions! But making grand predictions and having to be wrong about them consistently just shifts the focus so badly.

bsky.app/profile/adve...

I'm in this weird place of seeing models completely fall over at generalizing concepts that're all over the place in training data while also thinking that "model should explicitly use if p, then q" is cuuurazy

It's so odd because lecture title linked down-thread would make you think that he'd see "bag of tricks that a connectionist system mysteriously learns" is cool and yet...

big fan of neruosymbolic systems, bitter lessons, and "ai bubble bursts in 1 year" every 1 year.

Like this is awesome: digitallibrary.amnh.org/items/21dd36...

but instead he's "guy who makes predictions about the state of 'AI' in two years and is consistently wrong"

Yeah, the way that LLMs can make the most insidious code drives me bananas. cursor autocomplete has almost ruined my life for a week many times!

Developmental psych rules. It is absolutely a perfect match for producing better ML architectures, tasks, whole paradigms. But he just keeps saying things that are proven not true lol

What's funny about this is that, given his cognitive science papers, I'd probably be really interested in his takes if this weren't just the case.

Gary Marcus is mostly known for making incorrect predictions.

But I think this llm-cum-author was like seasoned-ish at programming. eh idek

If it weren't for the fact that 20 de facto juniors and 30 other researchers would end up using it, would have dumped 90% of the code from that specific ~LLM~. Took normal code and made it 100 times more convoluted and brittle.