Rachel Flood Heaton

I wonder how many others it happened to as well.

I just had a conference paper desk rejected by an automated system I believe to be an AI. The email said that the reasons for the desk rejection were listed at the end of the email. There were zero reasons listed. It's a reviewable paper. I'm not new at this. AI is really bad for reviewing.

I can buy twice as much chicken for the same price at Costco than I can at Cub, but I also have a place to store it. I have to admit that I have been kind of shocked at how expensive groceries are at the normal grocery store. Feels more like the shopping you do when you’re on vacation at a cabin.

Movie description of the B horror move Death PhD. Gawdawful movie.

Now this is a movie plot!!! Someone understands grad school.

AI philosophers in 2026:

Maybe it's me, and maybe it's slightly off topic, but is linear regression actually appropriate to those data? I feel like we're not starting out on solid footing.

Two things are happening instead: Contrived "LLMs sort of behave like people if you squint and ignore a lot of stuff" demonstrations, and suppression of clear falsifications of that idea.

ctrl-f for "psychologist", find a bunch of CS and EA people with no obvious formal training in psychology doing dumb shit based on nonreplicable social psych awww yeah

My first thought too. They were musing about nuclear testing after the treaty lapsed.

Moderna says FDA refuses its application for new mRNA flu vaccine The U.S. Food and Drug Administration is refusing to consider Moderna’s application for a new flu vaccine made with mRNA technology.

It is absolutely outrageous that Moderna’s flu vaccine was met with a “refusal-to-file” even after their approved their protocol with FDA and carried out the trial as agreed. This vaccine works better in older adults than the current flu vaccines.

apnews.com/article/mode...

I wish them good luck with that.

I'm going to look at these and follow up.

Okay I admittedly just skimmed that, but I will say, I think we have picked this up in a meaningful way that speaks to the part about inference, in humans and across different species. We don't say anything about motor activity. That's not our emphasis.

Okay but surfaces and geons are the higher order invariants!

There is a great deal of evidence in favor of parts-based accounts (even using photographic stimuli to test them). Can you justify the claim that the visual system does not use them? That's quite a strong claim.

Drawing of a cheetah and a second cheetah deconstructed into geometric volumes from the website of Kevin Bethune.

I think there is quite a lot of evidence that our world is built on geons. We know that there is visual priming in the parts of objects. Plus, even naturalistic stimuli can be decomposed into volumes.

This is, incidentally, the representational framework of the paper I shared earlier.

Biederman's Recognition-by-Components theory is an answer to this challenge. As I am sure you are aware, it explains how you can extract viewpoint invariant structural descriptions of parts in hierarchies of relations from the static image. Stereo is better, but one monocular vision is enough.

What I mean is that you don't have to choose. If you're more interested in architecture than learning/training, the right architecture with relatively little learning could easily handle both without any issue.

I'm not really trying to overextend, but trying to understand the implications. I think I am coming to the conclusion that the point where we actually disagree is whether the things some call artificial might in fact also be naturalistic.

I would potentially agree about the model building if I thought it was necessary to use a statistics-based deep learning framework to build models. Some of us would argue that we don't need that framework to build a model that can handle both types of stimuli.

I think there might also be a bias the other way: if the only stimuli a model can process are the ones it is trained on, then it necessarily excludes theories that allow for the perception of stimuli it cannot see.

I think everyone agrees that ecological validity matters. But that seems like a different issue than what seems to be a theoretical claim in this discussion.

I'm not sure that I agree it relies on a notion of modularity. A compositional representation is not really modular. The other thing that is getting lost is that such experiments aren't one and done. Individual results are combined with others and tested, along with ecological validity.

Sure, I did not evolve to see those. I also didn't evolve to watch Spongebob Squarepants (and amazingly I never have), but we would both expect me to be able to see and respond to it. Is it an implication of what you are saying that one should never test over novel stimuli outside the convex hull?

We see visual illusions in the physical world, too. I wouldn't call illusions are failures of perception, though. More like artifacts. Perception reliably keeps us alive, and inverse optics is ill-posed. It's unreasonable and impossible to expect perfection over an ill-posed problem.

I'm interested to hear a counter argument, but any shape or color a human can perceive, even interrupted, broken, or dotted lines and contours, seems to be inside that domain. We see and understand things like words and art. It seems to follow that all of the classic stimuli must then be included.

Fair enough. We never have enough time. I do think it might be worth a skim at some point if you are interested in a serious computational (not LLM) argument about the role of representation in inference for affordances.