iiinova — where your digital self keeps evolving

Sam Goree @samgoree.bsky.social · 08/06/2026

Though I'm annoyed the visual style of my website (warm off-white background with highly readable serif font) became the default for both AI generated and AI company websites. I've done a shallow redesign away from that style now.

my personal website https://samgoree.github.io, which has a warm off-white background.

anthropic.com, which shares the warm off-white background

000

Sam Goree @samgoree.bsky.social · 20/05/2026

Since everyone's talking about the atrocious research practices that arXiv encourages (and their new much-needed quality standards), I'm going to repost two passages from our 2024 paper on the attention economy of computer vision research. dl.acm.org/doi/abs/10.1...

Faster publication through arXiv. The digital materiality of the PDF affords an alternative
publication process that came up in several of our interviews (P4, P5, P8): the preprint server arXiv.
P5 says arXiv is a major source of anxiety for his students: “So my students...had the unpleasant
experience of, you know, finishing a paper when they’re just about to submit and they saw a paper on
arXiv that did almost the same thing. Like, oh my, months of work just went down the drain.” This
emotion, the sense of loss when a project becomes unpublishable because someone else got there
first, echoes Su and Crandall’s observation of “selective amnesia” [96], except in addition to old
papers quickly becoming obsolete because something else is better, current papers may become
obsolete because they are no longer first.

ArXiv changes the writing process as well. As P9 explains, “I find that the amount of time and
effort put into each paper has gone downhill. For my papers, I would put a huge amount of time,
especially in the intro...And recently I have been doing less and less of that and basically because the
students are saying, ah you know, the new kids they don’t even, they just skip the intro, they just go
directly to the method. So I feel like, oh my god, nobody’s even gonna read my beautiful prose!...But
also I think because the field progresses faster, papers become obsolete much quicker. So it might be
reasonable not to spend so much effort on a single paper if you know that in a year it will be obsolete.”
In other words, the author now writes differently in order to better fit the faster publication system,
spending less time on introductions in order to spend less effort on a paper which might quickly
become obsolete.

120

Sam Goree @samgoree.bsky.social · 25/02/2026

Other patterns of collapse follow different patterns, which we categorize into an eight-category framework. The T-SNE plots show the MNIST latent feedback scenario, which exhibits oblate explansion, and two retrained scenarios which both exhibit coherent contraction into wrinkled contraction.

T-SNE plots of three scenarios on MNIST, illustrating the latent manifold collapse across many generations of retraining or re-generating.

A table showing eight kinds of collapse according to three metrics: sigma-intra, M_LB and PR_G. Each is characterized by its local and global collapse or expansion.
Coherent Expansion exhibits local and global expansion, with both local metrics increasing
Wrinkled Expansion exhibits local expansion and global contraction, with both local metrics increasing
Anisotropic Expansion exhibits local expansion and global contraction, with M_LB decreasing
Oblate Expansion exhibits local expansion and global contraction with M_LB decreasing
Coherent Contraction exhibits local contraction and global contraction with all metrics decreasing
Anisotropic Contraction exhibits local contraction and global expansion with both local metrics decreasing
Wrinkled Contraction exhibits local contraction and global contraction with M_LB increasing
Oblate Contraction exhibits local contraction and global expansion with M_LB increasing

100

Sam Goree @samgoree.bsky.social · 25/02/2026

We use three metrics for dimensionality, and found, interestingly, that these feedback patterns result in very different kinds of collapse. But the feedback loops that exhibit directional contraction and obey the Markov property all exhibit the same kind of collapse, which we call neural resonance.

Spark plots showing the curves for each of five experiments.
LGR-MNIST exhibits CC and WC
LF-MNIST exhibits OE
LGR-ImageNet5 exhibits CC
LF-ImageNet5 exhibits WE
CycleGAN exhibits CE
Acronyms are explained in the next post

110

Sam Goree @samgoree.bsky.social · 25/02/2026

We explored this phenomenon in four modes -- the Lucier setting, CycleGAN, in a diffusion model with latent feedback (where we train a model to recover the latent input and loop those), and in a fully retrained diffusion model. Then, we study each loop's dynamics in latent space.

diagram of four experimental scenarios -- Lucier's feedback loop, CycleGAN, Latent-feedback models, retrained models, with diagrams for each.

100

Sam Goree @samgoree.bsky.social · 30/05/2025

There's a cursed world of fully-AI generated SEO spam websites now

110

Sam Goree @samgoree.bsky.social · 22/05/2025

Special thanks to @irisvanrooij.bsky.social, whose article on reclaiming AI for cogsci whose table gave me the idea for defining AI as a "history of practices reflecting different ideas of AI." link.springer.com/content/pdf/...

193

Sam Goree @samgoree.bsky.social · 03/05/2025

Ooh I wish this paper had been out when I was writing my thesis! Systematization is a concept that I was grasping for when trying to talk about how AI research fails to handle subjectivity. Especially this observation.

Moreover, both benchmarks' measurement instruments involve crowdworkers, who, in the absence of an explicitly systematized concept, must rely on their own understandings of these high-level definitions, which may be contradictory (e.g., whether factually true generalizations about social groups are stereotypes or not).

040

Sam Goree @samgoree.bsky.social · 09/03/2025

Oh I meant in the art theory sense, like the arrangement of shapes, light vs. dark, focal points, etc. Most professional paintings or photos have very deliberate compositions which give a visual rhythm and organizational unity to the art.

110

Sam Goree @samgoree.bsky.social · 09/03/2025

But AI art made with earlier models that work differently look totally different (and imo 100% better). E.g. the work of Helena Sarin, which (I think) is made with a GAN trained on only her own paintings/drawings x.com/NeuralBricol...

1704

Sam Goree @samgoree.bsky.social · 09/03/2025

I find it easier to see if I blur the images. Like, look how weird the composition is. There's also often a weird vignetting effect, where the edges are darker than they should be. Since all these AI images come from the same kinds of generators, we learn these tells.

3785

Sam Goree @samgoree.bsky.social · 09/03/2025

Oh this is related to my research. Generative models like this learn to take a random noise image (like tv static) and repeatedly reduce the noise, iteratively making the details more like a training image and more like the associated text prompt. A middle stage might look like these

18211

Sam Goree @samgoree.bsky.social · 06/02/2025

Turns out you can't. They're not equivalent. They just look kinda similar so people use the easier algorithm instead because it's discrete and little differences don't matter as much as speed. (image from Noah Snavley's very helpful computer vision course slides)

a scaled impulse minus a 2D gaussian *approximately* equals the laplacian of gaussian

030

Sam Goree @samgoree.bsky.social · 28/01/2025

While I don't know what the future holds, scaling is clearly not the only way towards better AI systems. DenseNet had ~5% of the parameters of VGG and performed better. Brains do all the same things on way less power. Clearly there are better ways out there, they just don't fit corporate timelines.

100

Sam Goree @samgoree.bsky.social · 26/01/2025

I'm teaching image processing this semester using this photo of Cornelius instead of the infamous Lena image

3302

Sam Goree @samgoree.bsky.social · 06/01/2025

Do you have an idea for an app, website, device or computational system? Would you be interested in a team of computer scientists and engineers make your project a reality? Stonehill Computer Science is looking for capstone project pitches! Details in images below.

If you are interested, send an email to sgoree@stonehill.edu with a brief description of your idea and we'll schedule your pitch to the class.

Details:

Each spring semester, the seniors in the Stonehill College Computer Science Capstone build a real working software application for a real client that solves a real problem. An application can be a mobile app that you use on your phone, a web app or website that you use in a browser, or a desktop program that you install on your computer - any kind of software. This year, for the first time, we have computer engineers in the class alongside computer scientists, so we can support a wider range of form factors.

Past examples include apps that:

help local businesses to increase foot traffic with prizes

manage reusable takeout food containers to reduce waste

help doctors and nurses manage patients in a post-operative recovery room

help students at Stonehill request exam accommodations online

help women refugees connect refugee camp services for the non-profit Women Refugee Route

manage vision-impaired work-study students for the non-profit Perkin's School for the Blind

create an augmented reality tour of the campus for prospective students

help users to reflect and pray and share their reflections with other members of their religious community.

Instructions for pitches:

Send us an email with a brief description of your app and we'll schedule your pitch to the class.

You should plan on spending 5-10 minutes on your “elevator” pitch to the class and we will assign your pitch to a slot on M,W,F @ 1:30-2:45 on 1/24, 1/27 or

1/29. If you have a conflict between 1:30-2:45, let us know and we will schedule around it.

If your idea is selected, you must commit to joining (in person or via Teams) the 1:30 class on Friday once every two weeks to give feedback to the developers. We are able to negotiate the day/time a bit, but if you (for example) teach MWF 1:30-2:45, we are not able to consider your project.

If your idea is selected, you should make plans to continue maintaining the app after May. In the past, clients have shared ownership of the app with some of the developers or hired new developers to continue work. We are unable to guarantee any bug fixes or technical support after the seniors graduate.

Depending on your pitch, if your idea is selected, we may need to set up a cost sharing agreement. We can provide labor and basic cloud computing and/or 3D printing costs, but we may ask you to contribute towards any hardware or software purchases.

We are looking forward to hearing from you!

Stonehill College Computer Science Capstone Team (Mike Salé and Sam Goree

162

Sam Goree @samgoree.bsky.social · 07/12/2024

My specific language (fwiw)

020