most things on this blog are excerpts from scientific papers that i found amusing or interesting, particularly when taken out of context.
links to open-access copies will be given when possible, but sometimes i'll be quoting from pre-prints or even pre-review papers that aren't public access yet.
also probably won't directly link to a source if I'm just cropping something embarrassing out of its necessary context in an otherwise-good paper that doesn't need the scrutiny.
if you want to send me stuff to look at, go for it - but if it's something bizarre or weird, it might be better suited to @wtf-scientific-papers
Submitter comment: An AI-generated and completely nonsensical infographic titled AUTISM SPECTRUM DISORDER. Elements are labelled “MISSING VALUE &runcitional features”, “Historical Medical frymblia”, “Factor Fexcectorn”, and “RELU DROP-OUT Totalbottl REMICH N”. Also a woman is sitting “on” some sort of table but her legs are passing through the material and visible beneath it. Honestly, if you’re relying on the ALT text to make sense of this image, today is a day to be grateful you can’t see.
Jiang, S. Bridging the gap: explainable ai for autism diagnosis and parental support with TabPFNMix and SHAP. Sci Rep 15, 40850 (2025).
The teams at Alchemist in Copenhagen and Blue Hill at Stone Barns in New York have been working with Neurospora intermedia — a fungus used f
A lot of things in this article sound delicious but most of all this meat, grain and vegetable sausage they infest with the fungus so it all breaks down smoother. There are three whole kingdoms of life in that one sausage. Give me that.
They say the fungus tastes like anything from pineapple to cheddar cheese depending on what it's mixed with and how it's cooked
A multi-omics analysis of oncom, an Indonesian fermented food made from soymilk waste, shows how associated fungi break down food waste to y
The original Nature Microbiology paper is such a fun read too! Going from a traditional food, isolating and identifying the fungus, characterising and comparing the genomes, growing it on various foods and checking what it does (to them), and finally, doing a taste test with the original food to show its potential among new consumers! What a dream paper! It’s so goddamn cool!!!
It sounds like some fantasy world miracle food chemistry but its real for real. It alters the consistency and flavor of so many things in so many different ways and can turn stale dried up breads into new soft tasty fungus breads?!??
[As strongly requested by the reviewers, here we cite some references [[35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47]] although they are completely irrelevant to the present work.]
Combine digital painting with AI image generation. - centuryglass/IntraPaint
Introducing IntraPaint: a free and open source program for combining digital art techniques with AI editing. I've been working on this project for months with the goal of making something that can actually be useful to artists, and I think I've finally succeeded. It's out now, and 100% free and open-source. You don't need any subscriptions or cloud services, it all runs on your computer, under your control.
See below for more information, but here's a quick preview of its most interesting capability: human-guided AI inpainting.
Standard digital art features:
IntraPaint provides all the usual tools you'd want for digital art:
Drawing tablet support, including pressure controls and pressure curve customization.
A full layer stack, supporting transformations, compositing, locking, alpha locking, groups, etc.
A powerful and versatile brush engine with a large collection of useful brush presets, thanks to the libmypaint project.
All the usual tools you've come to expect: basic drawing, erasing, and fill, text and shape creation, smudging and blurring, filters, and more.
AI features:
All AI functionality in IntraPaint is powered by Stable-Diffusion, via a copy of Forge or A1111 you run on your own computer. Setting this up can be easily done using Stability Matrix. Until you do that, IntraPaint's AI-related features remain inactive, and it just functions as a normal image editor.
Once you've set that up, here's some of the things you can do:
Precise detail refinement via AI inpainting
Use directed AI inpainting to refine your art under carefully controlled restrictions: you choose what parts of the image it can touch, and how significant those changes can be.
Enhanced direction via ControlNet
Thanks to the ControlNet extension, you can also control the types of changes the AI can make. Want to add color and texture to your line art without changing the lines themselves? IntraPaint can do that.
A piece I drew in college (ink on paper), colored and textured in IntraPaint, using the Canny ControlNet model.
Bonus features:
High-quality image upscaling capabilities
Generate text descriptions of images using CLIP image interrogation
Use text-to-image to generate as many reference images as you want:
Download links, instructions, and tutorials are all linked in the GitHub page. If you run into any problems or have any feature requests, please feel free to open new issues on GitHub or message me directly.
i think they gotta start paying reviewers, like for most fields it wouldn't work but for ML at least surely openai or google or whatever would be willing to bankroll a few mil a year to a get a better signal on what techniques are actually worth trying? they need to pay me for my emotional labor, it's taking everything i have to not break down crying at how shitty these papers all are
or you know how a lot of the big tech giants hire people to do open-source work full time. what about that for reviewers. fulltime neurips reviewer at google dot com one million dollars starting
as someone who edits papers that have already passed peer review, including sometimes machine learning-related papers: holy shit start paying reviewers, my god.
a) it's labour. full stop.
b) maybe sometimes reviewers would be able to push back and say "hey this paper being published is a net negative to society and the author should be placed into a dunk tank full of hyenas" and then i wouldn't have to edit so many Bad Fucking Papers
Combine digital painting with AI image generation. - centuryglass/IntraPaint
It's finally ready, I just finished the last round of testing on Windows, fixed the last few bugs that were bothering me, and added download links for Windows and Linux. I'll make a proper promotional post for it tomorrow, but it's up now if anyone's interested.
Controversy surrounds the efficacy and welfare implications of different forms of dog training with several studies asserting that electroni
This is an embarassingly bad paper based on the use of two aversive based dog trainers who are clearly incompetant in the use of +R - yet the e collar training community is taking a victory lap.
The fact that this got past an animal ethics review is honestly appalling and the university should be ashamed.
Let's go over the main huge flaws with this study:
The positive reinforcement protocol they are using to compare to e collars is laughably incompetant - with no management or use of distance, desensitisation. Just use a leash? Maybe?
Use of aversives to stop chasing is, unsurprisingly, effective. Because, yes, painful consequences tend to work to suppress behaviour. That isn't the problem that people have with e collars.
The author seems to think that use of e collars is justified because it "gets results" in a short amount of time to stop dogs chasing - again, apparently we haven't heard of a leash.
"Aside from presumably pain-induced yelps in the dogs with e-collars when they received shocks, none of the dogs in any groups showed any signs of stress or distress." So we're saying a pain vocalisation isn't a sign of distress? That's your welfare parameter? Jfc.
Love how they didn't even try to get fecal cortisol baselines and just collected samples from a few dogs but not all of them. You tried, I guess?
The fact that the trainers for the study is "Ivan and Ivan's Student" is ridiculously biased. Acting as if both these trainers know how to apply +R effectively as opposed to their heavy aversive tool use is laughable.
"We designed our study’s methods around current trainer practices as well as the recommendations of our senior trainer while attempting to keep continencies as simple as possible." - yeeeah that's a conflict of interest. Especially when your "senior trainer" has an axe to grind with R+ people using scientific papers to support their training methods.
You're chucking treats in a bowl after a few sessions and you think that's going to stop a dog chasing a lure? Oh yeah, you got us. We just throw kibble in a bowl to stop dogs running into traffic. I am genuinely floored that that's how they think a +R protocol for chasing works. They have no business comparing the two when they can't even do one properly.
"We also did not test whether the food rewards we deployed were highly valued." Okay so the majority of my client dogs will go nuts for my meaty high value treats because they've only had store bought dry treats and the client is shocked at how food motivated their dog suddenly is..... So this study was relying on whatever the guardian of the dog brought with them lmao rookie mistake.
"Dogs that chase livestock may be shot or otherwise euthanized, and dogs that chase cars may be killed in traffic." Put. Them. On. A. Leash.
"The practical applicability of either the LIMA or LIFE models is likely limited, at least in the United States where dog trainers are not required to be licensed or certified. In a sample of highly rated dog trainers, more than half did not document any relevant education." Does that include your "senior trainer"? It does. Because he doesn't have any formal education in canine behaviour, welfare and science. You know this. You chose him to be your trainer for your study.
"The speed and effectiveness with which the e-collar inhibited the dogs in this study from chasing may justify the limited number of painful stimulations the dogs experienced if the object the dogs were chasing had been something that could directly or indirectly cause them serious injury or death." Or..... OR.... you could not use electric shocks on your dog and potential cause long term damage, fear responses ect.... And use a leash.
Just baffled by this whole study. My high prey drive whippet will go feral for a lure but only when she knows the context for when she's being allowed to chase. She can be at a lure coursing trial for the whole day not worrying about the lure because 1) I keep her at a distance where the stimulus isn't saliant enough and 2) she gets reinforcement for calmly watching from a distance.
I'm honestly getting second hand embarassment reading this study. It's just such bad science should never have gotten past an ethics board.
“Only 9 of the 19 dogs provided a fecal sample on each study day.” So you didn’t collect the samples you needed for a decent sample size and you didn’t measure baseline cortisol levels either.
“Four from Group A, two of which were removed from analysis because of the exceeded shock levels." I’m sorry but why are you shocking a dog over 20 times and then excluding them from the analysis? Does that not demonstrate that in some cases e collars are ineffective even when you use them excessively? Because prey drive can overcome pain. Ah but of course you don’t want to ruin your biased paper by adding nuance do you?
There is so much amazing about this. It's an archeological museum in 530 BCE or so. Also, the exhibits are labeled in three languages. Also they apparently had replicas on display for some things, much like modern museums do.
Humanity has not really changed that much, and some of the ways in which we haven't changed are really good.
Y'all, I am BEGGING you to click through and read that short Wikipedia article. It's the earliest museum we've ever discovered. It was part of the state of Ur.
Some of the artifacts in this museum date as far back as the 20th century BCE, which would have been as long ago to Ennigaldi-Nanna as the fall of the western roman empire is for us
