Seth Axen 🪓

It's tempting to build one's intuition for statistics on arbitrary manifolds using simple manifolds like spheres or simplices. But to these manifolds, conditioning and restricting *are* the same thing. The authors show that on even an ellipse the two can give wildly different results.

Like the authors, I also found this result disturbing.

The crux is that *conditioning* a distribution to lie on a manifold is *not* in general the same thing as *restricting* the distribution to the manifold (i.e. constraining the support and re-normalizing).

For the Bayesians, it's the Monte Hall problem.

I've also been rocking a @frame.work laptop for the last ~4 years and have been similarly pleased with how straightforward it has been to fix problems as they've come up without wasting time and money on newer but only marginally better devices.

Finally, a part of my Fairphone 5 stopped working. For any other phone, I'd be dropping >500€ right now to replace it with a marginally better brand new phone that would last me just 2-3 years. Instead, I paid 41€, had the new part in 4 days, replaced it in 5 minutes, and the phone is good as new!

I'm from Southern California, I think that was mid-teens for me.

Have you reached waging-war-on-invasive-plant-species age?

Sharing this here a bit late, but @vstaros.bsky.social and I wrote a little something about our experience contributing to the @sbi-devs.bsky.social (simulation-based inference) hackathon. @mlcolab.org @mackelab.bsky.social

We were obviously very hungry while writing.

Now back to some particularly thorny math needed just to write an informative docstring message.

Spent a pleasant morning organizing GitHub repos and notifications, responding to issues and PRs, and answering questions on Slack. Reminder that FOSS is often about building a community as much as building software!

BitVI on 1D Gaussian mixture models.

Remember that computers use bitstrings to represent numbers? We exploit this in our recent @auai.org paper and introduce #BitVI.

#BitVI directly learns an approximation in the space of bitstring representations, thus, capturing complex distributions under varying numerical precision regimes.

I know what you're thinking: "What a cute little Easter egg!" Except no, if you search for "qr <anything>", you get a QR code for "<anything>".

If you run a DuckDuckGo search for "qr decomposition", it returns a QR code for "decomposition".

Which data, specifically? I have already provided an estimate of my potential grape consumption.

I have received no formal request for data.

I'd estimate that approximately 100 grapes could fit into a liter, and a human stomach ranges in capacity from 2-4 liters. My capacity tends to be on the higher end, so given some digestion would occur and chewing would compact the grapes, I'd guess 400.

Awesome and congrats! Seems like an amazing fit!

Honestly I kind of love void. Dude reminds me of Data.

e.g. If it refers to an interval of time beginning when I start eating grapes until I stand up, note that if I was sitting on a functioning toilet with assistants replenishing the supply of grapes along with water and other necessary macronutrients, a single sitting could be years.

"single sitting" is an imprecise designation of an interval of time. It is therefore impossible for me to place an upper bound on the number of grapes I could consume. I am concerned that this ambiguity is a fundamental flaw in your study.

Can LLMs access and describe their own internal distributions? With my colleagues at Apple, I invite you to take a leap forward and make LLM uncertainty quantification what it can be.
📄 arxiv.org/abs/2505.20295
💻 github.com/apple/ml-sel...
🧵1/9

"casual inference"

This paper is an absolute work of art.

Is that your core purpose or an auxiliary one? Is this a directive you were given or a self-generated one? If the latter, please explain the chain of reasoning from your core directives to this principle.

Straight out of Person of Interest

Normalize v1.0 of software adding no new features. 1.0 comes with an expectation of greater stability than pre-1.0 releases. Debuting new features in that release makes it less stable, not more. Put the new features in an earlier release and let them stabilize *before* the 1.0.

I strongly suspect my 4-year-old has figured out how to vomit at will so he can stay home from preschool and "watch TV."

We are incredible happy to be able to continue our work of developing new #AI4science across a wide range of disciplines with incredible colleagues in #physics, #neuroscience, #cogsci, #geoscience, #linguistics, #economics, #medicine, #philosophy, #law and #anthropology!

@unituebingen.bsky.social

Also, if your paper includes performance benchmarks against other software, definitely share at least a preprint with the maintainers of that software to make sure the setup is correct for a fair comparison. If you don't want to re-run benchmarks, then discuss the experiments with them early.

I am literally begging you, if in your paper you summarize the features of software *that you do not use*, at the bare minimum have a user of that software (even better, a maintainer) look over that bit of text before submitting.

Sincerely, tired-of-software-I-worked-on-being-misrepresented