David Pfau

An academic colleague just told me his students aren't using LLMs because he can't afford even the low tier monthly subscription for everyone. I'm not sure the folks making $Xm/yr at frontier labs understand just how much of a bubble they're in.

An academic collaborator just told me his students aren't using LLMs because he can't afford even the low tier monthly subscription for everyone. I'm not sure the folks making $Xm/yr at frontier labs understand just how much of a bubble they're in.

20 Sept

Having a seeded bagel on unceded land

20 Sept

We spent $100m building a robot that can pirouette at 5000 rpm and performs Swan Lake beginning to end in 30 seconds. Ballet is solved.

14 Sept

Ok, it's time to coin a law of social media discourse: when something happens that is legitimately worth getting angry about, social media will drive people to get angry about it for the wrong reason.

14 Sept

Prime example I can think of is all the discourse around data centers and water usage, but I see echoes of it in people getting upset at OpenAI because they "stole mathematician's ideas" (probably not) and not because they're destroying the soul of mathematics.

10 Sept

Do the Zoomers know about the Y2K bug?

10 Sept

I can't believe how many of my normie friends I had to explain superintelligence, AI safety, rationalism and x-risk to today.

15 Aug

Are people on here aware that on The Other Site all the remaining left wing people are talking about "Woke 1" as some wild thing that is very definitively in the past? Has that discourse made it over here at all, or does everyone here believe Woke is still alive and well?

15 Aug

To be fair they are mostly talking about it in a sort of "ok now that we can honestly look back and discuss some of the problems, let's figure out what we can do better for Woke 2" sort of way.

newscientist.comSingle-celled organism with no brain is capable of Pavlovian learningA trumpet-shaped, single-celled organism seems able to predict one thing will follow another, hinting that such associative learning emerged long before multicellular nervous systems

6 Apr

Using AI to cure cancer or fight climate change makes as much sense as using neuroscience to figure out AI. It's all yak shaving. If your goal is to "solve X" just go solve X, don't try to solve Y then use Y to solve X.

David Pfau reposted

Chris Simms

@chrisnsimms.bsky.social

13 Mar

This is an awesome discovery: A single-celled organism with no brain called Stentor seems capable of Pavlovian learning. Yes, it can actually learn to associate two things despite having no neurons. My latest for @newscientist.com. 🧪 #science #memory #learning www.newscientist.com/article/2519...

@drsimevans.carbonbrief.org

3 Mar

I miss when ideas mattered in machine learning.

David Pfau reposted

Simon Evans

12 Feb

NEW ANALYSIS: China's CO2 has now been 'flat or falling' for 21 months * Down in 2025 * Still below Mar 2024 * Clean energy wave a key factor If this is China's peak (TBC) it's the climate story of the century so far… www.carbonbrief.org/...

en.wikipedia.orgAlexander Abian - Wikipedia

30 Jan

This you, @truebe.horse ? en.wikipedia.org/wiki/Alexand...

30 Jan

"Alexander (Smbat) Abian (January 1, 1923 – July 24, 1999)[1] was an Iranian-born Armenian-American mathematician who taught for over 25 years at Iowa State University and became notable for his frequent posts to various Usenet newsgroups, and his advocacy for the destruction of the Moon."

20 Jan

One year, and we're on the brink of war with the rest of NATO. God help us after three more years of this.

9 Jan

I'm not really sure how to square this with the US spending 1% of GDP on datacenter construction.

Joey Politano🏳️‍🌈@josephpolitano.bsky.social

The US is losing blue-collar jobs for the first time since the early pandemic—year-on-year job losses have hit 145k as manufacturing, transportation, & mining industries lose jobs at a rapid pace while growth in construction has nearly zeroed out

8 Jan

If anyone DMed me recently - to use DMs for this app in the UK, I have to go through age verification, which I have not done yet on principle. If you are trying to reach me privately, email or Twitter DM works.

7 Jan

I'm starting to think The Blob might not have been such a bad thing after all.

8 Jan

Ok apparently some people thought I was talking about a movie. I meant the US foreign policy establishment in DC.

arxiv.orgExtracting Anyon Statistics from Neural Network Fractional Quantum Hall StatesFractional quantum Hall states host emergent anyons with exotic exchange statistics, but obtaining direct access to their topological properties in real systems remains a challenge. Neural-network wav...

Excited to share our latest preprint, on using deep learning to extract the weird fractional quantum statistics of 2D topological materials: arxiv.org/abs/2512.15872

This requires the calculation of multiple degenerate ground states. People have used neural networks to calculate *a* ground state for the FQHE, and have calculated fractional spin statistics in lattice models, but no one has done these calculations in the continuum before.

All-around fantastic work from Andres Perez Fadon and collaborators, and this shows how these calculations could be done for systems where we still don't know how to write down even qualitatively correct solutions.

nobelprize.orgNobel Prize in Physics 1998The Nobel Prize in Physics 1998 was awarded jointly to Robert B. Laughlin, Horst L. Störmer and Daniel C. Tsui "for their discovery of a new form of quantum fluid with fractionally charged excitations...

Excited to share our latest preprint, on using deep learning to extract the weird fractional quantum statistics of 2D topological materials: arxiv.org/abs/2512.15872

This behavior was so weird and unexpected that it ended up being the subject of the 1998 Nobel Prize in physics. No one even knew how to write down a wavefunction that explained this phenomenon until Bob Laughlin just dreamed one up one day. www.nobelprize.org/prizes/physi...

Rather than write down a solution through a stroke of genius, we trained a Transformer-like neural network to calculate these fractional spin statistics using "entanglement interferometry", and got answers more accurate than the Laughlin wavefunction.

Excited to share our latest preprint, on using deep learning to extract the weird fractional quantum statistics of 2D topological materials: arxiv.org/abs/2512.15872

Fundamental particles are either fermions or bosons. Fermions, like electrons, make up matter, while bosons, like photons, are force carriers. But in two dimensional materials in a magnetic field, things can start to get weird...

The collective excitations of many electrons behave a lot like particles...but instead of particles with integer spin (bosons) or half-integer spins (fermions) they behave like they have weird fractional spins like 1/3, 1/5, etc. This is the *fractional quantum hall effect*.

9 Dec 2025

Not sure why they bother asking for my age on this app, everyone knows everyone on Bluesky is exactly 39.

I am not at NeurIPS.

The first recorded case of COVID-19 was on November 30, 2019. ChatGPT launched to the public on November 30, 2022. What new horror was unleashed on the world yesterday?

cabinetmagazine.orgUtterance Is Place Enough | Frances RichardMapping conversation

20 Nov 2025

I'm a bit surprised that in the era of conversational LLMs, there hasn't been a revival of interest in the art world in Cohen, Frank and Ippolito's "Argument Drawings" www.cabinetmagazine.org/issues/2/ric...

arxiv.orgA Generalized Bias-Variance Decomposition for Bregman DivergencesThe bias-variance decomposition is a central result in statistics and machine learning, but is typically presented only for the squared error. We present a generalization of the bias-variance decompos...

New (sort of) preprint on arXiv today: a generalized bias-variance decomposition for Bregman divergences! arxiv.org/abs/2511.08789

Until one day recently, I dropped the notes into an LLM and...got the raw LaTeX right back out. In theory I could have done it myself in a few hours, but this was enough to get me over the "activation energy" to finally do it. So thank you, Frank and Gemini! (Please don't hate me, Bluesky)

I will say, it is a very odd feeling to finally tick something off your to-do list after 12 years...satisfying, but odd.

New (sort of) preprint on arXiv today: a generalized bias-variance decomposition for Bregman divergences! arxiv.org/abs/2511.08789

But then a weird thing happened...people found it anyway, and actually started citing it like it was a novel result! I guess I accidentally made the top Google result for "Bregman bias-variance" or something.

So I reached out to Frank Nielsen, who helpfully pointed me to the relevant literature. I meant to set the record straight, but by this point I'd actually lost the original LaTeX for the notes, and just never had the time to rewrite them...

New (sort of) preprint on arXiv today: a generalized bias-variance decomposition for Bregman divergences! arxiv.org/abs/2511.08789

This was actually a note I wrote way back in grad school, but never published or even posted to arXiv. After proving the result, I couldn't find a prior reference, but it seemed so basic that I figured it must have been done already. So I posted it on my website and moved on.

Trump's approval rating is now roughly comparable to what it was in mid-2020, at the height of COVID unemployment and George Floyd protests.

8 Nov 2025

I think that we should all strive to live our lives in such a way that, when we die, our greatest accomplishments are not overshadowed by constant bickering about exactly how racist or sexist we were.