In 1994 Fefferman proved that by analytically continuing a neural network with sigmoid-activations to the complex numbers, you can read off the weights of the network by looking at the pole structure of the function. www.math.stonybrook.edu/~bishop/clas...

Have almost developed literal hatred for even the sight of all these swarms of anons on social media. Any scroll around and it’s just clowns and cowards being "based."

Got a book to review from a UP for suitability for publication, and it seems really nice content-wise. Lots of algebra and cool stuff related to ML. But somehow every third chapter title has "AGI" in it. I am sure this has more to do with the press than the author, right?

proceedings.mlr.press/v54/abbasi-y... A paper by Yasin Abbasi Yadkori and others. Same theme. Also showed polynomial bounds, but under "nicely behaved" embeddings (which implies isoperimetry) and low curvature assumptions.

The TCS variety papers on mechanism design / auction theory / "Turing's invisible hand" are so interesting. They are always kind of hard-core, have a lot going on technically (you can't just gesture to yourself what they mean w/o careful reading), but eventually still have a simple story. This is in

It feels strange in a way to "observe markets" and "be curious about the futures open" when the oil shock (exacerbated by no small means by oil futures manipulates) has the full potential to turn into a catastrophe. But we have started seeing this now. Almost 10%, another 10% would be good. But

My sister keeps coming across a particular sort of picture of dogs and

Does Poincare inequality suffice for polynomial-time sampling? arxiv.org/abs/2603.25622

Not the screenshotting type, but I saw this continuously on the other site each time I logged in. I don't want to dunk on these guys, but the underlying sentiment has really annoyed me no end for a long time. The underlying latent factor now shows up in the various weekly "AI forecasts."

This wasn't clear to me and us at the time, but this paper actually contains a very general result using which you could prove any kind of universality result for all sorts of NNs. But it would like using a bazooka to attack a fly.

There is now a whole sub-industry around LLM routing, gateways, even compute arbitrageurs (e.g. inference dot net). This is a basic but nice study on arbitrage in such settings and implications for the ecosystem (e.g. price drops, market entry, revenue cannibalization etc.) arxiv.org/abs/2603.22404

A very polite collaborator trying to find a meeting time. I was going to say 5 AM research meetings and I go long back, but resisted, and just settled with "oh, sure, could work."

Evolution seems to have repeatedly converged on a similar architecture in the early olfactory system of many animals. We offer an efficient-coding framework to understand this. Out in @pnas.org now! Led by @jzv.bsky.social, driven by Juan Carlos, + help from Farhad. www.pnas.org/doi/10.1073/...

This image (from www.gregegan.net/SCIENCE/Elli...) sums up for me (roughly, intuitively) why there is SO(4) symmetry for orbits: the two equal-sized vectors A and B can be rotated *independently* while preserving the semi-major axis, and hence the energy, of a Kepler orbit. And so(3)+so(3) = so(4).

This is really fascinating, the story of how one of the oldest known writing systems was recently deciphered www.lrb.co.uk/the-paper/v4...

Very interesting work arxiv.org/abs/2603.20071

Looks like we will get this, but in slow motion. It would be good to just have a circuit breaker or two and get done with it. It's nice ("nice") that there's so much chaos that if it were to really roll over, the "told you so about the AI bubble" bros and ma'ams will get no airtime.

Recently connected with an old Aikido sensei who used to administer our tests. Found that he had moved to the city, so naturally decided to ask about training. The meeting was v pleasant, till the bit about training came up. He got furious (and made it v apparent), but with a smile throughout. :))

Cool stuff. Linear RNNs can be simulated by NC circuits with a O(log* n) overhead over optimal polylog-depth parallel computation (like transformers). Nonlinear RNNs have an additional multiplicative O(log⁡ n) depth overhead. arxiv.org/abs/2603.03612

Our paper “Multifidelity Simulation-based Inference for Computationally Expensive Simulators” has been accepted at ICLR 2026! 🥳 We hope this can be a practical solution for anyone analysing and doing inference on computationally expensive simulators. Paper: openreview.net/pdf?id=bj0dc...

Starting to think more and more that the spectre to deeply understand and design against is not Gestell/Gestalt vis-à-vis technology, but hyper-financialization. In some ways enframing is only the condition of possibility, but hyper-financialization is the dominant instantiation.

Came across this picture in a social media post (not my picture), and thought woah, had no idea Flannery O'Connor wrote that much, should get it. Just an initial search immediately

I really dislike the free usage of tags like "boomer" "millennial" "gen whatever." It might be because I have too many "boomer-coded" traits, one of them is simply calling people whenever I think there is a need to talk, without warning, rather than doing the whole to and fro text garbage.

Preprint alert: Simulation-based validation of Bayes Factor computation with @paulbuerkner.com and S. Stroppel. We bring lessons learned in SBC to validation of BFs. arxiv.org/abs/2508.11814 The idea is simple: simulate data from the models, fit and see if the inferences are calibrated. 1/10

Nice perspective on PPI from a survey sampling angle. arxiv.org/abs/2603.19160

If yields are a barometer, if Powell wasn't a lame duck, markets would have taken a 18 December 2024 style dump today (which would be good, but we just don't get it anymore these days).

Haven't even managed to submit my reviews yet and here people are getting desk rejects. Will need to do a night out and submit today given that a grad student has a paper with me. What a mess. Got an unusually hard to read stack -- I think I know why...

This book is a great introduction to Bayesian statistics cswr.nrhstat.org #rstats #rcpp #cpp #stats

NSF Update through March 13, 2026 1/2

I was going to say there hasn't been an antipope for more than 500 years, and that there could be a chance to have a gay anti-pope. But then realized that historical anti-popes were usually respected figures, and could be considered fairly legitimate political claimants.

[1/4] Our paper “Characterization of Exponential Families of Lumpable Stochastic Matrices” (with Shun Watanabe) has been accepted for publication in Information Geometry. Preprint: arxiv.org/abs/2412.08400

Habermas was v interesting to me. I learnt a great deal from his writings, at first about systems/social theory, then as a v particular style of thought*. This is the end of an era. He always felt like a historical figure, in the canon of great German intellectuals, not someone who was still alive.

The main problem with daylight savings is that for about a fortnight or so, I keep calling my parents 45 minutes / 1 hour before the time they usually fully wake up.

A different way of saying that the generalized Laplacian Y - M is PSD (it satisfies Lx = 0). I am not sure if this holds for generalized Schrodinger operators as used to define this magical invariant (see the matrix definition here). en.wikipedia.org/wiki/Colin_d...

The New School was a good place, sad to see it fully implode in the past few years, although the writing had been on the wall for longer.

Now out in Journal of Neuroscience: www.jneurosci.org/content/earl... How are motor memories and sensory information combined when you are playing the piano? Congrats to Amin and Medha! I know having a paper out may seem small and insignificant against all that happens in the world right now!

How does the brain build a memory? A common assumption is that the neurons activated during an experience collectively form the memory engram. In our new Nature Neuroscience paper (finally out!), we show that this is not the case. www.nature.com/articles/s41...

Quite neat - answers some questions I had (at least in one dimension): arxiv.org/abs/2603.08939 'Shape-constrained density estimation with Wasserstein projection' - Takeru Matsuda, Ting-Kam Leonard Wong

What is the interplay between representations learned from (language) surface forms alone, and those learned from more grounded evidence (e.g.,vision)? Excited to share new work understanding “Cross-modal taxonomic generalization” in (V)LMs arxiv.org/abs/2603.07474 1/

You cannot make this thing up

No time to meet today, perhaps tomorrow? I will block some time.

Something more has been happening in the oil market.The worst thing you can do in a supply shock is to block open and free market pricing. De facto price controls will turn into shortages if all this doesn't buy too much time.

NASA smashed a spacecraft into one of a pair of asteroids. the impact shifted the pair’s orbit around the sun: www.nytimes.com/2026/03/06/s...

I had posted this late night. But then later removed as I wanted to keep spamming in check (but forgot I talk to myself here). Too bad I didn't have the stomach, was a 53% net profitable trade had it been closed by 4 PM ET today (so in less than 16 hours).

The new conformal prediction book now seems to be final after a bunch of updates: arxiv.org/abs/2411.118...

During COVID this guy used to write his name as "Dr. Vinay Prasad MD MPH, Associate Professor" (the full thing). Started out as a mask crank (working up a sweat how no RCTs had been done, so it was a no go), then graduated to being a professional (and increasingly cynical) contrarian shaped by SM.

It's interesting that the sort of financial advice my father always gave me (from childhood) seemed to be too paranoid, overly defensive and hyper-hedged, focused on hard assets, developing multiple income streams and multiple fungible skills. It seemed out of touch and for a different world, but

The outcome of the Iran-US-Israel war would be a major deciding factor when these two countries again enter a state of war. Iran's presence and regime posture was a key reason the previous war did not expand into Armenia proper. Iran saw any action on Zangezur as a threat to their security, and they

Barely started unboxing books and already hurt my back. This is gonna take months.