Bruno Neri

If you wanna change the world, start off by making your bed. If you make your bed every morning, you have accomplished the first task of the day and it will encourage you to do another task ... #leadership youtu.be/pxBQLFLei70

*Alice goes to Japan!* The book is available for purchase from Amazon.co.jp! As always, price is almost all printing (except a ☕). 🙃 amazon.co.jp/-/en/微分可能な不思議の国のアリスの冒険-ニューラルネットワーク設計入門書-第1巻-Simone-Scardapane/dp/B0D9QHS5NG/

Keep going down the rabbit-hole with Alice guided by @PetarV_93 @mmbronstein @TacoCohen #geometricdeeplearing

This paper kicked off our team's studies into the intricate relationship LLMs have with confidence. Now landed in @NatMachIntell 🚀 Give it a read, esp if you enjoy CogSci-style analyses of LLMs 🧠 Thoroughly impressed by Dharsh's leadership on this work! More outputs soon 👀

@PetarV_93 Nice news! Some deadline in mind?

new book chapter in the air

"A Mechanistic Analysis of Looped Reasoning Language Models" by @HughBlayney , @arroyo_alvr , Johan Samir Obando Ceron, @pcastr, @AaronCourville, @mmbronstein , @epomqo Paper: arxiv.org/abs/2604.11791 #machinelearning

*Learning without training: The implicit dynamics of in-context learning* by Dherin et al. An interesting work that tries to explains in-context learning by "contextual" low-rank updates on the MLP components. arxiv.org/abs/2507.16003

There's finally a community implementation of Neural Assets (in PyTorch)! Go check it out 👇 Neural Assets was one of the first (and maybe even the first scalable?) solution(s) to the long-standing problem of multi-entity consistency in visual generative models. One of the most fun projects I had the chance to work on (with the amazing @Dazitu_616).

If you want to improve your skills about flow matching and diffusion model take a look at this amazing course by @peholderrieth

A guide to remembering everything you read: bit.ly/4rxUHIr v/@CuriousMindsHub

@neribr @2ptmvd Link to the paper arxiv.org/abs/2602.18428…

@docmilanfar @2ptmvd sorry 😅

"The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning" by Mojtaba Sahraee-Ardakan, @2ptmvd , @docmilanfar #machinelearning

"Categorical Flow Maps" by @daan_roos_ , @osclsd , @FEijkelboom , @mmbronstein, @wellingmax, @ismaililkanc , @LucaAmb , @jwvdm Paper: arxiv.org/abs/2602.12233 #machinelearning

"A Researcher's Guide to Empirical Risk Minimization" by @LarsvanderLaan3

"Did you ever play ball, Max?" "No, never have." "But I make it a little more fun to watch, you see." RIP Robert Duvall

Our paper "Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin" has been accepted to ICLR 🎉 This work connects two puzzling phenomena in LLMs that were previously studied separately: attention sinks and compression valleys, showing they're fundamentally linked through massive activations. We propose a three-stage theory of how LLMs structure computation across depth, which also helps explain why intermediate layers work best for embedding tasks while generation needs full-depth processing. Great collaboration with @arroyo_alvr, Enrique Q, @fedzbar, @epomqo, @mmbronstein, and @ylecun.

The Elusive Concept of Time. Your instincts treat time like a background meter the whole universe shares. Relativity does not take time away, it forces you to earn it operationally. Events are points. Motion is a curve through them. A clock is not a metaphor, it is a worldline with a number attached to it. Two observers can disagree on which distant events are simultaneous, and nothing contradictory happens, because causal structure is still pinned down by light cones and invariants. Even in weak gravity the rule shows up. A clock deeper in a gravitational potential ticks more slowly relative to one far away. Near a compact object the difference becomes hard to ignore. Add rotation and spacetime itself picks up a twist. That twist is frame dragging, not a new force, just geometry telling you that time and angle are coupled in a rotating spacetime. In the animation You are watching a geometry lesson disguised as a black hole scene. The fabric is a visualization of the field shaping clock-rates and the paths light can take. The ripples are driven by local proper time, so their phase visibly slows as you approach the horizon. The accretion disk is lensed through Kerr ray tracing, and its brightness is pushed by redshift and beaming so the approaching side can flare while the receding side dims. Beacon points at different radii pulse at different rates, so you can see time dilation without any labels. The bead ring is a redshift tracer, with intensity scaled by a g³ proxy so deeper emission arrives weaker and shifted. The math breakdown Start with what a clock actually measures. Proper time τ is the accumulated time along an observer’s worldline. In special relativity, the invariant interval is ds² = c² dt² − dx² − dy² − dz² Along a timelike path, dτ = (1/c) √(ds²) = √( dt² − (1/c²)(dx²+dy²+dz²) ) If the observer moves with speed v, so dx²+dy²+dy²+dz² = v² dt², then dτ = dt √(1 − v²/c²) That is time dilation as geometry. The moving clock accumulates less τ between the same pair of events. Now add gravity. General relativity replaces the flat interval with a metric gᵤᵥ that depends on position: ds² = gᵤᵥ dxᵘ dxᵛ For a stationary clock in Schwarzschild geometry (mass M), the time component is g_tt = −(1 − 2GM/(rc²)) If the clock sits at fixed r (no spatial motion), ds² = g_tt c² dt², so dτ = dt √(1 − 2GM/(rc²)) Closer to the mass means a smaller factor, so the clock ticks more slowly relative to a clock far away. That is the rule used to drive the fabric phase in the animation. Now connect time to light. A gravitational field shifts photon frequency. Between an emitter at rₑ and an observer at rₒ, f_obs / f_emit = √( (1 − 2GM/(rₑ c²)) / (1 − 2GM/(rₒ c²)) ) For a far-away observer rₒ → ∞, f_obs / f_emit = √(1 − 2GM/(rₑ c²)) Deeper emission arrives redshifted. Lower frequency. Lower energy per photon. In the render, the disk intensity uses a Kerr-derived redshift factor g (clipped for stability). The bead ring uses a simple radiative proxy I_obs ∝ g³ I_emit to make that effect visible. Finally, why rotation looks like a twist. A rotating black hole is Kerr geometry. The key structural change is a nonzero g_tφ term, which couples time to angle. That coupling is frame dragging in equations. Near the hole, being stationary is not the same notion everywhere, because the local inertial frames are being pulled around the spin axis. So the moral stays clean. Time is not a universal substance flowing everywhere at one rate. It is what clocks accumulate along worldlines. Light cones constrain what can influence what. Invariants are what everyone agrees on. The rest is operational detail that only feels universal because our daily corner of the universe is slow and mild. #GeneralRelativity #Gravity #FrameDragging #BlackHoles #Spacetime

At St. Peter Basilica

for the first time, eeml treads mountainous terrain 🏔️ while being very close to the seaside 🌊 not to mention it's the best AI summer school in EE & beyond! come join us, it'll be amazing! @clarelyle @dorothychou @ninamiolane @sedielem @apsarathchandar @re_rayne @fedzbar