🔥Telegram İfşa

AI Training Load Fluctuations at Gigawatt-scale Risk of Power Grid Blackout? 108GW Large Load Queue Tesla Megapacks Supercapacitors Gigawatt-scale batteries PyTorch No Power Plant Blow Up semianalysis.com/2025/06/25/ai-…

Out today in nature: we show that individual neurons have diverse tuning to a decision variable computed by the entire population, revealing a unifying geometric principle for the encoding of sensory and dynamic cognitive variables. nature.com/articles/s4158…

Representational drift without synaptic plasticity biorxiv.org/content/10.110… #biorxiv_neursci

Investigating action topography in visual cortex and deep artificial neural networks biorxiv.org/content/10.110… #biorxiv_neursci

New theory: Categorization is Baked Into the Brain It is not at the end of sensory processing but a core computation that begins at the earliest stages, occurring throughout the cortex as predictive feedback shapes feedforward processing. doi.org/10.31234/osf.i… #neuroscience

Just published my new article, "Marketplace": my first attempt at efficient GPU training without backprop 😄🎉 I've been considering eliminating backprop for a while. I had an idea, experimented for two weeks, and it worked! Here's how it works:

In a classic study of "mental rotation", Shepard and Metzler (1971) found that the time to compare two 3D cube-made objects was proportional to their angular difference. But *what is going in the brain* during this process? 🔗 Metzler & Shepard (1971): semanticscholar.org/paper/Mental-R…

Glad to announce, the second article is here 😄🎉: Marketplace V2 is all you need: A training algorithm on par with backprop that needs only forward pass Believe it or not, this time, it's on par with backprop! It has a great potential to revolutionize ML training!

Researchers were able to freeze learning in mice. This great recent work shows the striatum is required for skill learning but not skill use (through optogenetics). Understanding the memory-making mechanism helps understand the core idea. Here is my toy model and notes:

Why don't we have a programming language for deep learning where dtypes and shapes are part of the type system?

Grid cells do not operate as a global positioning system but rather estimate position within multiple local reference frames Kevin Allen nature.com/articles/s4159…

The primate visual system is a marvel of nature, inspiring the convnet. But how did it evolve? Do all highly visual mammals possess a ‘ventral stream’—a hierarchy of brain areas with increasing selectivity for complex forms? In our new study, we tackled these questions using

“Recurrent pattern completion drives the neocortical representation of sensory inference” by Hyeyoung Shin et al. Nature Neuroscience nature.com/articles/s4159…

We believe what we see, but we also see what we believe. This is why we see illusions - our prior knowledge of the sensory world tricks us into seeing things that aren't there. How does this happen in the brain? In this study... nature.com/articles/s4159…

Very interesting finding regarding the cortical code! Remarkably, it was published recently using data from 2017. I’m curious about the spatial distribution of these neurons, how are they organized across the cortical sheet and in terms of laminar profile.

Intelligence is the ability to acquire new knowledge and using it, as effectively and efficiently as possible, to solve new problems the entity was not exposed to before. Solving in-distribution problems is automation, solving out-distribution problems is intelligence.

I’m super excited to finally put my recent work with Tim Behrens on bioRxiv, where we develop a new mechanistic theory of how PFC structures adaptive behaviour using attractor dynamics in space and time! biorxiv.org/content/10.110… 1/8

New blog post: "The Not-so Bitter Lesson" If you ever wanted to know what the tiny corp , DSPy, SAT solvers and fusion reactors have in common, go ahead and read it.

active inference on thermo hardware is the future

Hello Thermo World.