“10,000 times more efficient” I think they cooked

Guys, Moore's law just became irrelevant. Beff – e/acc started out working on quantum computing and realized "this is dumb there's a better way" and then just launched a thermodynamic processor. It's 10,000x more efficient at converting joules into tokens. This means

Huge news from Gill Verdon and the Extropic team. The next wave of compute is here - and we are honored to have Extropic as one of the earliest NAIA members last year. Here’s to the future - and the infra it will take to Reindustrialize America.

Extropic have unveiled their thermodynamic sampling units or TSUs TSUs differ from CPUs and GPUs by producing samples from a programmable distribution. This new form of computing powered by their products XTR-0, X0, and Z1 The era of thermodynamic computing is here

Seriously though, check out the website. You won't regret it. Several blogs, docs, and code now available. Extropic.ai

There is absolutely no fundamental reason we build AI the way we do today. There certainly is a radically different approach that is orders of magnitude more energy efficient. I’m going to find it before I die arxiv.org/abs/2510.23972

Beff – e/acc Maxwell Ramstead New paradigm ahead (active inference priced in)

Active Inference is exciting as a paradigm because it’s an architecture that can deliver genuine world models (‘deeper’ than an LLM’s linguistic gloss). This a technological trajectory that can get us to a kind of AI that could legitimately be understood as AGI.

I take back everything I said about the algorithm. Turns out all I needed to do was actually ship.

well if strawberry says so...

Today, Extropic did something most AI hardware companies won't: they open-sourced the foundation of their entire technological approach. THRML is a Python library for simulating and accelerating generative AI models through thermodynamic computing. At first, that might sound

Viscerally experience the inputs and outputs for yourself github.com/extropic-ai/th…

Liron Shapira 1. You clearly don't know anything about probabilistic machine learning. That's ok; it's a graduate-level topic, and I wouldn't expect you to. 2. The 10,000x claim is actually ultra-specific and backed by a paper with a ~40-page appendix of machine learning and physics

Beff – e/acc I hear an independent researcher replicated the DTM experiments using thrml and now that's open sourced too👀 github.com/pschilliOrange…

My 2c about Extropic's paper. 1 - Yes, it's a small scale demonstration BUT they solved 2 of the most important issues of probabilistic computing: (a) scaling beyond a single global energy and avoiding exponentially long mixing time and (b) avoiding using bulky RNGs 1/n

Meet the XTR-0 A way for early developers to make first contact with thermodynamic intelligence. More at: extropic.ai

The foundations have been laid. Now it's time to scale. Excited for the Thermodynamic Intelligence takeoff ahead.

We’re honored to be backed by cyber•Fund and featured in this post on their investment thesis: cyber.fund/content/noumen… Digital AI stacks simulate randomness on hardware meant to erase it. In contrast, our approach treats noise as computation, enabling the radical acceleration

The hardest hitting question I got on launch day: how much do you bench?