The startup Extropic.ai headed by cult leader Twitter X personality Beff Jezos (Guillaume Verdon) came out of stealth mode recently, on the heels of a $14.1M Series Seed funding round in Dec'23 led by Steve Jang of Kindred Ventures, who is known for backing Uber and Coinbase.
Extropic claims to be developing a novel AI hardware acceleration platform that harnesses the intrinsic noise and fluctuations in analog circuits to perform probabilistic computations more efficiently than traditional digital processors.
The company's new web site features a brown color theme with a floating, amorphous, shapeshifting turd. Presumably, a quantum thermodynamic turd.
Hilarity then ensued on the Internet.
But I also think it's totally unfair to compare Extropic to Theranos. Totally, totally, totally, totally unfair.
I took a read of the published litepaper to try to understand the claims. While I am by no means an expert in... <checks notes> ... "parameterized stochastic analog circuits"... or <checks notes again> ... "the most energy efficient neurons in the universe [paraphrased]"... here is my understanding of the claims.
Their key claims are:
- Using parameterized stochastic analog circuits to directly implement Energy-Based Models for generative AI tasks... hahaha. Think of it as a noisy electronic circuit that you can "tune" to generate useful patterns of randomness for certain computing tasks. On today's digital processors, generating these random samples requires significant computation and energy consumption. With Extropic's approach, they are "free".
- Achieving significant improvements in speed and energy efficiency compared to running sampling-based algorithms on digital hardware.
- Targeting specialized low-volume, high-value applications with superconducting chips operating at cryogenic temperatures.
- Developing room temperature semiconductor devices as a longer-term path to address larger markets, although this claim seems questionable.
- Building a software compilation layer to map EBMs to their analog hardware.
As for what they have currently built, the litepaper provides limited evidence. They show a microscope image of an "early device that tested several possible superconducting neuron designs" but don't provide performance metrics or comparisons to digital baselines.
The claim about room temperature semiconductor devices is particularly dubious and seems to contradict other aspects of their pitch.
Earlier in the litepaper, they emphasize how thermal noise and fluctuations are a key enabler of their approach, allowing them to harness intrinsic randomness for probabilistic computing. They even say "devices must be physically small and low power to be strongly affected by [thermal fluctuations]."
But then they propose semiconductor devices operating at room temperature, where thermal noise would be much higher than in their cryogenic superconducting chips. It's hard to see how these room temp devices could maintain the energy efficiency and sampling speed advantages touted earlier.
The semiconductor proposal seems like an attempt to have their cake and eat it too - to claim huge performance gains from specialized low-temp analog hardware but then also promise a path to mass market chips. But realistically, the room temperature semiconductors would likely face the same issues around thermal noise disrupting digital logic that they claim will limit traditional computing.
It feels like an attempt to appeal to investors by promising a best-of-both-worlds scenario. But the two paths (cryogenic superconductors vs. mass market semiconductors) have very different engineering constraints that make a unified approach tenuous at best.
The claim of room temperature superconductors has been a recurring theme in the history of physics, with many past claims turning out to be premature or unsupported. However, (I think) Extropic is not claiming to have developed room temperature superconductors.
What (I think) Extropic proposes is using superconducting devices at cryogenic (very low) temperatures for their most efficient AI accelerators. Separately, they suggest developing semiconductor devices operating at room temperature as a way to scale their technology, but these would not be superconductors.
The idea of room temperature superconductors has been a holy grail in physics for decades. Superconductivity allows electrical current to flow with zero resistance, but it typically requires extremely low temperatures, which limits practical applications. Many previous claims of high-temperature or room-temperature superconductivity have failed to hold up under scrutiny.
I am sure we will (have to) hear about this startup for some time as the AI wave continues, we'll see if the claims hold up to actual use and commercial interest.