Bubbly Solutions
Who knows if dice rolls, paint splashes, yoyo swings, or spinning tops can represent a higher-purpose problem?
We can often use analog physics to solve computing problems. The earliest example is the millstones in 17th-century European windmills. James Watt later used the same idea to regulate and "smooth out" the steam engine. This loop.
- If the steam valve bursts open, the turbine spins faster.
- As it spins faster, the balls swing out.
- That closes the valve more and more, cutting the steam.
- Turbine slows down.
Since then we've used many simple physical objects, lives, and phenomena to solve "higher-purpose" problems.
Say you need network cable across a new town with as little wire as possible. Classic Minimum Spanning Tree (MST). At least linear time, even when you already know every pairwise distance.
There's an O(1) way, though. Soap films. Dip the points in soapy water and surface tension makes the membranes always take the shortest path. Physical and natural phenomena like this often hand us an extremely efficient or reliable way to solve "grand" optimization problems.
Recently, Extropic announced that it is "building a full-stack hardware platform to harness matter's natural fluctuations as a computational resource for Generative AI."
How can we design a complete AI hardware and software system from the ground up that thrives in an intrinsically noisy environment?
Pretty exciting idea. Their claim is that biology computes efficiently with noisy, random processes. Energy-based models (EBMs) fit low-data regimes well, but they need sampling, which is hard on digital hardware. So Extropic is building analog circuits that run EBMs directly and claims big gains in runtime and energy efficiency over digital computers. They start with nano-fabricated superconducting devices at low temperatures, plus room-temperature semiconductor devices for a wider market. Extropic is also building a software layer that compiles abstract EBM specs into the hardware control language, so big programs can split across analog cores.
But many people raised concerns
- Analog circuits for AI acceleration aren't new. IBM and Intel have been experimenting for years.
- Analog circuit design is much harder than digital VLSI design, with mixed-signal issues, nonlinear behavior, and device-level problems.
- Even with a great team, yield and variability are real problems in stochastic circuits. Every chip behaves differently, so consistent performance is hard.
- Verifying and testing these circuits needs new EDA (Electronic Design Automation) flows and methods that don't exist yet.
- Getting to production will be tough. It means building a new stack from scratch and competing with a mature, heavily optimized digital silicon industry.
- It's suspicious that Extropic hasn't released any public software simulations or benchmarks, a basic due diligence step in hardware.
- Superconducting Josephson junctions may give some gains in a lab, but they need cryogenic cooling, which makes mass-market packaging very hard.
- Is the sampling speedup really the bottleneck to AGI?
- Maybe Extropic is riding the hype around esoteric approaches to justify its valuation.
It poses an interesting question of knowledge on how information is entangled throughout the universe. Perhaps in the near future we will have dedicated units that run a tiny physics experiment inside, record the result, and extrapolate it into a more human-oriented, useful interpretation.
Say your CPU has a dedicated Soap Processing Unit. It blows foam bubbles, reads the coordinates of every intersection, and magically solves MST in O(1).
Who knows if dice rolls, paint splashes, yoyo swings, or spinning tops can represent a higher-purpose problem?
What if the way the spinning top oscillates and swivels represents a spaceship's optimal trajectory?
No backlinks found