Pinned post
Now that I have a captive audience…(welcome new followers!) a monster thread on what my paper says, the approximations and the caveats… (1/aleph)
View quoted post on X
@sineatrix
Pinned post
View quoted post on X
In 2009, Kontorovich & Lagarias called it "scandalous" that no one could prove infinitely many Collatz numbers take at least the average time to reach 1.View quoted post on X
New preprint: half of the numbers that reach 1 do. Their stopping times are Gaussian. zenodo.org/records/230468…
Researchers show a thermodynamic system can generate images without a single neural network call.View quoted post on X
It's called Generative Thermodynamic Computing.
Right now, every image generator works the same way. It uses massive, energy-hungry neural networks to mathematically "denoise" static until a picture appears.
It requires billions of parameters. And millions of dollars in GPU compute.
But this paper proved that is completely unnecessary.
They replaced the neural network with a physical thermodynamic system.
Instead of writing code to calculate the denoising, they just let physical matter evolve over time according to the laws of physics.
The information isn’t stored in software weights. It is encoded directly into the physical dynamics of the system itself.
But here is the craziest part. How does it learn?
It minimizes heat.
They trained the system by forcing it to reverse a trajectory from noise back to structure. In physics, doing this efficiently means producing the absolute minimum amount of heat.
By simply optimizing for the lowest heat emission, the system naturally learned to generate perfectly structured data from pure noise.
Zero neural networks. Zero artificial noise injection. Zero active software control.
It just naturally synthesizes structure out of chaos, exactly the way physical matter forms in the universe.
If this gets built into analog hardware, the implications are staggering.
You wouldn’t need software to run a generative AI. You would just turn on a physical chip, let it naturally relax, and watch it generate data.
We have spent billions trying to make software simulate reality.
It turns out it's infinitely cheaper to just let reality compute the software.
View quoted post on X