A Closer Look at LightSolver’s Laser-Based Processing Unit (LPU)

LightSolver are shameless charlatans and scammers of the highest order. We may one day see both founders in jail for defrauding their investors.

Let's just look at two of their false claims:

> Taking on the “3-Regular 3-XORSAT Challenge,” LightSolver solved the problem in polynomial time, while all other state-of-the art classical and quantum computers solved it in exponential time.

That is absolutely 100% false in numerous ways.

First of all, they solved this in a simulation, not on their actual LightSolver hardware. So that means that classical computers can solve this problem in polynomial time just fine. Their hardware is totally superflous.

That's not at all surprising because 3R3X is solvable in cubic time with Gaussian elimination or conjugate gradient methods! It's a thorny problem for a class of optimizers, but it poses no problems in general.

Everything they say is just a lie. For example "Faster than TensorFlow". Then you click, and it's a paper comparing their method to one particular max-sat solver implemented in TF, RUN-CSP. And they implement a totally different approach, again, as a simulator on a classical computer. Which then turns out to be faster (not surprising, RUN-CSP is not state of the art in performance, nowhere near). In no sense does this mean they're faster than TF. They're just lying.

Hm, they seem to be making huge theoretical claims: it seems they're claiming it outperforms classical Turing machine methods by an exponential (non-polynomial) factor. I think this would be a violation of Church-Turing thesis and in of itself quite revolutionary. It's odd this conclusion didn't receive more attention.

On the surface, since they're not using full quantum coherence properties, it seems this should be possible to simulate in a classical computer just fine (in "polynomial time" in some sense) which make it unclear to me why the exponential speedup should exist.

Nonetheless, I think this kind of machine could be useful and more efficient even without any exponential advantages (which to be frank I'm dubious of).

> In three recent trials testing the performance and accuracy of the new LPU, the device showed promise against its supercomputing and quantum computing counterparts in the following ways:

They proceed to cite three not-yet-peer-reviewed manuscripts by the co-founders [1,2,3]. In all three papers, they use a simulation of their LPU (that runs on CPU or GPU), and compare to some classical heuristics on an AWS instance.

I read [1] and the results are... preliminary at best. They compare their simulation to a single deep-learning-based heuristic called RUN-CSP on tiny MAX-2-SAT instances (<350 variables). Deep learning is great but it is generally not SOTA on combinatorial problems. Still, RUN-CSP is not bad, and its authors seem more reasonable; in their paper, they write: "Despite being generic, we show that [RUN-CSP] matches or surpasses most greedy and semi-definite programming based algorithms and sometimes even outperforms state-of-the-art heuristics for the specific problems." Note that RUN-CSP is indeed generic and was by no mean fine-tuned for MAX-2-SAT.

In summary, LightSolver only exists as a simulator on GPUs, that simulator was only tested on very niche problems, and even then probably not against the state-of-the-art. I am trying to be polite here, but to be clear: the BS-o-meter is strongly in the red.

[1] https://arxiv.org/pdf/2302.06926.pdf

[2] https://arxiv.org/pdf/2207.09517.pdf

[3] https://arxiv.org/pdf/2209.03788.pdf

This sounds very similar to the device from LightOn.AI that was discussed here at the end of 2021.[1]

That device used random projection[2] to do the work. Daralthus pointed out in the other thread that it's really useful for dimensionality reduction.

I'm not sure if that's the same operating principle in this device or not.

[1] https://news.ycombinator.com/item?id=29701976

[2] https://en.wikipedia.org/wiki/Random_projection