In other words, gradient descent isn't good at combinatorial optimisation. I'm sure the research is better but the hype in the blog post leaves a bad taste.
There must be a version of Rich Sutton’s Bitter Lesson that applies to alternative computing like this, along with all the other exciting specialised hardware we've seen come and go over the years, like expert systems, optical computing, neuromorphic computing, etc.
Something like:
General purpose commodity silicon with rapidly evolving software generally beats specialised hardware.
Software is just so much faster to iterate and improve than hardware. AI is also improving it too (eg AlphaEvolve).
Specialized hardware may give a single, significant improvement that grabs headlines but in the long term, compounding small improvements win.
I don't think they are even referring to gradient descent here. I think they are referring to systems like AlphaEvolve where they use LLMs to give an informed/heuristical guess to try to tackle an otherwise insurmountable search space.
In hardware Prolog/Kanren/expert systems? That would be possible with libre microcode for Intel, and not this spyware corporate shithole we are living it.
We would be able to switch microcode at boot and set one for security, another one for C performance, others for Lisp performance and so on.
> [...] quantum-inspired computing built on CMOS technology [...]
So at the heart of the solution is some FPGA that does something (close to?) quantum computing and that helps exploring exponential search space in somewhat feasible way? Is the gist that we might have stumbled upon a practical application of QC? And if so, what's the secret sauce if not lots of qbits? A new algorithm? Is it just hype?
Can someone that understands quantum computing please comment?
This is not quantum computing - "quantum-inspired" could just as well be used to describe a process like simulated annealing. The problem they are solving here is a problem often used as a benchmark for quantum computing, but the approach is purely classical.
So this isn't quantum computing (in the qubit sense), but instead a different computer architecture (demonstrated on an FPGA) that's based on Fowler–Nordheim (FN) quantum tunneling (a real physical effect, used in flash memory, but simulated here).
From the paper:
> The FN-dynamics may be realized either by a physical FN-tunneling device or via a digital emulation of the FN-tunneling dynamical systems. In this work, we employ the digital emulation to achieve the precision required for simulated annealing in the low-temperature regime.
With a "real" (read: analog) FN device, you potentially get large speed ups and even larger cost/energy savings, because the physics is essentially working for "free" -- that's the quantum part.
What's unclear is how scalable the autoencoder architecture would be with analog FN devices today (though the authors suggest this a positive, as it lends itself to 3d architectures as we're hitting the limits of planar scaling).
This is not especially related to quantum computing. Neuromorphic computing uses an algorithm that tries to replicate how the brain works and then in this case implements it and runs it on an FPGA. There are quite a range of papers on this concept and multiple companies are doing just this to show their work. It is often used as it should theoretically avoid such a brute force approach.
I had a long ELI16 session with Claude about it, and the way I understand it is that they
- use Ising machines to describe a certain problem into clauses, storing system state (e.g. spin of something) in variables
- then use a neural network layer where each neuron determines the value of one clause
- then for each state item, use the neuron output to determine if flipping that state would improve the overall system score
- and then use FN-like "noise" to determine whether to flip or no
If the energy landscape of the problem is pretty local, this is guaranteed to find a good solution to the system, using way less compute than brute-forcing it.
They have replicated a neuromorphic algorithm (brain like) on a FPGA, but this implementation at this scale is doubtful to have any improvement over a brute force effort. Quite a few companies feel this is the way forward, although the end goal would be potentially better using photonic chips than qubits and obviously better than an fpga.
The title is especially buzzword based with minimal meaning for the actual paper.
In other words, gradient descent isn't good at combinatorial optimisation. I'm sure the research is better but the hype in the blog post leaves a bad taste.
There must be a version of Rich Sutton’s Bitter Lesson that applies to alternative computing like this, along with all the other exciting specialised hardware we've seen come and go over the years, like expert systems, optical computing, neuromorphic computing, etc.
Something like:
Software is just so much faster to iterate and improve than hardware. AI is also improving it too (eg AlphaEvolve).Specialized hardware may give a single, significant improvement that grabs headlines but in the long term, compounding small improvements win.
I have Bruce Sterling’s Ascendaries: The Best of Bruce Sterling” and… the reality is somewhere here in his stories…
Or take Charles Stross and his Accelerando book.
Do you think that teams behind such projects are avid readers and just fulfill the sci-fi stories? :)
We would be able to switch microcode at boot and set one for security, another one for C performance, others for Lisp performance and so on.
So at the heart of the solution is some FPGA that does something (close to?) quantum computing and that helps exploring exponential search space in somewhat feasible way? Is the gist that we might have stumbled upon a practical application of QC? And if so, what's the secret sauce if not lots of qbits? A new algorithm? Is it just hype?
Can someone that understands quantum computing please comment?
From the paper:
> The FN-dynamics may be realized either by a physical FN-tunneling device or via a digital emulation of the FN-tunneling dynamical systems. In this work, we employ the digital emulation to achieve the precision required for simulated annealing in the low-temperature regime.
With a "real" (read: analog) FN device, you potentially get large speed ups and even larger cost/energy savings, because the physics is essentially working for "free" -- that's the quantum part.
What's unclear is how scalable the autoencoder architecture would be with analog FN devices today (though the authors suggest this a positive, as it lends itself to 3d architectures as we're hitting the limits of planar scaling).
...
Crickets
...
[0]https://www.nature.com/articles/s41467-026-71937-4
I'm only commenting on the title. I like their work.
Is there some code or results from experiments where we can see the speed up?
[0]https://github.com/aimlab-wustl/NeuroSA-HO
Can't compute.
Help.
- use Ising machines to describe a certain problem into clauses, storing system state (e.g. spin of something) in variables
- then use a neural network layer where each neuron determines the value of one clause
- then for each state item, use the neuron output to determine if flipping that state would improve the overall system score
- and then use FN-like "noise" to determine whether to flip or no
If the energy landscape of the problem is pretty local, this is guaranteed to find a good solution to the system, using way less compute than brute-forcing it.
The title is especially buzzword based with minimal meaning for the actual paper.