I think this is about Ising Computers. I can't judge whether or not the worth of this paper.
But here are some good video introduction for what Ising computers are and how they work by
Aaron Danner :
https://www.youtube.com/watch?v=mD-0VpNSJA0&list=PLXb3r5ny8_...
Ising Computers #1: Introduction
Ising Computers #2: The Number Partitioning Problem
Ising Computers #3: The Max-Cut Problem
It's an alternative way of computing, by setting up physical system, letting them evolve, and looking what state they evolve to.
You are setting problem by defining a system of coupled harmonic oscillators. Statistically (Boltzmann) after a long time it should settle in a configuration of low energy state, where the energy function is defined by the values of the coupling constant you set up.
It has a lot of similarity with quantum computing but none of the weirdness and you can simulate them numerically on standard computer instead of using real hardware to study them.
The whole title is a buzzword cluster, until proven otherwise.
Which tasks, in particular, does it do better? Not as in "it could do them better", but actually there are benchmarks. If they are, they are buried beneath marketing; if not - well, we have our answer.
What is "thinks like nature"? Spin systems, are no more (or less) nature than transistors.
That said, I am all for exploring various systems for computation and simulation - I think there is a lot to discover.
Yeah, I mean it's obviously meant to be a marketing pitch but it's not a very good one.
> The hardest computational problems are not waiting for faster chips – they are waiting for machines that compute in a fundamentally different way.
Surely they don't actually believe that, right? Like you say the benefits must be limited to specific shapes of problems (not all of "the hardest" ones), and the whole history of computing is about how faster chips is an excellent answer to difficult computational problems.
> and the whole history of computing is about how faster chips is an excellent answer to difficult computational problems.
I don't really disagree, and I am definitely not taking their marketing pitch seriously. Yet, you could look at the same computation history and interpret it as an economically constrained hill-climbing around an idea that was simple enough to work reliably (von Neumann architecture) and that worked and scaled so well that we were rarely forced or desperate enough to move conceptually far away from it.
Sufficiently general digital computers can simulate other computational models, so I think 'faster' is ultimately the end game, but for some classes of computation, as you also noted, we may need to go for analog hardware, (maybe) quantum devices, optical interconnects, and so on.
Bret Victor has a talk about this, more or less: [0]
That's interesting, thanks. I only read the abstract so far but was immediately reminded of this recent HN submission[1] and the whole thing that certain ideas go together, and so they are adopted together, but the resulting bundle of ideas might be poorly suited to certain problems.
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.
It requires O(N^4) evaluations to compute Fisher Information Matrix for N-dimensional parameterization of the problem in original formulation. But there are closed form solutions and more economical representations of covariance matrix (LoRA, hehe).
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.
> General purpose commodity silicon with rapidly evolving software generally beats specialised hardware.
All of the Amiga people are sighing right now, as they recall how their beautiful, elegant system synergistically designed with custom chips was outpaced by CPU/memory brute force in the early 90s.
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.
Amusingly the nature paper is also an incredibly dense wall of hype terms but actually appears to have substance. It's like a weird alternate reality where a scam artist attempting to fleece gullible investors took things too far and performed rigorous science.
> [...] 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.
the use of 'quantum' appears to be tagging onto the potential of quantum annealers (which have repetitively [1] [2] been shown to be classically tractable) while trying to mimic a kind of quantum tunneling, ie the annealing schedule, without any kind of promises about exponential speedups etc. Quantum annealers themselves have few promised advantages for general combinatorial optimization problems without significant changes to extant hardware paradigms [3]
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'll have to see if I can find references to an older effort on previous learning algorithm optimizations with FPGAs in the loop - it must be 20+ years old by now. The algorithm did indeed optimize the toy problems that it was setup to optimize, but it exploited non-digital, analog electronic characteristics of the individual FPGA to do it, so the solutions were not portable to any other FPGA - even of the same model.
Edit: There it is, Adrian Thompson evolution of tone generators, 1997.
Seems like the key elements in this are the use of a neuromorphic autoencoder (instead of a 'regular' one), plus Fowler–Nordheim annealing dynamics and Ising energy minimization so that the system is not just passively settling, it's being taken through a controlled search process designed to avoid premature trapping and scale to higher-order combinatorial optimization problems. [1]
A 'regular' autoencoder is a neural network trained to compress data and then reconstruct it.
A neuromorphic autoencoder is instead implemented using brain-inspired computing elements like spiking neurons, event-driven updates, local interactions, sometimes specialized hardware. In this paper, looks like the autoencoder is being used as a structured energy-minimizing circuit for an Ising optimization problem. The architecture manipulates Ising clauses rather than only pairwise spin interactions.
Ordinary artificial neurons compute matrix ops such as y=f(Wx+b), while this uses artificial neurons that accumulate input, which emit a spike when they cross a threshold, like biological neurons (event driven neural dynamics).
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.
I think this is about Ising Computers. I can't judge whether or not the worth of this paper.
But here are some good video introduction for what Ising computers are and how they work by Aaron Danner : https://www.youtube.com/watch?v=mD-0VpNSJA0&list=PLXb3r5ny8_... Ising Computers #1: Introduction Ising Computers #2: The Number Partitioning Problem Ising Computers #3: The Max-Cut Problem
It's an alternative way of computing, by setting up physical system, letting them evolve, and looking what state they evolve to.
You are setting problem by defining a system of coupled harmonic oscillators. Statistically (Boltzmann) after a long time it should settle in a configuration of low energy state, where the energy function is defined by the values of the coupling constant you set up.
It has a lot of similarity with quantum computing but none of the weirdness and you can simulate them numerically on standard computer instead of using real hardware to study them.
The whole title is a buzzword cluster, until proven otherwise.
Which tasks, in particular, does it do better? Not as in "it could do them better", but actually there are benchmarks. If they are, they are buried beneath marketing; if not - well, we have our answer.
What is "thinks like nature"? Spin systems, are no more (or less) nature than transistors.
That said, I am all for exploring various systems for computation and simulation - I think there is a lot to discover.
Yeah, I mean it's obviously meant to be a marketing pitch but it's not a very good one.
> The hardest computational problems are not waiting for faster chips – they are waiting for machines that compute in a fundamentally different way.
Surely they don't actually believe that, right? Like you say the benefits must be limited to specific shapes of problems (not all of "the hardest" ones), and the whole history of computing is about how faster chips is an excellent answer to difficult computational problems.
> and the whole history of computing is about how faster chips is an excellent answer to difficult computational problems.
I don't really disagree, and I am definitely not taking their marketing pitch seriously. Yet, you could look at the same computation history and interpret it as an economically constrained hill-climbing around an idea that was simple enough to work reliably (von Neumann architecture) and that worked and scaled so well that we were rarely forced or desperate enough to move conceptually far away from it.
Sufficiently general digital computers can simulate other computational models, so I think 'faster' is ultimately the end game, but for some classes of computation, as you also noted, we may need to go for analog hardware, (maybe) quantum devices, optical interconnects, and so on.
Bret Victor has a talk about this, more or less: [0]
[0] https://www.youtube.com/watch?v=8pTEmbeENF4
You're describing The Hardware Lottery: https://arxiv.org/abs/2009.06489
That's interesting, thanks. I only read the abstract so far but was immediately reminded of this recent HN submission[1] and the whole thing that certain ideas go together, and so they are adopted together, but the resulting bundle of ideas might be poorly suited to certain problems.
[1]: https://news.ycombinator.com/item?id=48237163
Yeah, I was reminded of the Rockwell Retro Encabulator a little bit.
https://www.youtube.com/watch?v=RXJKdh1KZ0w
> Explores what AI cannot
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.
[1] https://en.wikipedia.org/wiki/Natural_evolution_strategy#Nat...
It requires O(N^4) evaluations to compute Fisher Information Matrix for N-dimensional parameterization of the problem in original formulation. But there are closed form solutions and more economical representations of covariance matrix (LoRA, hehe).
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.
“neuromorphic computer that combines quantum-tunnelling physics with a brain-inspired architecture to find solutions to hard mathematical problems”
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? :)
> General purpose commodity silicon with rapidly evolving software generally beats specialised hardware.
All of the Amiga people are sighing right now, as they recall how their beautiful, elegant system synergistically designed with custom chips was outpaced by CPU/memory brute force in the early 90s.
I'm not sure whether these FPGA codes count as specialized hardware.
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.
> a neuromorphic computer that combines quantum-tunnelling physics with a brain-inspired architecture
This ought to be the most rhetorically compressed, stacked-legitimacy-seeking hype phrase I've ever seen in a tech description.
Amusingly the nature paper is also an incredibly dense wall of hype terms but actually appears to have substance. It's like a weird alternate reality where a scam artist attempting to fleece gullible investors took things too far and performed rigorous science.
> [...] 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.
the use of 'quantum' appears to be tagging onto the potential of quantum annealers (which have repetitively [1] [2] been shown to be classically tractable) while trying to mimic a kind of quantum tunneling, ie the annealing schedule, without any kind of promises about exponential speedups etc. Quantum annealers themselves have few promised advantages for general combinatorial optimization problems without significant changes to extant hardware paradigms [3]
[1] https://arxiv.org/abs/2503.05693 [2] https://arxiv.org/pdf/2507.22117 [3] https://arxiv.org/abs/2008.09913
No it's just analogies. It's a normal FPGA.
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'll have to see if I can find references to an older effort on previous learning algorithm optimizations with FPGAs in the loop - it must be 20+ years old by now. The algorithm did indeed optimize the toy problems that it was setup to optimize, but it exploited non-digital, analog electronic characteristics of the individual FPGA to do it, so the solutions were not portable to any other FPGA - even of the same model.
Edit: There it is, Adrian Thompson evolution of tone generators, 1997.
https://en.wikipedia.org/wiki/Evolvable_hardware
> Can someone that understands quantum computing please comment?
...
Crickets
...
Very well then, could someone that _got used to quantum computing_ please comment?
Higher-order neuromorphic Ising machines—autoencoders and Fowler-Nordheim annealers are all you need for scalability[0]
[0]https://www.nature.com/articles/s41467-026-71937-4
OK, this is just ridiculous now. Cut it with all this "all you need" crap.
I'm only commenting on the title. I like their work.
The unreasonable effectiveness of regurgitated partial titles.
ah that's the actual paper the OP is about! took me a bit. thanks for the link.
Seems like the key elements in this are the use of a neuromorphic autoencoder (instead of a 'regular' one), plus Fowler–Nordheim annealing dynamics and Ising energy minimization so that the system is not just passively settling, it's being taken through a controlled search process designed to avoid premature trapping and scale to higher-order combinatorial optimization problems. [1]
A 'regular' autoencoder is a neural network trained to compress data and then reconstruct it.
A neuromorphic autoencoder is instead implemented using brain-inspired computing elements like spiking neurons, event-driven updates, local interactions, sometimes specialized hardware. In this paper, looks like the autoencoder is being used as a structured energy-minimizing circuit for an Ising optimization problem. The architecture manipulates Ising clauses rather than only pairwise spin interactions.
Ordinary artificial neurons compute matrix ops such as y=f(Wx+b), while this uses artificial neurons that accumulate input, which emit a spike when they cross a threshold, like biological neurons (event driven neural dynamics).
[1] https://www.nature.com/articles/s41467-026-71937-4
I wonder how this compares to thermal wells?
https://extropic.ai
They seem to work in a similar way, sampling from chaotic datasets to find the lowest energy state.
Is one fundamentally more scalable? More efficient?
Computers that don't require cooling - so cool - literally!
This isn’t even a research paper.
Is there some code or results from experiments where we can see the speed up?
Paper is linked on the page (doi.org link redirects to Nature), code here[0]
[0]https://github.com/aimlab-wustl/NeuroSA-HO
There are a lot of buzzwords in there. Does it work?
We should ask Stuart Hameroff for help then.
This reads like the paper from the Sokal affair.
It really does. The verbiage just reeks of gratuitous buzzword grandiosity.
So many ... words... big words ...
Can't compute.
Help.
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.
Yes, I actually believe that if we really want to build AI and physical AI, we need this. I'm working on this for a while. vantar.xyz