TLDRAIBBQ: Claude redesigned eight simulated quantum circuits, reducing operations by up to 17× and 12× by reconstructing temporary data and replacing an algorithm with a better one.
"This work was led by Andi Gu as part of Anthropic’s STEM Fellows Program, with mentorship from Alek Kemeny, a quantum physicist at Anthropic, in collaboration with Susanne Yelin (Harvard University) and Mikhail Lukin (Harvard University and QuEra Computing). Read the full paper on arXiv."
"We thank Pablo Bonilla, Arjun Kar, Nazli Ugur
Koyluoglu , Rohan Mehta, Varun Menon, and Vinay Ra-
masesh for feedback on the manuscript. AG conducted
part of this work under the Anthropic STEM Fellows
Program. Work at Harvard was supported by IARPA
and the Army Research Office under the Entangled Log-
ical Qubits program (Cooperative Agreement Number
W911NF-23-2-0219) and by the National Science Foun-
dation (NVQL grant PHY-2410716)."
The whole Anthropic commercial appears to be hallucinated.
This builds on the Kickmix quantum language from Google. Essentially, Kickmix is an assembly language for classically simulatable quantum circuits, and finding more efficient circuits is basically a game of code golf with the assembly instructions. That means LLMs benefit from the closed loop of verification, and can rapidly test and discard ideas. It's not too surprising that previous academics didn't try to optimize to this level, and a good task to automate. See ecdsa.fail for another example of circuit optimization using the Kickmix language.
It's a nice illustration of the Envelope Principle: AI is good at filling in the blanks, where the knowledge is 'within the envelope' of things already discovered but the area itself is still not fully exhausted. Real progress happens by expanding the area of the envelope, not by coloring in the blanks.
People (most people?) have spent their entire decades long careers coloring in the blanks within the envelope of human knowledge. I think it’s kind of shitty to suggest that they weren’t making “real” progress with their lives, just because AI can now do the same thing.
I'm not sure of your background, but FWIW everyone in my PhD program in my elite CS program had varying levels of anguish and angst when realizing that their efforts were mostly suited for "coloring in the blanks." The PhD Grind [1] felt like a poignant capture of those feelings.
this is what research was for the majority of people, even those talented enough to do it. there's a reason we remember the names of people who invent whole ass new fields, like einstein, grothendieck, turing, riemann, etc.
the hope might be that AI will fill in the remaining envelopes and let us focus on building new theories. but that's like trying to build a castle in the sky without being grounded and understanding how castles on earth are built. Imo, it won't work, but maybe I'm wrong. It's not like people need to know how to write assembly these days.
> Real progress happens by expanding the area of the envelope, not by coloring in the blanks.
Respectfully disagree. What one person may classify as expanding the envelope, another one with broader knowledge may classify as coloring in the blanks. Much knowledge in one field comes from cross-pollination from other fields.
Let me give you an example I just learned about. The database community spent decades trying (to some extent) to solve the implementation of materialized views to arbitrary complex SQL queries. Then, a few years ago, someone with experience in digital signal processing brought their expertise to the DB field and completed the puzzle. [1]
Someone with DB expertise but no signal processing expertise might say that is expanding the envelope, but another person with experience in both may say it is coloring in the blanks.
AI work now may seem like coloring in the blanks, but given it's vast expertise in multiple fields, my bet is that for most people it will look like pushing the envelope very soon.
That is a very interesting example but it serves very much to illustrate my point: this is latent knowledge, we effectively already have it but we just haven't put 2 and 2 together. Unfortunately AI excels at such problems because it oversees a much larger slice of our knowledge than most individual humans. I've always said that cross domain application of knowledge is the easiest way to make progress and I haven't really seen anything conflicting with that to date.
So while you start off with 'disagreement' in fact it appears we agree far more than we disagree. Yes, it is important that this work is done. But you clearly illustrate that the envelope did not in fact expand. But all the time scientists around the world do in fact expand the envelope of our knowledge and one of my fears is that by undercutting a lot of such knowledge work by putting this digital combine to work on it that we will lose the infrastructure to do the rest. We'll see a lot of progress in a very short time and then we'll find out that because we took away that class of problems that it has become almost impossible to keep the infrastructure, incentive structure and educational paths in place to be able to do the next round.
You see a similar thing happening in software development. All of a sudden it has become a lot harder to start as a career. And lots of people are not even going to try this direction if there is no first rung on the ladder.
At this point, it’s better to ponder a question: Are there fundamental reasons why a search trajectory during reinforcement learning can’t leave the envelope? And: what is the envelope, anyway? A metaphorical surface in a metaphorical space of ideas with no defined procedures to make it into an actual metric space?
Reinforcement learning can definitely (in principle) leave this envelope, but it depends crucially on whether the feedback generates new information from reality, rather than just repackaging what humans have already discovered in the form of RL environments and densified rewards. So the question is not about RL per se, but what is the information source, what is the bandwidth of information it supplies, and then only finally the effectiveness of RL in distilling information that into models.
> In other cases, it found that existing circuits included costly components that were needed to handle only extremely rare inputs, and it redesigned them without those components.
Why the ellipses? It's not like they're sweeping this under the rug, it's pretty explicitly stated for which optimizations this applies. It's a perfectly standard optimization strategy.
Their second example features it for example, but as an additional optimization that saves memory. The compute optimization part doesn't need this.
We have excellent approximations of the full quantum chemistry that work really well, e.g. DFT. It's the gold standard for chemistry today. But these edge cases, although rare, do show up. They have a nasty tendency of showing up in biology especially, where reactions often happen on a knife's edge of the energy landscape, for efficiency reasons.
Chemists are interested in quantum simulation because it would be fully accurate and be an actual substitute for ground-truth experiment. With these approximations, you still wouldn't know if what you're simulating matches reality.
i just skimmed the article, but it does sound like it would be detectable if the case you had is one that the algorithm doesn't handle. In any case, it sounds like it is very rare that you have one of those cases vs existing heuristics, which is a simplification every time.
Typical quantum simulation papers do not target 100% accuracy, they target "chemical accuracy [1]". That said, chemistry simulations do tend to be less tolerant of approximations because you are sampling an unknown distribution and so can't easily tell if a shot had a failure. But the linked article is about a subroutine for Shor's algorithm, where it is very obvious if you succeeded or not because you either recovered the private key or you didn't.
Yes strictly speaking it's a different computation, but the differences are so small that swapping between then won't meaningfully affect the outcome of Shor's algorithm.
This is one of the most important things to realize when optimizing quantum algorithms: approximations are overpowered. The difference between 99.99999% accurate and 100% accurate can be a factor of 10 in cost. And there's lots of places where a 90% accurate algorithm is more than good enough.
(warning: pile of AI) I tried autoformalizing my advisor’s decades of papers in https://github.com/emberian/graphplay he studied “Grover-like” problems that evade some of the no-go theorems. these AI are a lot better at connecting domains and ideas for algorithmic insight than we have previously had available (the state of the art being not much better than Llull’s circles). I expect we’ll see a big burst of early progress.
I haven’t started yet, but modeling various diffusion and other operators inside the framework might find some Grover-speedups to some ML primitives of interest.
> A new reversible algorithm
>
> Designing a quantum circuit involves two constraints that ordinary programming doesn’t have. The first is that a quantum circuit can’t make “if-then” decisions the way an ordinary program can. An ordinary program can check a value and then run only one of two pieces of code. But a quantum computer runs on a superposition of many inputs at once, some of which would run the first piece of code and some of which would run the second, so the circuit has to include the operations for both every time it runs.
If you squint, this looks like SIMD programming (for a quick introductory reference, see [1]) where there is no branching logic in the lowest level. Looking from far away, it seems like high-performance programming is converging from these two points, isn't it? One with the current classical hardware, the other with the new quantum one.
No, it is not like SIMD. It's more like an ALU, in that there needs to be gates present for each possible operation (and then muxers/demuxers route through the correct one).
Something that is like SIMD in quantum computing is that if a block code supports a transversal operation then it is far cheaper to apply that operation to every qubit in the block than to selectively target one of them with the operation.
37 comments
[ 0.27 ms ] story [ 95.0 ms ] threadThis is the level where the AI develops quantum computing to make itself smarter (in the game)
"This work was led by Andi Gu as part of Anthropic’s STEM Fellows Program, with mentorship from Alek Kemeny, a quantum physicist at Anthropic, in collaboration with Susanne Yelin (Harvard University) and Mikhail Lukin (Harvard University and QuEra Computing). Read the full paper on arXiv."
The actual paper (https://arxiv.org/pdf/2610.22367):
"We thank Pablo Bonilla, Arjun Kar, Nazli Ugur Koyluoglu , Rohan Mehta, Varun Menon, and Vinay Ra- masesh for feedback on the manuscript. AG conducted part of this work under the Anthropic STEM Fellows Program. Work at Harvard was supported by IARPA and the Army Research Office under the Entangled Log- ical Qubits program (Cooperative Agreement Number W911NF-23-2-0219) and by the National Science Foun- dation (NVQL grant PHY-2410716)."
The whole Anthropic commercial appears to be hallucinated.
[1]: https://dokumen.pub/the-phd-grind-a-phd-student-memoir-3.htm...
the hope might be that AI will fill in the remaining envelopes and let us focus on building new theories. but that's like trying to build a castle in the sky without being grounded and understanding how castles on earth are built. Imo, it won't work, but maybe I'm wrong. It's not like people need to know how to write assembly these days.
This is a pretty shitty thing to say, and if you think about it for even a few minutes, you'll understand why.
Respectfully disagree. What one person may classify as expanding the envelope, another one with broader knowledge may classify as coloring in the blanks. Much knowledge in one field comes from cross-pollination from other fields.
Let me give you an example I just learned about. The database community spent decades trying (to some extent) to solve the implementation of materialized views to arbitrary complex SQL queries. Then, a few years ago, someone with experience in digital signal processing brought their expertise to the DB field and completed the puzzle. [1]
Someone with DB expertise but no signal processing expertise might say that is expanding the envelope, but another person with experience in both may say it is coloring in the blanks.
AI work now may seem like coloring in the blanks, but given it's vast expertise in multiple fields, my bet is that for most people it will look like pushing the envelope very soon.
[1] https://arxiv.org/abs/2203.16684
So while you start off with 'disagreement' in fact it appears we agree far more than we disagree. Yes, it is important that this work is done. But you clearly illustrate that the envelope did not in fact expand. But all the time scientists around the world do in fact expand the envelope of our knowledge and one of my fears is that by undercutting a lot of such knowledge work by putting this digital combine to work on it that we will lose the infrastructure to do the rest. We'll see a lot of progress in a very short time and then we'll find out that because we took away that class of problems that it has become almost impossible to keep the infrastructure, incentive structure and educational paths in place to be able to do the next round.
You see a similar thing happening in software development. All of a sudden it has become a lot harder to start as a career. And lots of people are not even going to try this direction if there is no first rung on the ladder.
[1] https://www.reuters.com/world/anthropic-quietly-sets-up-biol...
So it’s not performing the same computation…
Their second example features it for example, but as an additional optimization that saves memory. The compute optimization part doesn't need this.
Chemists are interested in quantum simulation because it would be fully accurate and be an actual substitute for ground-truth experiment. With these approximations, you still wouldn't know if what you're simulating matches reality.
[1]: https://chemistry.stackexchange.com/questions/62001/why-is-c...
This is one of the most important things to realize when optimizing quantum algorithms: approximations are overpowered. The difference between 99.99999% accurate and 100% accurate can be a factor of 10 in cost. And there's lots of places where a 90% accurate algorithm is more than good enough.
I haven’t started yet, but modeling various diffusion and other operators inside the framework might find some Grover-speedups to some ML primitives of interest.
If you squint, this looks like SIMD programming (for a quick introductory reference, see [1]) where there is no branching logic in the lowest level. Looking from far away, it seems like high-performance programming is converging from these two points, isn't it? One with the current classical hardware, the other with the new quantum one.
[1] https://mitchellh.com/writing/everyone-should-know-simd
Something that is like SIMD in quantum computing is that if a block code supports a transversal operation then it is far cheaper to apply that operation to every qubit in the block than to selectively target one of them with the operation.