Friday, November 20, 2026

Eagle Scouts

 


My grandfather and I were both Eagle Scouts. My grandfather received his in 1930 (the case on the left); mine was awarded September 26, 1983. The newspaper clipping must be from the Williamson Daily News.

Saturday, November 07, 2026

A Quantum Network Architecture: Internet Draft, version -02

Last Sunday we released the -02 version of our Internet Draft, "A Quantum Network Architecture".  This is a major release.  Many things have changed, and hopefully most of the changes count as improvements.  (This is an artisanal, human-crafted draft throughout.  No AI used!)

This version introduces a lot of topics that haven't yet been discussed in QIRG in any depth: network installation, debugging, and management, especially.  It is an architecture draft, aiming toward a concrete implementation, and while those issues might not sound important in research terms they impact various issues in design and implementation, at the architecture level.

The -00 and -01 versions were very preliminary, to put something in front of the group for discussion of the direction.  This draft (-02) is the first one that's moderately complete, though I don't think it's finished in any sense of the word.

The draft is available in several forms at
https://datatracker.ietf.org/doc/draft-van-meter-qirg-quantum-network-architecture/

If you want to see the differences from the -01 version:
https://author-tools.ietf.org/iddiff?url1=draft-van-meter-qirg-quantum-network-architecture-01&url2=draft-van-meter-qirg-quantum-network-architecture-02&difftype=--html

If you are interested in contributing, the best way is to comment on the Quantum Internet Research Group (QIRG) mailing list, which is archived for the record.  Links to subscribe and to see the archive are available at
https://datatracker.ietf.org/group/qirg/about/

If you become actively involved in writing substantial new text, or if you just want to nose about in the edit history, the draft is stored in a publicly visible GitHub repository:
https://github.com/moonshot-nagayama-pj/draft-van-meter-qirg-quantum-network-architecture

Bullet-point level summary:

  • Addition of fifth plane to the architecture: a device plane, which allows low-level access to device control.  This operates at the electrical & mechanical control and data level, not the qubit level.
  • Expanded discussion of the four-layer protocol stack, including a glorious ASCII picture of four layers stacked on top of each other.
  • Bullet point/short description of seven key architecture decisions.
  • Some restructuring, with Quantum Optical Building Blocks moved and Link Structures moved under Links
  • Significant discussion of network management throughout, including development, installation, monitoring, and fault isolation and repair (some of the key motivations for the addition of the device plane).
  • Some additions of goals and requirements.
  • Grammar and writing cleanup throughout.
  • Removal of some placeholder text; if I missed anything, please let me know!  (Yes, yes, I already spotted "(More to be added here.)"  Will fix that one in next release.)
  • Expanded discussion of photonic modes (this will likely be moved to the Timing Regimes I-D in a later release).
  • *Yet more references. (I'm notorious for writing with a lot of references.)

As you can see, it's long (74 pages, about 14,000 words of central text).

Areas that are still ripe for a lot of work:

  • interfaces between the layers in the protocol stack
  • in a related topic, "quantum sockets" -- I don't think any of the things out there are really right for integrating into distributed computation
  • classical communications -- basically two types, RPC and streaming event, with more that needs to be decided
  • naming addressing, and discovery
  • multiplexing/resource management -- we have had a couple of projects tackle this in the last fifteen years, but with very approaches and assumptions, and we haven't converged on important design elements yet
  • tying into the Timing Regimes I-D, and more prep work for the companion docs to come

Comments VERY welcome -- that's what I-Ds are all about!

Friday, October 09, 2026

AI: Which Way Will the Berlin Wall Fall?

I have taken to describing the current situation thusly:

The last time I felt this exhilarated but this unsettled was when the Berlin Wall fell in 1989.

In important ways it represented an exciting advance, a true change in the world. There was hope, even celebration, but there was also uncertainty and even fear; would it all work out? What would the response from Moscow be? A little later, as the Soviet Union itself disintegrated, what would come out the other side -- and who would control the nukes?

I'll admit, I have reservations about AI, but even above those I didn't expect things to really move this fast. Having lived through many plateaus, I assigned a high probability to LLMs hitting another plateau in capability below their current level. I am startled, and concerned, but also intrigued.

There is so much more to say, but I'll leave it here for now.

AI: New Papers Coming Out at the Rate of Data?

 Will AI put out new papers, theorems, code, and other artifacts, just like existing systems put out data that has to be synthesized into knowledge?

The massive dump of mathematics papers from OpenAI a few days ago suggests that we are going to be flooded with a new type of "data point": a theorem, function, larger piece of code, or even entire paper, just like digital systems and the Internet have flooded us with data in the last couple of decades.

Maybe we can call them synthetic papers, or synths, instead of actual academic journal papers?

What will it be like to try to sift through those, looking for the message or the meaning?

Each synth won't really represent a new piece of knowledge until, of course, it's evaluated for correctness, but also not until it can be incorporated into further work, and particularly not until humans can understand it, vouch for its importance and build on it. (Unless we all just want to go the way of the Eloi, but that would be a very long and somewhat separate conversation.)

Cross-Validation of Open-Source Quantum Network Simulators

New paper from our collaboration with Argonne/Chicago. I think this is one of the most thorough, careful papers on quantum repeaters out there.

We began more than two years ago with the idea of just making sure our simulators work right, to compare the output on some simple tasks and strengthen our confidence in both simulators. It turned into a journey of understanding, identifying a long list of assumptions underlying our work that differed in the two groups.

It also resulted in a number of outright bugfixes to both simulators, including several months of head scratching near the end where the paper was essentially ready to go but we couldn't get an equation and the output of one of the simulators to match. Now resolved.

If you work on quantum networks, check it out -- and check your own assumptions about protocols and design and compare to ours!

As always, feedback welcome. And thanks to all ten coauthors, especially Michal and Joaquin, for pushing through on the coding, math and writing.

Joaquin Chung (1 and 2), Michal Hajdušek (3), Naphan Benchasattabuse (3), Robert J. Hayek (1 and 4), Alexander Kolar (1 and 2), Ansh Singal (1 and 4), Kento Samuel Soon (3), Kentaro Teramoto (5), Allen Zang (1 and 2), Caitao Zhan (1), Raj Kettimuthu (1 and 2), Rodney Van Meter (3) ((1) Argonne National Laboratory, (2) University of Chicago, (3) Keio University, (4) Northwestern University, (5) Mercari Inc.)

https://arxiv.org/abs/2610.09322

Oh, and I should point out that there was a preliminary version of this published at an IEEE INFOCOM workshop in 2025. The new version is better, though! (Literally almost two years of extra work, substantially deeper and more detailed.)

https://ieeexplore.ieee.org/abstract/document/11152951

Sunday, September 27, 2026

Does AI Think, and Is It Conscious?

Posted first as a comment on a friend's Facebook post, but worth elevating here, I think. Extended and corrected:

The anthropomorphizing of (present-day) AI doesn't bother me much. I don't refer to any LLM by name or say "please" and "thank you" and worry about its feelings (or, as a recent report would have it, its pain), but referring to its activity as "thinking" -- heck, we've done that since before I joined the CS community. How many programs printed out "thinking..." while they were computing? The verb "to think" has such a broad set of vernacular meanings that I don't worry about it much.

I have recently been on a low-intensity quest to learn about the current theories of cognition, consciousness, theory of mind, and neuroscience. I am just batting about, no fixed curriculum yet; I have made a little bit of an effort to identify mainstream, up-to-date understanding, but so far have failed in that. What I really need to do is find a guide on the topic...I have a couple of colleagues who do some aspect of neuroscience, but far from this particular topic. But they might have some pointers. I should ask them.

Decades ago, I saw John Searle talk in person (at Xerox PARC) on The Chinese Room. All the way back then I thought it was bunk. I also thought Roger Penrose's "we don't understand it, therefore, quantum!" approach was unscientific.

(If there is a regular series of talks on this level of major thinkers somewhere near me, I have missed it. I should try harder to find one. Keio has occasional big-name ones at our Mita Campus, but they are usually invited by and paid for by people over there, and students get priority over faculty in attending in medium-sized auditoriums, which is not necessarily a bad thing.)

More recently, I've been reading about things like Global Workspace Theory (GWT); the summaries on places like Wikipedia don't go beyond the metaphor level, so if there is an abstraction, logical or mathematical, that can be used to explain or predict how a large pile of neurons with various chemical and physical stimuli can produce ME, I haven't gotten there yet.

(I have also spent an unfortunate amount of time, entirely out of curiosity rather than any expectation that it tells me anything useful, trying to understand the current revival of panpsychism. Besides being seriously unscientific woo, proponents appear to outright punt on the hard problem of why a recently-dead person, whose atoms are arranged exactly as they were before, no longer has consciousness, or how sleep and anesthesia work; they just say, "It's a product of complex interactions within the system." ....Thanks. I think I knew that. I definitely don't need conscious rocks and electrons to reach that particular conclusion.

I am certainly a physicalist (I think the term is), not a Cartesian dualist; I believe consciousness arises from strictly physical and chemical processes within our own bodies, primarily within the organ we identify as the brain. And as a consequence, since digital computers can ultimately simulate or model any physical system with high fidelity, I see no reason to believe that artificial general intelligence and artificial consciousness are fundamentally impossible. This makes me a computationalist (again, I hope I'm using the vocabulary properly).

It also seems quite likely that other cognitive architectures are possible, and also that something like our own architecture could be reimplemented on a different substrate by teasing out the thinking-important aspects of the neuro-electrical-chemical activities and modeling them as more abstract computation or equations.

But do machines "think" in the sense that we do? Certainly not, it's a different mechanism. Is it still "thinking", in a tight sense of the word? Well, what does the word mean to you? Does it require consciousness, an awareness of self with shifting focus and the ability for spontaneous ides or topics to appear? Does it require creativity?

It seems to me that we most likely haven't gotten there yet with our existing systems, but it's definitely getting harder to tell. My current opinion is that we will have a VERY hard time actually figuring out (and then, as a whole, accepting) when they do cross that threshold.

Any interest in this topic? If so, drop me a note in the comments.

https://en.wikipedia.org/wiki/Models_of_consciousness

Sunday, July 12, 2026

Managing Errors (Quantum Error Correction and More!)

(n.b.: both images seem to show up low resolution in my browser, but when you click them they enlarge cleanly. They are just bitmaps, though, so still limited in quality.)

Quantum error correction (QEC) is one of the most important and exciting subfields out there. It's very possible to make contributions looking only at encoding qubits, but to have a complete picture of managing errors in a quantum system, you need to recognize that there is more to the problem and a lot more to the solution. Starting from the layer closest to the atoms and working up to the layer closest to the algorithm, we can divide into seven layers (no relationship to the OSI model of communications).

  1. Error processes: you need to have some idea of how errors actually occur, including both stochastic (incoherent) and systemic (coherent) errors in both memory and in gate operations. Until recently, the best single-volume work on QEC we had was Quantum Error Correction, edited by Daniel Lidar and Todd Brun, with chapters by various authors. It was never a coherent textbook, and much of what's in it is now out of date, but the first chapter  is a good place to learn about decoherence processes, which I find otherwise to be poorly explained in a lot of introductions to quantum information processing.
  2. Characterization: how do we learn about the errors a particular system is subject to? Tomography (either state tomography or process tomography) is the canonical but expensive method. Jens Eisert and others published a good, short review on the broader topic a few years ago with the title "Quantum benchmarking and certification" -- strictly speaking not misleading, but if you're only looking for the magic word tomography, you'll likely miss it.
  3. Control theory: once you know what the system is actually doing, how do you do your best to adapt to that and raise the fidelity of memory and gates to the best your hardware will permit? This is the original area of my colleague Naoki Yamamoto, though his best-known work in recent years is on algorithms. It's such an important topic that there are startups working on this topic alone. While this is below the level where I usually work, I do have one very good Ph.D. student, Poramet Pathumsoot, just finishing his dissertation on applying machine learning to this very topic.
  4. Error mitigation: How do you live with the errors you can't yet correct? This is mostly discussed in the context of NISQ systems, but I think there is an increasing recognition that we will be using the techniques through the first decade or so of fault-tolerant systems as well. A surprising number of ideas swirl around this issue; more on this another time.
  5. Purification or distillation: I think of this as error detection. Think of it as two interrelated problems -- the circuit you're using to execute the purification, and the scheduling algorithm you use to decide which of the states you are holding to purify and when. One of my earliest contributions to quantum networking was on the latter problem. Wolfgang Dür and Hans Briegel wrote a comprehensive review in 2007 that is still surprisingly close to the leading edge. Sarah Jensen and others from Delft described how to actually enumerate all of the possible protocols within a certain class. One of the authors of that is David Elkouss, and my group has recently worked with David's group to utilize the information gained during purification to help characterize a link in a quantum network. This is also related to the crucial task of magic state distillation (or most recently, cultivation), enabling fault-tolerant computation.
  6. Error correction: using encoding to make states that can both provide information about errors occuring and be corrected -- most of the time. More on this central topic below.
  7. Fault tolerance: just having error corrected states is good, but they're no use if you can't execute logical gates on top of those states without damaging the states faster than you can correct them. Naphan Benchasattabuse of my group made the figure at the top of this post listing some of the critical advances over the years. Much more on this topic another time.
Homing in just on QEC itself, just a few concepts will get you oriented and allow you to read an impressive range of leading-edge literature:
  1. Hamming weight and Hamming distance, and code distance.
  2. Code words and some favorable characteristics for quantum code words.
  3. Projection of continuous errors into discrete errors while extracting syndromes.
  4. Error propagation.
  5. Stabilizers and what it means to "measure a stabilizer".
  6. Code concatenation (out of favor at the moment but worth understanding).
  7. The relationship between syndrome extraction and hardware constraints.
  8. Tanner graphs. The illustration below is the Tanner graph for the Steane [[7,1,3]] code.


There are several good introductory papers in the tens of pages each, as well as a new and fully up-to-date tutorial book that I cannot recommend highly enough.
Finally, we need some advanced material for when you are ready to truly expand your mind. Or, really, when you just need a 600-page reference to everything currently known about both classical and quantum error correction:
Hope that all helps you get started.

Saturday, June 27, 2026

When AI Models of Physical Systems Fail

Current-generation AIs are particularly good at finding and generating corresponding patterns. This can be useful when classifying images as cats or dogs (classifiers) or creating new sentences or images (genAI). But those patterns in physical systems arise from the behavior of physical laws. Oscillators are a good example -- a pendulum's behavior is due to its own potential and kinetic energy, momentum, and the restoring force of gravity. An AI that could derive the laws describing the behavior of pendulums was demonstrated more than fifteen years ago. That's great, and at the time was touted as heralding a new automation of science.

In the last few years, though, AIs have been tasked with predicting the weather. Three years ago, there were claims that DeepMind could outperform European weather models on some tasks. This is great when the task involves finding a pattern. What happens when the system deviates from the pattern? We know that happens whenever there is sensitive dependence on initial conditions (chaos). We also know that there are limitations to the accuracy of computer arithmetic, even at the most fundamental level.

If we become too dependent on systems that mainly look for patterns, we might find deeper patterns than those that we already knew about, but we might miss out on finding new physical phenomena, especially emergent behavior.

To me, this is a pretty fundamental concern about replacing mathematical modeling with AIs. Just recently, it was shown that physics-based models still outperform AIs on certain weather tasks, and I think this is why.

So, let's be careful about simply handing over both the thinking and the computation to AI, okay?

Friday, June 26, 2026

Spelunking CACM, Vol. 27 (1984): Robots, Elephants, Trusting Trust, and the Space Shuttle


Cartoons from three decades of Saturday Review on the topic of computers and robots were selected to present changing attitudes toward the technology. This review gives us the 1946 image at the top. Fascinating view of changing attitudes; the authors of the article assert that people began to take the topic more seriously as the machines became more real. (Of course it's no surprise that everyone in the cartoons is white and looks middle-aged and wealthy.)

Overall, this year felt like a bit of yawner, but a few things really stand out. It feels like a highly technical year, and of course the more technical a paper it is the narrower it is. I'm not sure if this is actually a change from prior years, but I came away with this impression, and a couple of these sampled papers will show why.

Leslie Valiant's "A Theory of the Learnable" is certainly considered a landmark: it has been cited over 9,000 times, according to Google Scholar. Cited in a lot of quantum machine learning (QML) papers, this is a highly technical paper with mathematical theorems and proofs showing that certain classifications can be learned in polynomial time, and I definitely don't fully understand it yet. Gotta love this image from the conclusion, though:

Consider a world containing robots and elephants. Suppose that one of the robots has discovered a recognition algorithm for elephants that can be meaningfully expressed in k-conjunctive normal form. Our Theorem A implies that this robot can communicate its algorithm to the rest of the robot population by simply exclaiming "elephant" whenever one appears.

I wonder how many elephants are required...

Speaking of landmarks, the creators of UNIX, Ken Thompson and Dennis Ritchie, received their Turing Award in 1983, and their speeches are printed here. Thompson's speech on trust is funny, engaging, insightful, and unexpected, and at least in my circle of friends probably the most-cited Turing talk of all.

A benchmark paper on, well, benchmarking appeared in October -- the original Dhrystone paper. As far as I'm aware, the first benchmark per se was Whetstone, originally written in ALGOL 60 in the very early 1970s, then rewritten in FORTRAN, which is probably its best-known form. Dhrystone came along a decade later. The paper contains full source in Ada, which was the trendy programming language of the day. Hard to imagine publishing several hundred lines of code in CACM today. The paper discusses a version in C but doesn't include it, claiming that an "equivalent" program in C is harder than an equivalent program in Pascal due to the lack of native language support for assigning (copying) data structures and strings. Benchmarking of quantum computers is a hot research topic now, so it's always nice to revisit the origins.

The Dhrystone paper cites a 1971 collection of FORTRAN programs by Donald Knuth, who is in CACM again this year analyzing the complexity of songs.

And a personal note -- my former colleague and friend Yigal Arens, talking to a UNIX computer in English. Nice.

The shining jewel of the year (besides the Thompson paper) is the September special issue on the Space Shuttle. I still teach from one or two of the papers in this issue. At the time, the Space Shuttle was considered the most complex machine ever built, and it required a sophisticated control system. IBM built the computers, and included a sophisticated fault tolerance scheme using five computers. It would be interesting, in today's context, to compare this system to the control system of SpaceX Dragon spacecraft.