Panel: The Evolution of Quantum — Annotated Lecture Notes & Key Takeaways

I sat in on a roughly twenty-minute panel Q&A at MIT that turned out to be less a lecture than a conversation among three people who have known each other for decades: a moderator who clearly enjoys needling his friends, and two physicists who spent their careers wandering across fields before landing in quantum computing. The title promises an account of how the field evolved, and it delivers on two levels at once — how the technology got here, and how two very non-linear careers did.
 

Two Physicists Who Kept Changing Fields

The moderator opened with biography rather than physics, and it set the tone. He'd only learned that day that Jay grew up in what he called "the Bronx of Australia," and he sketched Jay's route from theorist working on quantum foundations — the question of why quantum mechanics describes the world at all — to running a large superconducting hardware program. Hartmut, meanwhile, has a physics PhD like Jay, but detoured through neuroscience, vision and AI before arriving at quantum. The moderator described the lab Hartmut now leads as one of the country's storied research institutions, the place credited with hard disks, floppy disks and magnetic storage. (The intro got a little tangled about which institution did what, so I'm reporting the framing rather than vouching for it.)

Then the first real question: what would you tell a student who wants a career as unusual as yours?

Hartmut's answer was about timing. He likes entering fields well before they're fashionable, before everyone piles in and the competition becomes suffocating. He was doing what we'd now call self-driving cars in the 1990s and AI in the early 2000s, and he described those areas as most enjoyable when they sit just past the edge of where mainstream attention is. There's a second claim buried in there that I found more interesting: big questions are easier to bet on than short-term ones. It was reasonable, in the nineties, to predict that self-driving cars would eventually work; it would have been much harder to name the hot topic two years out. The moderator compressed the whole answer into a single word — fun — and Hartmut accepted it without objection.

Jay's answer was about a different kind of appetite. He said he dislikes the theorist-versus-experimentalist split, and his rule has been to pick whatever he doesn't yet understand and jump into it. His example was concrete: he started out firing lasers at atoms and measuring cross-sections, then realized he didn't understand the equations underlying what he was measuring. So he approached Howard Wiseman, a quantum optics theorist, and asked to switch to theory. Wiseman took him on. Then Jay decided he didn't actually want quantum optics — he wanted to build a quantum computer — and Steve Girvin at Yale took an optics person and let him learn superconducting qubits and condensed matter from scratch. The advice that comes out of that story is blunt: don't be deterred by difficulty, get good mentors and talk to them, ignore people who tell you it doesn't make sense, and deliberately take on the hardest problems you can find.

I'd flag one thing as a listener. These are both survivorship stories. Neither panelist says anything about the people who jumped into the wrong field at the wrong time and paid for it, and "find a mentor who gives you a chance" is partly a story about luck in who says yes.
 

"Rational Exuberance" and Quantum's ZX81 Problem

The moderator then went after the obvious charge: isn't quantum wildly overhyped, and isn't "quantum AI" worse?

Jay's response was to concede the popularity and deny the conclusion. Yes, the field is crowded, but crowding doesn't use up the good questions. And there's a harder fact underneath: a fault-tolerant quantum computer does not currently exist. That absence is itself the argument — an unfinished machine means unfinished work. He added that even after such a machine exists, the algorithm side stays thin; there are famous algorithms already, and many more waiting to be discovered.

Hartmut reached for a phrase he'd picked up from the economist Mohamed El-Erian: rational exuberance. The moderator misheard it as irrational, and the correction came back immediately — rational. The reasoning was historical. Railroads and the early internet weren't hype in the sense of evaporating; they were real, they needed enormous capital, and some of that capital was destroyed by people who arrived too late. AI, he noted, is something most of us now use daily. Overinvestment certainly happens at these moments, exactly as it did with railroads, but that's a different failure than a bubble built on nothing. His verdict was that quantum today sits in that same category, and the field deserves the celebration.

Then the moderator turned the screw, and this was the sharpest stretch of the session. Honestly assessed, he said, today's machines aren't economically useful — they don't beat what you could have done without them. He compared the current Google and IBM hardware to the Sinclair ZX81 and the BBC Micro, those early home computers whose descendants were utterly invisible from the vantage point of 1981. His question: is our moment like computing right before Bob Noyce and Intel, and are you two the Noyce figures of this generation, or is that person sitting in the audience?

It didn't land. One of them heard "noise," asked what the moderator meant, then asked when Noyce was founded, and finally said he hadn't understood the question. The Noyce comparison was never answered. That's a genuine gap in the discussion, not a small one — the analogy between 2020s quantum hardware and 1970s silicon was the most interesting thing anyone said, and it died on a mishearing.
 

Chasing Quantum Advantage: Echoes, Trackers, and a Commodore 64

Jay did eventually push back on the economic framing, and his distinction is worth holding onto. If the test is an algorithm that returns money to an enterprise, he agreed the machines aren't there. But he offered a second test: doing science on devices we can't easily simulate. That, he argued, is a legitimate scientific question in its own right, independent of whether a CFO would pay for it. He openly invited disagreement about what "advantage" should mean and about whether empirical demonstrations count or whether everyone should just wait for fault tolerance. And he stressed that the space for students to invent new numerical and heuristic algorithms is essential — this is not a field where the algorithmic map is drawn.
 

Hartmut's echoes and the NMR test

Hartmut's concrete example was the quantum echoes algorithm his team has been running. The algorithm is ready, he said; it's been demonstrated on small cases; and it can extract properties of small molecules that nobody currently knows. The moderator interjected with a teasing "in principle" / "in practice," and to his credit Hartmut took the point rather than dodging it: for the actual end use, you could still get the same answer from a classical machine today.

What has to change is computational volume, meaning lower error rates, before a working NMR spectroscopist — someone with no particular interest in quantum computing — would pick a quantum box simply because it's the better instrument. His estimate for that was around two years, not much more, and the reasoning was a threshold argument: once a molecule has more than about fifty spins, classical simulation gets very expensive, while their demonstration sat at fifteen. So there's a real gap, but you don't need the largest machine anyone has on the roadmap to close it.

He also said, plainly, that he isn't an NMR expert and can't predict whether that community will ever adopt this as a standard protocol. I appreciated that. It was the most honest sentence in the room.
 

Jay's tracker and the pace of iteration

Jay's route is different from Google's, and he pointed students at a public quantum advantage tracker to follow it. His main methodological complaint is about speed: he wants these early scientific demonstrations to move from slow, paper-mediated cycles to something like rapid iteration, where a claim goes up and the community responds in weeks.

He said the community has identified at least three routes toward advantage, and he singled out Scott Aaronson's "peak circuit" framing — hide a signal, then find the peak. The story he told is a miniature of how the whole field currently works. IBM found the peak faster than a tensor method. Someone from AI then produced a black-box algorithm that found it roughly a hundred times faster still. Quantinuum and BlueQubit subsequently posted results finding the peak faster on actual quantum hardware. Open question: can the AI method beat that too?

A second example ran the same way. IBM did observable estimation on circuits similar to Google's, aimed at correlated dynamics, and identified observables that Flatiron tensor methods couldn't simulate — the simulators conceded the point. Then a student at Madison, Wisconsin, ran just a single iteration instead of the full dynamics, stripped out much of the interference, and showed it could be simulated after all. Jay's reaction was notably unbothered: that's science, let it run, and it's more entertaining than usual because it's a live rally with the rest of the community.

The moderator then supplied a story Jay had skipped, and it's the one I'd remember longest. IBM had published a much-admired paper using quantum computers to address a fundamental question about the Ising model. A student then reproduced the same result on a Commodore 64, with no quantum computer involved. IBM kept iterating anyway, and today there's a public advantage board that shows IBM ahead. Jay cut in mid-story: "none of them show advantage," he said — the honest framing is "advantage quantum," a direction rather than a scoreboard. That exchange is the single most useful calibration device in the whole session.
 

Trillions or Billions? The Money Question They Both Rejected

The moderator called this the diciest question and aimed it at Hartmut: getting to real, return-on-investment advantage is often said to require silicon-scale money — trillions over a decade, not billions. How does the field get there?

Hartmut declined the premise. Expensive, certainly, but he put the order of magnitude in the billions. Then he complicated his own answer in a way I found revealing. Both he and Jay run superconducting programs, a decision made for good reasons, but Silicon Valley runs on "only the paranoid survive," so they spend time watching other modalities. Misha Lukin had been sitting in the room; QuEra had come up earlier; neutral atoms have made impressive progress. Google invested in QuEra precisely to have a front-row seat on where that technology goes. Hedging costs money, he said, but it doesn't get you to trillions.

Jay's answer located the cost question in design philosophy. Keep funding smart students who work on codes, he said, because the size of the bill depends on whether you have to reinvent the CMOS industry or can ride on top of it — or, in Lukin's case, ride on the existing laser and control industry. He chose superconducting qubits partly because the microwave industry is already enormous: phones, radar, CMOS everywhere. Fit your design inside infrastructure that exists and the number stays in the billions. Then came the twist I keep thinking about — maybe the first generation of quantum computers will mainly teach us how to build the second one, and that's when the figures could climb into the trillions. He still refused the fatalism that we won't have a fault-tolerant machine by the end of the decade. The moderator's synthesis was fair: both men are arguing for cleverer routes rather than more force, which is presumably why they were saying it at MIT.
Here's my reservation. Neither of them answered the question that was asked. They were asked about the aggregate investment an entire industry needs to clear the hump; both answered about the cost of their own program. Those aren't the same figure, and Jay's own remark about first-generation machines teaching us to build the next ones concedes that the curve may steepen later than the current budget cycle.
 

Peripherals, Fridges, and the Parts Nobody Owns Yet

The last substantive question was the most fun: forget new modalities, better cryo controllers and better qubits — what peripheral would make your life easier? The moderator offered his own nostalgia, the cassette tape on a TRS-80 Color Computer that let him load his programs, and observed that there's no quantum equivalent.

Jay gave three. First, quantum information science that extracts more from whatever hardware exists — better error correction, algorithms that accomplish things more simply. Second, treat quantum as a subroutine inside classical computing, and actually merge the HPC and quantum communities instead of running them as separate tribes. Third, and this was the one he clearly worries about at night: his group has a picture of ten dilution refrigerators linked together, and if one of them fails you have to cool the entire system back down. You can imagine sitting there cycling fridges forever and never completing a run. Breaking into a millikelvin vacuum chamber is close to impossible, but space physics has shown it can be done at 4K. The peripheral he wants is an interface that lets you move things in and out without wrecking the coldest stage. Solve that, he said, and you strip a great deal of risk out of the big demonstrations — after which the imagination can run: coupling different machines together, feeding quantum data directly into a quantum computer, algorithms that assume quantum inputs, linking his system to Lukin's or Vladan Vuletić's.

Hartmut endorsed all three and added the systems view. This is a full stack with an enormous number of components, and even a well-funded organization like Google cannot explore a trade-off space that large. His example was materials: they still use aluminum, tantalum has produced strong results elsewhere, and they'd like to try it — but there are only so many things one group can test. That, he said, is where universities come in. He also noted that error correction is moving: the surface code, dominant for years, is now being challenged by smarter codes, which should yield more effective machines. The moderator's closing observation was that theory is reaching engineering practice unusually fast, largely because industrial groups are pushing it.

Then the parting lines. Jay's first attempt was a cheer — go quantum, go AI — followed by a joke about abandoning all of it for neuroscience. His serious version was better: don't assume everything is done, come in and challenge it, good ideas will always emerge. Hartmut's was the crowd-pleaser: you'll have fun, and probably do well financially, in quantum AI.

Two things I'd want a second pass on. The Noyce analogy never got its answer, and I'd have traded a fair amount of the tracker discussion for it. And the claim that the surface code is being displaced by smarter codes arrived as a single sentence with no names, no numbers, no example — for a room full of students who might want to work on exactly that, it deserved thirty more seconds.

What I took away is a strange emotional combination. The people closest to the hardware are the most eager to say the field is real and the most reluctant to say it has arrived: no fault-tolerant machine yet, no advantage on the board, maybe two years to a first genuinely useful scientific result, and a funding picture measured in billions only if the cleverness holds up. That mix of confidence and refusal to overclaim is rarer than it sounds, and it's the most useful thing a student could take from the panel.