The Evolution of Quantum Science — Annotated Lecture Notes & Key Takeaways

These are my notes from the transcript of the QMIT launch event at MIT, written for a friend who couldn't be there. Nothing here is a verbatim record, and where the audio was garbled I say so.

If you missed it, the event was three things at once: a celebration of MIT's history in quantum science, a pair of roadmaps from the two companies with the largest machines, and a genuinely unexpected swerve into the biology of consciousness. Jay Gambetta of IBM and Hartmut Neven of Google Quantum AI each got about fifteen minutes, then were supposed to sit down with Ike Chuang for a conversation — that last part never made it into the transcript I read, so what follows is really two talks plus an introduction.
 

Why QMIT, and Why 1981 Is the Reference Point for Quantum Science at MIT

The host — a senior MIT figure, though the transcript never quite says who — opens by thanking the people who pushed QMIT into existence, including Ian Waites, Anantha Chandrakasan and Nergis Mavalvala. (Small listening note: the host thanks someone called Donna early on and refers to Dana a minute later. Almost certainly one person and one transcription glitch.)

The framing move is interesting. Instead of claiming a founding date, the host offers the summer of 1981 and MIT's first Physics of Computation conference as a "reference point" — roughly fifty researchers from physics and computing in one room, described as the unofficial starting gun for quantum applications. That's a nice rhetorical choice because it sidesteps the usual origin-story argument and quietly claims forty-five years of continuity instead.

QMIT is pitched as built on problem-driven missions, meant to live in shared intellectual space first and shared physical space eventually, with the job of picking the right questions and mapping a route to solutions that are transformative and accessible. I wrote down "accessible" with a question mark, because the talk never says which sense is meant: accessible to non-specialists, or affordable to institutions outside the handful that can build these machines. Those are very different programs.
 

Gambetta's Case: Fault Tolerance Has Become an Engineering Problem

Gambetta says up front that he's skipping IBM's résumé to make two arguments: the community can build a fault-tolerant machine before this decade ends, and it's already time to think about what lies past that. The concrete target is a system by 2029, which the transcript spells as "IBM Starlink" (I'd bet the audio said Starling) rated at 200 logical qubits and 100 million operations (5:04).

He spends most of his time on the "why we're confident" part. First come his criteria for what a fault-tolerant machine even is — framed as a fault-tolerant cousin of the DiVincenzo criteria, but explicitly as guidance rather than commandments. A real architecture where every logical qubit can be addressed; a universal set of operations; an adaptive design so measurement and correction feed forward into later operations; modularity, because he wants to build with the CMOS tooling that exists today rather than wait for tooling that doesn't; and efficiency, meaning as little qubit overhead as possible.

That last criterion drives the whole story. Around 2019 IBM walked away from the surface code and went hunting for something cheaper, hiring heavily from the error-correction student pool, MIT included. They landed on a family of LDPC codes. In 2024 they demonstrated a memory, took the obvious community jab — that's only a memory — and then this year published a full set of operations that stays efficient, plus an architecture around it. The remaining objection to fancier codes is the decoder: can you extract an answer fast enough to be useful? His answer is a real-time belief-propagation decoder, and the assembled framework is described as logical memories, logical processing units and universal adapters wired together in a modular layout. He name-checks Ted Yoder and Andrew Cross as MIT graduates whose work is visible in it.

The section that actually made the engineering feel concrete was about the shape of the thing. The memory is a donut, and superconducting qubits have historically been stuck with nearest-neighbour coupling, so building a donut in a commercial fab means breaking it across several planes and linking across distance. That forces packaging work: bump bonds, multi-level wiring, and the decision to split the qubit across wafers, one for the qubits, one for readout, one for I/O. The transcript renders one component as "cellular conveyors," which I'm fairly sure is a mangled version of through-substrate vias.

Three enabling results follow: coupling preserved out to about 14 millimetres with no visible gate degradation; vertical wiring through a low-loss dielectric so a link can go up a plane and back down; and qubit reset in roughly ten nanoseconds at an error around three in a hundred thousand. Then the systems layer — commercial fab and automated EDA, a comparison shot of the current Heron processor against a new error-correction chip that reportedly looks like older-generation CMOS, very large cryogenic assemblies with desegregated fridges, a target of roughly 3,000 qubits per cryogenic environment, a move from coax to superconducting cabling, cold CMOS control at about one fiftieth the power, and the decoder running on AMD FPGAs within a few nanoseconds.

I liked the recurring line that this is systems work, not physics. It's the most persuasive thing in the talk. It also left me with the number I wanted most: no physical qubit count, no stated physical error rate that 200 logical qubits would require. I'm not saying the claim is wrong, just that the audience was asked to take the headline on faith.

He does push back on his own optimism at the end. Is error correction finished? No. He points at concatenated inner and outer codes, at Google's own "Yo code" (the transcript is unclear here and I couldn't resolve the term), at the idea of tuning codes to the algorithm you actually want to run, and at stacking algebraic codes on top of LDPC. Then post-selection: keep full fault tolerance at the base, but on the outer layer sample across time, run space-time checks, and throw away the bad runs. Two results get cited — 140 logical qubits in what the transcript calls "GHC states" (almost certainly GHZ) with roughly double the improvement over the decoded version, and Clifford-driven graph states with what he calls ridiculous gains.
 

What Comes After Fault Tolerance: Quantum as One Part of a Bigger Machine

Gambetta's second half is about refusing to stop at the fault-tolerant machine. His picture of future computing is CPUs, GPUs and QPUs cooperating, with a neat division of labour: CPUs handle addition, GPUs handle tensors, QPUs handle circuits. He leans on a Dyson line about tool-driven revolutions — new instruments reveal things nobody has explained yet, while concept-driven revolutions just re-explain what we already had.

The chemistry example is where he lands it. Reaction rates depend exponentially on energy differences, so a small error in the energy estimate makes the rate estimate useless; quantum computing promises accurate answers through phase estimation in the long run, and in the near term something more modest — drop the quantum processor into a classical workflow as one subroutine that feeds the rest. RIKEN on iron-sulfur molecules, Cleveland Clinic on something the transcript garbles, Lockheed on methane-related molecules, Boeing on their own applications. His point is that the interesting metric is no longer how much quantum you use but how the classical and quantum portions are mixed.

Then networking. A quantum networking unit that attaches to their best processors through a tunable device; an L-coupler for tying modules together; separate cryogenic environments joined by a switchable microwave channel around 4 kelvin, in a project started with "Fermi" (I'm assuming Fermilab, though the transcript only says Fermi); a transduction unit to convert microwave photons to optical photons, with Caltech and Q-NEXT at Argonne; and eventually microwave photons handed off to flying microwave photons. Of everything in either talk, this is the part I'd most want another hour on — it had by far the fewest concrete numbers attached.
 

Neven's Roadmap: Six Milestones, Quantum Echoes, and a Post-Quantum Deadline

Neven opens with Google's mission line — quantum computing for otherwise unsolvable problems — and a six-step roadmap ending at a million physical qubits, which he expects to yield somewhere between one and a few thousand well-protected logical qubits. They're about halfway.

The historical recap is refreshingly blunt about the messy parts. The 2023 error-correction demonstration, the one that finally answered a thirty-year challenge, was "a little bit anticlimactic": the logical error rate came down by about 4%, and nobody was impressed. The 2024 Willow run fixed that — every step up in code distance cut the logical error roughly in half, and the curve dropped below the physical qubit error rate. He makes the odd but memorable claim that this experiment created "billions and billions of dollars in value," because quantum startups facing delisting suddenly had something to point at. Whatever you think of that framing, it's a reminder that these results move markets, not just journals.

The result he says got less attention is the one he finds most encouraging: protect only one of a qubit's two degrees of freedom and you can push to much larger code distances, which took the logical error rate down to around 10⁻¹⁰. His summary — that all they have to do now is go from this one-dimensional experiment back to a two-dimensional one — is the single sentence in the whole event where I'd most want to interrupt. That "all" is carrying an enormous amount of weight (21:54).

On hardware: 105 qubits at the Willow announcement, 500 now. Coherence went from roughly 20 microseconds in 2022 to routinely above 300 microseconds. At a modest distance-5 code, logical error rates are approaching 10⁻⁴. And the number he clearly cares most about is lambda, the error suppression factor per code-distance step: barely above 1 in 2023, about 2 in 2024, now approaching 4. His philosophy is stated plainly — other groups have more qubits or longer coherence, but the point is whether the whole thing works as a computer, so they optimise for system-level experiments.

Give him credit for naming his worst problem himself. Those beautiful chandelier photos of a dilution refrigerator are, in his words, an engineering abomination, because every coupler and every qubit needs its own wire and you cannot scale that to a million qubits. Hence the acquisition of MIT spinout Atlantic Quantum to work on cryogenic control electronics. Milestone three is one well-protected logical qubit at one-in-a-million error; where they once thought that needed a distance-15 or 17 code (a 15×15 to 17×17 grid of physical qubits per logical qubit), faster processor progress suggests distance 11 or even 9 might get there.

The software half turns on a distinction worth writing down: hard is not the same as useful. Their random-circuit-sampling benchmark from 2019 is unambiguously hard — the classical cost has grown at a double-exponential rate, and by 2024 the best supercomputer would need something like 10 septillion years. But he concedes the standard criticism, that it solves nothing anyone on Main Street cares about, then flips it into a gate: a machine that can't do this well won't do anything practical either. That's a necessary condition dressed up as a recommendation, and I don't think it fully answers the objection.

Between the two poles sit what he calls "discovery nos" — small discoveries, not Nobel-scale ones. Traversable wormholes, non-abelian anyons, time crystals, lattice gauge theories: real papers in Nature and Science, real physics learned by preparing a state on a programmable device and studying it, which is more or less Feynman's original pitch. All of them, he admits, could still have been simulated on a large classical machine.

The October result on the cover of Nature is meant to be both hard and useful, and he walks through the trick. The naive way to make a computation classically expensive is to build a big enough circuit — but the state becomes so entangled that any local observable averages out and the signal falls off exponentially. So they refocus: run a unitary, run it backwards, return to the start, and in the middle poke one qubit with what they call a butterfly operator, then watch how that perturbation propagates to a distant probe qubit. That correlator is an out-of-time-order correlator, or OTOC; the algorithm is nicknamed quantum echoes (28:41).

Two things back the hardness claim. Ten person-years of red-teaming, trying the best classical algorithms against it. And an experimental protocol: insert a cut mid-circuit that puts a dephasing gate on every qubit, and see whether the result degrades. If it does, interference is genuinely doing the work. The plain OTOC passed that test but weaker than they hoped, so they ran the forward-backward sequence twice — OTOC-squared — which raised the sensitivity and, on paper, opened up much more long-range interference. Nature's editors apparently called it a weak algorithm; Michel Devoret reframed it as the difference between single-pass and multi-pass interferometry, and Neven's own analogy is that it could become a kind of LIGO for molecules. For usefulness, there's a companion paper using these signals to learn molecular structure from nuclear magnetic resonance data, which is a concrete, checkable claim rather than a vibe.

He closes on a deadline. A colleague — the transcript says "Greg Gibney," which I suspect is Craig Gidney — showed this year that factoring 2048-bit RSA integers with Shor's algorithm needs on the order of a million physical qubits, which lands exactly on Google's milestone six. The roadmap therefore ends at a cryptographically relevant machine. And weaker schemes are exposed first: Bitcoin's 256-bit elliptic curve is named directly, with the advice that there's no time to procrastinate on post-quantum migration.
 

The Xenon Detour, and the Questions I'd Have Asked From the Floor

Neven gives himself permission, because it's a launch event, to get audacious. The AI portion is easy to follow: his lab is called Quantum AI because he expects quantum resources to make AI stronger, and one of the first practical contributions is training data. AlphaFold worked partly because it inherited fifty years of patient curation in the Protein Data Bank; he claims one second of usable fusion plasma data would cost something like a trillion dollars to produce today, and that a programmable quantum machine could generate equivalent training sets far more cheaply.

Then the turn. Despite AI's success, he says, there are blatant holes in our understanding of natural intelligence — he corrects himself to say sentience. What physical process implements the experience of a toothache, or of smelling a rose? He jokes that asking this is career-ending unless you're already tenured, and argues the question has a kind of ontological primacy: our account of the Big Bang may look different in a century, but people will describe their toothache the same way. He finds it an embarrassment that we've made so little progress.

With outside colleagues he has formulated a conjecture that quantum processes need to be considered — that quantum processes are what create conscious experience — and there is a paper that both derives the conjecture and, more importantly in his telling, lays out an experimental program to test it. If you knew which process implemented experience, he suggests, you could in principle link a human brain to a quantum computer and shape the conscious experience of the person attached.

The first baby step is xenon. They expose biological systems to different isotopes of the noble gas: brain organoids, which he describes as lentil-sized blobs of generic cortical tissue, and fruit flies, in an experiment he likes because a high schooler could repeat it (39:45). Roughly forty flies in a syringe, 10% air, 90% xenon of a chosen isotope. At one atmosphere nothing much happens; push the plunger to about three atmospheres and the flies drop, which he calls the Glastonbury festival effect; pull the plunger back and within seconds they're flying again. The striking part is that different isotopes need the plunger pulled different distances — the flies behave as isotope detectors. Someone named Wolfgang in the audience is credited with having doubted exactly this. Why xenon at all? Because it's an anesthetic, and after roughly 180 years of use we still don't know how anesthetics work.

Here's where I stopped following, and I want to be careful because I'm not a neuroscientist. The isotope result, if it replicates, says something real about anesthetic potency and nuclear spin. The path from there to "quantum processes create conscious experience" was not laid out — the transcript even has him say he "will have time to explain it" and then not explain it, which I assume is a mis-transcription of "won't," but either way the mechanism never arrived. The further suggestion about sculpting someone's experience by wiring a brain to a quantum computer is, on the evidence shown, speculation piled on speculation. I'd file it as the most interesting thing in the room and the least supported.

The other loose threads I'd have asked about, briefly: the 10⁻¹⁰ result is one-dimensional and the jump back to two dimensions is anything but routine; random-circuit sampling is a gate, not a product; IBM's 2029 target came without the physical error rate it depends on; and both roadmaps are, unsurprisingly, roadmaps drawn by the people building the machines. Still, the overwhelming impression coming away is that quantum science has quietly shifted disciplines — the argument on stage was mostly about packaging, wiring, cryostats and decoders, which is a strange and encouraging place for the field to be.