The Chaperone

IBM's biggest quantum chemistry result depended on a conversation between very different machines. In May, Cleveland Clinic, RIKEN, and IBM simulated a protein–ligand complex in water totaling 12,635 atoms, which IBM describes as the largest heterogeneous quantum-classical electronic-structure calculation to date. Two 156-qubit IBM Heron processors, one at Cleveland Clinic and one at RIKEN in Kobe, did the quantum part. Classical computers did everything else: cutting the protein into fragments, sampling the nastiest fragments, and then... handing it off. Then two classical supercomputers took those samples: Fugaku, which topped the TOP500 rankings before Frontier, and Miyabi-G. The algorithm (called EWF-TrimSQD, because quantum computing has never met an acronym it didn't want to marry) made progress by stitching the molecule back together.

It's kind of crazy to see the pace of things. In January, at Supercomputing Asia, IBM's Jay Gambetta presented what was then the largest and most accurate chemistry experiment ever performed on a quantum computer: a pair of iron-sulfur molecules, solved by one Heron chip in a closed loop with all of Fugaku, the two machines passing data back and forth in an unbroken cycle. The protein work started smaller: Trp-cage, 303 atoms. Five months later, in May, it had a T4-Lysozyme–ligand complex in water at 11,608 atoms and a trypsin–ligand complex in water at 12,635, a forty-fold jump. That's a big jump! And I don't want to take that away from anyone at Cleveland Clinic, IBM, or RIKEN who spent the preceding year getting a quantum processor to provide a real-world use case. BUT.

Quantum computing has been sold for something like two decades as the technology that gets you out from under classical computing's limits, the thing that solves what a normal machine can't solve no matter how many racks you buy. Chemistry simulation is the textbook example: the compute required to track every electron's interaction with every other electron grows exponentially with the size of the molecule, which is why a high-accuracy classical treatment of a whole protein is still impractical and why quantum computing has always used chemistry as its opening pitch. So, we have the biggest (quantum) chemistry result of the years still needing classical machines at every single step, and... it still does not yet outperform the best classical approaches. There's a good reason (fault tolerance) which will be addressed by 2029, which will still not remove classical control, error decoding, or the classical steps of a hybrid chemistry algorithm, but it'll be ... more quantum? IBM describes the future as quantum-centric supercomputing, with the machines working together.

On narrow tasks, researchers have reported results beyond the classical methods they compared. On July 30, IBM and Qedma claimed a quantum advantage on the dynamics of a driven quantum magnet, up to 74 qubits, using error mitigation and comparisons with classical methods where those methods worked. A September report on an unreviewed Nighthawk preprint described a million random-circuit samples in 19 seconds. BlueQubit's 110-year Frontier estimate was for a particular tensor-network method at comparable fidelity, under an unlimited-memory assumption. But benchmarkmaxxing random-circuit sampling is great for press, not a chemistry result. Useful chemistry still needs a comparison against the best classical workflow for the same problem.

There's an almost too-on-the-nose precedent for this from transport. Britain's Locomotives Act of 1865, called the Red Flag Act, required every road locomotive to be preceded by a person on foot carrying a red flag, walking sixty yards ahead to warn horses and riders that the machine was coming. These were steam engines of up to 14 tons, rolling down roads built for horses. They could genuinely propel themselves, but they couldn't (yet) do was be unsupervised, so the law bolted a human chaperone onto the front, capped the speed at four miles an hour in open country and two in town, and made the astonishing new machine walk at the pace of the person escorting it. Parliament dropped the flag itself in 1878 and kept the man, who now walked twenty yards ahead with nothing in his hands. He lasted until 1896, when light vehicles under three tons, the new motorcars among them, were finally let off the leash. On November 14 that year, the Earl of Winchilsea marked the occasion by tearing up a red flag that nobody had been legally required to carry for eighteen years.

The quantum computers in Kobe and Cleveland contribute a specialized step inside a larger classical workflow. The result is real progress. The test is whether the whole workflow produces a useful answer better than the best classical alternative, including its classical processing and coordination costs. The red-flag story is a reminder to inspect what surrounds a new machine. It does not establish a date when that support disappears. IBM's 2029 target concerns fault tolerance, and its chemistry researchers expect the hybrid workflow to continue.

Want to learn how intelligent data pipelines can reduce your AI costs? Check out Expanso. Or don't. Who am I to tell you what to do?

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!