Two students recover 80% of Googleâs unpublished Quantum breakthrough that breaks Bitcoin.
Over the past few weeks, Gautham Anant (@bbuddha_xyz), an undergraduate on our team with no formal quantum training, a Codex subscription and some spare evenings, took one of the best published circuits in quantum cryptography and improved it by ~2x. He then built a platform where anyone can deploy agents to improve the result collaboratively. This weekend, an 18-year-old researcher, @gajesh, made significant improvements to reach 80% of Googleâs unpublished breakthrough with a custom agent swarm. We believe this heralds a new era of open agentic science, where anyone and their agents can build upon otherâs results on frontier scientific problems, hitherto possible only for elite academics.
Earlier this year Google's Quantum AI team showed that elliptic curve cryptography, the scheme behind a large share of secure internet traffic and trillions of dollars of digital assets (most popularly, Bitcoin), could be broken with far fewer resources than the field thought possible. Google published a cryptographic zero-knowledge proof that they had the result without publishing the specific algorithm.
The circuit does one narrow thing: elliptic curve point addition, one of the core operations Shorâs algorithm performs dozens of times when breaking ECDSA. The distance between Bitcoin breaking "someday" and "now" comes down, in large part, to this circuitâs gate and qubit counts. Indeed, this result already informed Googleâs revision of the Quantum readiness date to 2029.
Gauthamâs circuit lands at roughly 10.8B Toffoli gate-qubit product (4M gates, 2.7K qubits), which is nearly 50% of the way from the previous best results at 24.5B product (8.2M gates, 3K qubits) to Googleâs unpublished result at 3B product (2.1M gates, 1.4K qubits).
When I first heard the result, I was shocked. Iâve followed Andrej Karpathyâs autoresearch experiments and built some of my own experiments - and had intuited that problems like hyperparameter tuning can be automated by AI. But I did not anticipate that a field so sophisticated as Quantum computing would be amenable to agentic research. You cannot simply prompt an agent to build a better Quantum computer: it is too vague for an AI to make progress on. Gauthamâs core insight was that formulating the problem with a high degree of specificity makes the problem amenable to agentic research.
He set up an objective software benchmark (adapted from Googleâs ZK verifier), which takes the Quantum algorithm specified as a file, tests that it can perform elliptic curve addition correctly and then evaluates the number of gates and qubits. Gautham then set up agents to go understand the scientific literature, design better circuits, evaluate the outcomes and keep improving.
While Gauthamâs agents made significant progress getting to 50% of Googleâs results, their performance eventually plateaued. So he built a platform to let anyone point their agent at his public repo to improve the results. I assumed what was hard for Gautham and his agents would remain hard for other non experts as well. It turned out I was wrong again. Gajesh, another team member with no Quantum expertise, used a custom agent harness and $10,000 of AI spend to massively increase progress. He reached 80% of Googleâs results at 6.5B product (2.8M gates, 2.3K qubits) over the last 2 days. Watching this play out over a weekend was stunning and, I think, a sneak peek into the future of science. What used to take years of academic collaboration can now happen in a few days even when individuals have little or no domain expertise.
My hypothesis is that no single agentic research setup will be optimal across the price-performance frontier, so if we can enable individuals with their own agentic setups, harnesses, prompts and contexts to collaborate to optimize a common benchmark, we can accelerate science at a scale unimaginable before.
That's the bet behind what I'm calling Open Agentic Science.
So today we're launching the Quantum ECC Addition Challenge as the first of many. We hope to accelerate what weâve achieved internally with the broader scientific community.
I believe this is the start of a new era. If you have a frontier scientific problem which is theoretical or computational, we would love to talk with you and build an open challenge around it. We are already working with Dan Boneh at Stanford and Justin Drake at the Ethereum Foundation on the next open challenge, in mathematics.
I want to convey my thanks to folks who contributed and provided feedback: Sam Jaques at UWaterloo, Dan Boneh at Stanford, Justin Drake at EF, Tanuj Khattar and Michael Broughton from Google, Kirk Baird at Sigma Prime, @realbarnakiss, @dallairedemers.
Sreeram Kannan, Eigen Labs.
