They aren't using an exact simulator (since they use >100 qubits), but an MPS. If there was honest-to-god-quantum-advantage in there, one might expect something like this. That said, I remain very skeptical...

You got me! (This is a joke, dual rail transmons are the future!)

Both the Puri group from Yale and J. Claes from Logiqal published exciting results today with some overlapping ideas , so check their papers out as well:
arxiv.org/abs/2509.05212
arxiv.org/abs/2509.05232

The bottom line is : magic states generation is now super cheap for architectures with long idling time and long range connectivity.

Read our updated paper here:
arxiv.org/abs/2502.01743

This can reduce the number of atoms used for addition during factoring by almost half, reducing the number of active atoms by almost 8 Million!

The most important point is that we can now get a 1E-7 fidelity magic state with over 75% success probability on Rydberg atoms with p=1E-3. This is high enough fidelity for factoring, and over 20x more efficient compared with resource estimates done THIS YEAR

Starting with a small surface code makes the expansion to a larger surface code super cheap, which is why it becomes so much more efficient that the original Cultivation paper

We used the fold transversal gates of on the surface code to measure a non-Clifford state and project the code into a useful magic state. This time we only need a single patch to make it work