IBM Quantum Leap
Fault tolerant qubits slash transformer training time
IBM's Quantum Starling roadmap surprisingly never mentions AI or machine learning, focusing instead on fault tolerance and supercomputing, yet the implications are obvious. The 127 fault tolerant qubits cut gate errors, letting Monte Carlo samplers and gradient scanners sprint through training loops, while a goal of one hundred million circuit gates before 2030 could shrink transformer pretraining from days to mere hours. Once fault tolerant machines do the same work in far less time, energy per model trained should fall sharply. Watch the video →