Classical computers correct errors by making copies and comparing them. Quantum computers cannot: copying an unknown quantum state is impossible, and looking at a qubit destroys what it held.
So error correction has to fix mistakes without ever reading the data. The trick is to spread one logical qubit across many physical ones and then measure only the relationships between them - whether neighbours agree - never the values themselves. Those measurements reveal that an error occurred and where, while leaving the encoded information untouched.
Below a certain hardware error rate, adding more physical qubits makes the logical qubit better. Above it, more qubits make things worse. That crossover is the threshold, and crossing it convincingly is the milestone the entire field is organised around.
When you see "fault-tolerant" in an announcement, this is what it refers to.