differential technological development
If you cannot stop a road from being built, you can still insist the brakes and traffic lights are ready before the cars are fast. Differential technological development is that idea applied to AI: since halting progress outright is unrealistic, the strategy is to deliberately speed up the protective, risk-reducing technologies relative to the dangerous, risk-increasing ones. It is not about stopping AI; it is about changing the order in which capabilities arrive.
Concretely, in AI this means trying to make safety-relevant progress, alignment techniques, interpretability tools, reliable evaluations, oversight methods, keep pace with or ideally lead raw capability gains, rather than lagging behind them. The motivation is the capabilities-versus-alignment gap: if the ability to build powerful systems races ahead of the ability to align and verify them, we deploy things we cannot yet make safe. Differential development asks each actor, and the field as a whole, to invest disproportionately in the safety side so that the protective tools exist when the powerful capabilities do.
There are honest tensions here. Some safety research requires building or studying capable, even dangerous, systems, so the line between safety work and capability work can blur, the same advance can cut both ways. There is also a coordination problem: a single cautious lab that slows its capability work may simply cede ground to less cautious competitors, which is why differential development connects to governance and ideas about reducing racing dynamics. It is a guiding principle and an aspiration, not a precise formula, and reasonable people disagree about how to apply it in specific cases.
A lab decides that before scaling a model to a new capability level, it will first invest heavily in evaluations and interpretability tools able to inspect such a model, so the safety machinery is ready in advance rather than scrambled together after release. Sequencing safety ahead of capability like this is differential development in action.
Deliberately ordering safety tools to arrive before the capabilities they must check.
Differential development is an ideal that is hard to execute cleanly. Safety and capability research often overlap, and a lone actor slowing down can simply lose ground to others, which is why the idea is usually paired with governance and coordination rather than relied on alone.