The AI safety debate has been dominated by two camps: the “accelerationists” who want to build as fast as possible, and the “decels” who want to pause or slow development until safety catches up. I want to argue for a third position that doesn’t get enough attention: that developing more capable AI systems is actually the best path to understanding and controlling them.
The Interpretability Argument
Anthropic’s recent sparse autoencoder research demonstrates my point. We can only understand how AI systems work — identify their internal features, map their reasoning circuits, predict their failure modes — by studying models at scale. Smaller models don’t exhibit the emergent behaviors that make AI both powerful and potentially dangerous.
If we had paused AI development at GPT-3.5, we would never have discovered the interpretability techniques that Anthropic is now using to make Claude safer. Progress in safety requires progress in capability.
The “Strategic Pause” Fallacy
The argument for pausing assumes that a pause would be globally coordinated. It won’t be. Even if the US and EU implemented strict AI development regulations, China, India, and the UAE would continue. A unilateral pause doesn’t make the world safer — it just shifts development to jurisdictions with fewer safety requirements.
What “Going Faster” Actually Means
“Going faster” doesn’t mean deploying unsafe systems. It means: (1) investing heavily in safety research alongside capability research, (2) running red-team exercises at scale before deployment, (3) building evaluation frameworks that keep pace with model capabilities, and (4) being transparent about risks and limitations.
The goal shouldn’t be to slow down AI development. It should be to make safety development go just as fast — or faster.