Watch: Unlocking hidden skills inside Google's model

Artificial intelligence models often possess hidden skills that no prompt can unlock. A recent study found that training a tiny, randomly selected fraction of an AI's weights can suddenly switch these capabilities on. To test this, researchers experimented on Google’s smallest Gemma model. They tried twenty different prompts to get the model to sort five numbers. Even with the best prompt, it succeeded only one percent of the time. But after training just one thousand of its one hundred million weights—that is just a tiny fraction of a percent—the model sorted the numbers correctly ninety-seven percent of the time. The same approach easily fixed other tasks. It got the model to output clean formatting ninety-nine percent of the time using just thirty trained weights, and it unlocked the ability to reverse word lists. However, for harder tasks the model never knew in the first place, like double-digit multiplication, training these random weights barely helped. This suggests the technique is not teaching new skills, but rather waking up dormant ones already inside the model. An analysis showed this awakening does not rely on a single magic switch. Instead, it comes from many small, coordinated adjustments, particularly in the parts of the model that handle attention. By nudging just a handful of random connections, we can access abilities that prompting alone leaves completely locked away.