AI models can now be built without a common stability fix, but they are harder to tune

Researchers found that some newer AI models can work without a common component that stabilizes training. This means these models might be more efficient, but they are also more sensitive to how they are set up.

For years, AI developers relied on a specific technique to make sure their models trained reliably. This paper shows that some newer designs can skip that step, which could make them faster or use less computing power. But it also means these models are trickier to get right, potentially limiting who can build and use them effectively.