Interference

Once upon a time, not long ago, when scaling was the practice, just spendin’ dough, the laws were firm: “more neurons, good”. And people were just trainin’, big as they could. There was a scientist way up on the north side, who wondered, "Why these models need to be so wide? Could we go smaller and spend less cash? Or would they just learn far fewer tasks?” He’s from Goodfire, name of Ekdeep Singh, and soon he was joined by his good friend Jing. They gathered up a crew, Stanford, Anthropic, MIT, Kempner, “Yo, let’s study this topic.” They started out with basic scaling laws, and spotted a prediction that gave them pause: Given infinite compute, they’ll be task combos, that big models nail but small ones fumble. That seemed intuitive, almost tautological, but they said, nope, merely phenomenological. We need much more than this formal prediction. Let’s find the causes, make sure it ain’t fiction. Think about it for a minute and it’s easy to see: examples flowin’ in, varied ease and frequency. Your bigger models master all the common easy cases, gradients go to zero, learning enters total stasis. Now should a hard example from a seldom task arrive, its updates are remembered, and your learning can still thrive. For your smaller models, things just don’t work out. Simple frequent tasks get updates, all else: wiped out. For a toy setting, they proved this handily. Just check out Theorem Four and its corollary. Now, in that toy setting, experiments match theory, but what about for real models? That’s the big query. So they pre-trained up some OLMos, all the way to 4B, injecting weirdo tasks, varied frequency. From the theory they surmise, this will come as no surprise, We have come to realize, that it really is the size. The tiny model tries, but capacity’s the skies, and so it always dies. We already gave the whys. But the results open wise eyes: The big OLMos learned the most infrequent tasks, battled grad collisions, kept on standing fast. And when the team sought features, in all these LMs, the causal interventions showed really clear trends: as the neuron count goes up and the task gets more pervasive, geometry gets richer, the whole picture’s more persuasive. For those tiny OLMos, on the other hand, they failed the novel jobs; interference was their end. So the bigger models get all of the glory. And this is the way we have to end this story. They were only small models in an AI dream. They couldn’t learn the features, that’s a real clear theme. This ain’t funny so don’t ya dare laugh. Your stuck with endless scaling, that’s just the math. But: being data-centric might be the path. To save you from the scaler’s costly wrath. Good-night. Knock it out the park, Jing, knock it out Jing. Knock it out the park, Singh, knock it out Singh. Knock it out the park, Jing, knock it out Jing. Knock it out the park, Singh, knock it out Singh.