Gabriel Pimenta de Freitas Cardoso

Great point, and I agree that this is where the discussion becomes most interesting.

It is not possible to say where the decline bottoms out, since neural networks learn representations through superposition rather than assigning individual neurons or subnetworks to specific domains. In other words, it is not possible to isolate parts of the network and claim that they are responsible for domain X or domain Y, as these domains are encoded through overlapping representations. What we can affirm is that it depends on the relationship between breadth of domains in the knowledge base and the number of parameters of the model (the greater the parameters, the greater the capacity to retain knowledge). To answer this question, we would need some way of measuring the amount of domain knowledge contained in the training data, as well as the degree of overlap between those domains in the latent space learned by a highly non-linear neural network. I believe this question is currently intractable.

Regarding the point at which investing in specialized content ceases to pay off, I believe that this only happens when the objective is to build a genuinely more general-purpose model. If the goal is to use the available resources (parameters) to provide shallow responses across a broad range of topics, then it makes more sense to train on a dataset with broad coverage but relatively little depth in each domain. On the other hand, if the objective is to build a model that solves a finite and relatively small set of tasks, more specialized data will always be the better choice. Naturally, as you increase the size of that task set, you eventually return to the first scenario of wanting a general-purpose model.

The same reasoning applies to the question of when we should pursue greater focus: whenever the scope of the application is itself focused. It is fundamentally a design choice. If you are training a model for open-ended tasks that may receive any type of input, it is better to have a model that responds superficially across many topics than one that responds deeply to only a few. Conversely, if the application has a well-defined scope, it will always be better to train a model specialized for that scope, making more efficient use of its resources instead of spending them on knowledge that will never be required.

That said, these are very relevant questions, and we hope to address at least part of them through mechanistic interpretability, which is one of the research areas within our R&D efforts. The techniques in this field allow us to observe internal circuits within models, sparsely used neurons, neurons that activate under specific conditions, and other internal mechanisms. This may help us better understand how models work internally and answer some of your questions more directly.