So then isn’t the issue with how the tokens are encoded (embeddings)? It wouldn’t be an issue with tuning the model parameters, because stochastic gradient desc

So then isn’t the issue with how the tokens are encoded (embeddings)? It wouldn’t be an issue with tuning the model parameters, because stochastic gradient descent will always find a local maxima or minima.