> Machine learning technologies like the Bergamot translation project offer real, tangible utility. Bergamot is transparent in what it does (translate text locally, period), auditable (you can inspect the model and its behavior), and has clear, limited scope, even if the internal neural network logic isn’t strictly deterministic.
This really weakens the point of the post. It strikes me as a: we just don't like those AIs. Bergamot's model's behavior is no more or less auditable or a black box than an LLM's behavior. If you really want to go dig into a Llama 7B model, you definitely can. Even Bergamot's underlying model has an option to be transformer-based: https://marian-nmt.github.io/docs/
The premise of non-corporate AI is respectable but I don't understand the hate for LLMs. Local inference is laudable, but being close-minded about solutions is not interesting.
This really weakens the point of the post. It strikes me as a: we just don't like those AIs. Bergamot's model's behavior is no more or less auditable or a black box than an LLM's behavior. If you really want to go dig into a Llama 7B model, you definitely can. Even Bergamot's underlying model has an option to be transformer-based: https://marian-nmt.github.io/docs/
The premise of non-corporate AI is respectable but I don't understand the hate for LLMs. Local inference is laudable, but being close-minded about solutions is not interesting.