> exclusively search-based rewards so that the model isn't required to compress a large proportion of the internet into their weights.

That just gave me an idea! I wonder how useful (and for what) a model would be if it was trained using a two-phase approach:

1) Put the training data through an embedding model to create a giant vector index of the entire Internet.

2) Train a transformer LLM but instead only utilising its weights, it can also do lookups against the index.

Its like a MoE where one (or more) of the experts is a fuzzy google search.

The best thing is that adding up-to-date knowledge won’t require retraining the entire model!

Yeah that was my unspoken assumption. The pressure here results in an entirely different approach or model architecture.

If openAI is spending $500B then someone can get ahead by spending $1B which improves the model by >0.2%

I bet there's a group or three that could improve results a lot more than 0.2% with $1B.

> so that the model isn't required to compress a large proportion of the internet into their weights.

The knowledge compressed into an LLM is a byproduct of training, not a goal. Training on internet data teaches the model to talk at all. The knowledge and ability to speak are intertwined.

I wonder if this maintains the natural language capabilities which are what LLM's magic to me. There is a probably some middle ground, but not having to know what expressions, or idiomatic speech an LLM will understand is really powerful from a user experience point of view.

Or maybe models that are much more task-focused? Like models that are trained on just math & coding?