I think the biggest problem is it’s very low ability to “test time adaptability”. Even when combined with a reasonning model outputting into its context, the weights do not learn out of the immediate context.

I think the solution might be to train a LoRa overlay on the fly against the weights and run inference with that AND the unmodified weights and then have an overseer model self evaluate and recompose the raw outputs.

Like humans are way better at answering stuff when it’s a collaboration of more than one person. I suspect the same is true of LLMs.