GRAM approximates data filtering and the question is just as applicable there: Is data filtering useful when models can re-derive, continually learn, or in-context learn?
I would argue yes, data filtering (and therefore GRAM) are still useful in this setting. Lets just stick to in-context learning case, e.g. you feed an advanced LLM huge amounts of knowledge on virology and it can derive instructions for building bioweapons in-context. For closed models, the developer could still monitor the in-context information and flag it before the model provides dangerous instructions. It also seems like preventing the model from learning virology is better or equal to forcing it to rederive things in-context. It’s strictly better in terms of the compute needed (e.g. doesn’t need to ingest millions of tokens of virology context).
GRAM approximates data filtering and the question is just as applicable there: Is data filtering useful when models can re-derive, continually learn, or in-context learn?
I would argue yes, data filtering (and therefore GRAM) are still useful in this setting. Lets just stick to in-context learning case, e.g. you feed an advanced LLM huge amounts of knowledge on virology and it can derive instructions for building bioweapons in-context. For closed models, the developer could still monitor the in-context information and flag it before the model provides dangerous instructions. It also seems like preventing the model from learning virology is better or equal to forcing it to rederive things in-context. It’s strictly better in terms of the compute needed (e.g. doesn’t need to ingest millions of tokens of virology context).