precisely — we need to train alternative models from the ground up on more diverse cultural contexts, instead of trying to retrofit a westernized model in non-western contexts
This is the right framing and I think about it constantly. The retrofit approach, fine-tuning a Western-trained base model on local data, is better than nothing but it is architecturally compromised from the start. You are trying to correct value misalignment at the surface while the deep structure of the model remains shaped by the corpus it was originally trained on. It is like translating a concept that does not exist in the target language and wondering why something is lost.
On who is working on this seriously: ILINA under Cecil Abungu is doing some of the most rigorous thinking on African-context AI development rather than adaptation. The MASAKHANE community has been building African NLP infrastructure from the ground up for years and is probably the closest thing to what you are describing in practice. My own work on GENSCORE, building culturally situated mental health NLP for Hausa, Yoruba, and Pidgin English communities using lived-experience corpora rather than translated Western instruments, is a small piece of this puzzle applied to a specific domain.
But to be honest, no one is doing this at the scale the problem demands. The compute costs of training frontier models from scratch put it out of reach for most Global South research groups without major institutional backing. Which is part of why the governance and funding conversation matters as much as the technical one. If the resources to build genuinely diverse foundational models only flow to labs in San Francisco and London, the alignment problem remains a Western problem with a Western solution applied everywhere else.
Worth a much longer conversation. What is your context for the question?
“Alignment to whose values?”
precisely — we need to train alternative models from the ground up on more diverse cultural contexts, instead of trying to retrofit a westernized model in non-western contexts
wonder who’s working on this?
This is the right framing and I think about it constantly. The retrofit approach, fine-tuning a Western-trained base model on local data, is better than nothing but it is architecturally compromised from the start. You are trying to correct value misalignment at the surface while the deep structure of the model remains shaped by the corpus it was originally trained on. It is like translating a concept that does not exist in the target language and wondering why something is lost.
On who is working on this seriously: ILINA under Cecil Abungu is doing some of the most rigorous thinking on African-context AI development rather than adaptation. The MASAKHANE community has been building African NLP infrastructure from the ground up for years and is probably the closest thing to what you are describing in practice. My own work on GENSCORE, building culturally situated mental health NLP for Hausa, Yoruba, and Pidgin English communities using lived-experience corpora rather than translated Western instruments, is a small piece of this puzzle applied to a specific domain.
But to be honest, no one is doing this at the scale the problem demands. The compute costs of training frontier models from scratch put it out of reach for most Global South research groups without major institutional backing. Which is part of why the governance and funding conversation matters as much as the technical one. If the resources to build genuinely diverse foundational models only flow to labs in San Francisco and London, the alignment problem remains a Western problem with a Western solution applied everywhere else.
Worth a much longer conversation. What is your context for the question?