Personalized, heterogeneous foundation models
Perhaps language is not solved just yet?
Language and sparsity
Since before the LLM era I always thought there is a tension between open-domain general knowledge and personalization/diversity. Each individual human has separate and specific prior knowledge (through genes, and lived experience — memories), and a set of goals/needs which drive communication. I think that humans don’t have word/response “distributions”, rather the next action (including speech) they take is fully conditioned (in a statistical sense) on all innate and prior experience. These latent variables are missing when pre-training on internet text which I think also affects sample efficiency.
Separately, by learning from text outputted by millions of humans there will necessarily be a homogenization of language. An LLM is not modeling how humans use language, it is modeling language in general. I have this hunch that this is precisely the issue that is bottlenecking true creativity and discovery. Scale was needed to attain general language capabilities, but true creativity may be orthogonal to scale, or perhaps, beyond a certain point, more data hurts creativity.
I think this hints at the AI of the future being more human-like in sample efficiency. I don’t want to make strong claims here, but a truly astounding system would be able to have the same conversational quality as e.g. GPT-5 while being trained on much less data than the whole internet, which may unlock true creativity too as a side-effect. Moreover, by following this line of logic it is clear that a single model will never be the end goal, since it would be quite domain-limited — as individual humans are. Thus, perhaps future AI will be much more like a collection of many varied models/agents trained on various data sources/domains. Perhaps developing current MoE-style models more could provide a path towards this, by introducing data partitioning per expert, etc.
In any case the distinction between many smaller models trained on different data (like humans) or one giant model trained on all human data is really materialized in the communication bottleneck. Smaller models/agents are forced to communicate with (ideally human) language, akin to a group of humans cooperating to form a larger entity (e.g. country, company) while a single big model does not enforce such a bottleneck, communication is embedded in hidden state propagation, analogous to the brain’s various regions communicating through synapses directly.
It is not entirely clear which approach is better, i.e. the degree of coupling between subparts of a larger system. Intuitively a single model that has no communication bottlenecks should be more capable, but it could be that interfacing bottlenecks, by forcing cognitive development of efficient information transfer and negotiation/alignment, result in better overall capability of the system. In essence sparsity sort of provides a sliding scale for this. The sparser the connective structure between two networks, the higher the information bottleneck imposed on the communication channel.