See the big shiny things we see in “AI” today are driving by a single paradigm change at the time, think convolutions for image processing and transformers for LLM. Progress could come from new forms of hardware (as it tends to btw, more so than actual algorithms) like we started using GPUs. The current trend shows that it makes sense to build the hardware more like we build the models (neuromorphic hardware), this way you can save orders of magnitudes of energy and compute so that it operates more like the brain. This is only an example of what else could happen, it could also be that language models stop improving as we are nearing the limit of language data apparently.