For hosted models I’ve noticed that Bard has been quietly improving. Was awful at first but now seems to be closing in.

Actually-open AI with open weights and open source is innovating rapidly too, especially when it comes to making models more efficient on smaller hardware. We have fast local mixtures of experts now. Rumor is that Llama3 is training as well, and that it may be some kind of MoE.

I feel like focusing on efficiency first for local models rather than shooting to beat GPT-4 is the right approach, because without major efficiency improvements a shot for simple dominance will just yield a model only wealthier people can afford the hardware to run.

OpenAI still seems in the lead but it’s shrinking.