Pedro Domingos quips: 'new idea' called RNNs will power next-gen LLMs

pmddomingos · x · 2026-09-05

ML scholar Pedro Domingos posted a tongue-in-cheek take: "I have this amazing new idea called recurrent neural networks. They blow transformers away if you do them right. The next generation of LLMs will be based on them." The sarcasm nods to the ongoing revival of recurrent-style architectures (linear attention, state-space models) in the LLM research conversation.

Original post →

More from Fun

Fun channel →