@Plinz@simone_m_romeo I think there's a risk in the current pivot. We were sold the chase for a mega-mind AGI via scale but the frontier now lives within nanosecond-interlinked dumb-ish agent societies. These could have a large impact, no matter how limited their individual constituents.
@beffjezos Isn't it structural? If progress is achieved through swarms of "dumb"-ish agents then networking / latency can be expected to dominate just as it does for training no?
@beffjezos@DavidSacks What frontier? It just changed from billion dollar training runs to billion dollar n8n instances. From mega-mind to virtual society. LLM's emergent properties were as obscure to @DavidSacks as they are to the dudes in the AI Labs, societal structures will be much easier to tune.
@elonmusk So it's either concentration or evaluation via emotional resonance. Feels like we've already been there, maybe the initial OpenAI distributed AI intuition was right. Embed the agents graph in the societal graph and retain some link to reality fitness.
@dwarkesh_sp@johnschulman2 Next token prediction is progressing logarithmically isn't it ? Seems recent headlines have more to do with scaling downstream via agent societies / superhuman amoebas.
@dwarkesh_sp Agent societies seem like a far more programmable object class than neural nets, and one for which we actually do have epistemological frameworks to grow capabilities & intelligence. All aboard The Ends of Invention
@elonmusk