@LudditeDesign Full duplex audio transmission and full duplex interactive speech models are completely different problems. One is multiplexing bytes on a wire and was solved decades ago. The other requires a model running continuous inference and deciding when to respond/delegate/etc.
@silasalberti Both methods leads to different leaps. Pretraining scaling is like sprinkling magic dust, capabilities emerge which are difficult to predict a priori. RL is for making the model cracked at a specific shaped problem.
@danshipper@every This matches my experience and points to the intelligence jump we get from larger pretrains. 5.6 is likely an Opus-sized model with much better RL. Fable is a 10x bigger model, so it exhibits genuinely novel intelligence.
@GovTimWalz They didn’t wear those clothes though. Their parents made that choice for them. And they’ll continue to push choices on them that don’t match up with western ideals.