AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Reward Is Not Enough, Right?
The ability to model phenomena at multiple levels of abstraction or description is really, really important. And also doing it with the joint embedding prediction architecture on the latent space. Also, action space abstractions in time, as well as concept abstraction. I was a huge fan of that paper. It did introduce uncertainty quantification as part of the, you know, predicting what is not observed from what is observed. So it's a bit... I mean, one thing I did think at the time though is that it does start to resemble a handcrafted cognitive architecture again.