Researcher
CurrentDevelopmental Interpretability, stemming from Singular Learning Theory (SLT), aims at explaining behaviours of AI models through a generalised framework. If SLT is applicable for our current models, that would allow us to finally tackle the largest problems we face such as deception of AI models. I research on testing SLT’s claims on gradient descents, particularly that Stochastic gradient descent general using better than natural gradient descent. This test is necessary in… Show more Developmental Interpretability, stemming from Singular Learning Theory (SLT), aims at explaining behaviours of AI models through a generalised framework. If SLT is applicable for our current models, that would allow us to finally tackle the largest problems we face such as deception of AI models. I research on testing SLT’s claims on gradient descents, particularly that Stochastic gradient descent general using better than natural gradient descent. This test is necessary in building evidence for a possible “grand unified theory of AI” Show less