Akarsh Kumar

Akarsh Kumar × Tim Scarfe

LLM-guided evolution beat 96% of roughly 300 human-written Core War programs, showing how objective verification and cumulative selection can amplify weak model priors.Artificial-life research extends discovery into possible physics and intelligence, but its strongest behaviors occupy a narrow region of rule space, making automated search valuable while leaving AGI a long-term, unresolved wager.

Tim Scarfe × Dr. Duggar × Kenneth O. Stanley × Akarsh Kumar × Keith

PicBreeder shows that identical outputs can conceal radically different machinery: evolved networks factor a skull into reusable components, while conventional SGD produces “total spaghetti.”If fractured representations make adaptation, continual learning and transformative creativity costly, growing sparse, protected modules through open-ended selection could be dramatically more efficient, though the “trillion-dollar question” remains an unproven research agenda alongside continued LLM scaling.

Tim Scarfe × Kenneth O. Stanley × Keith Duggar × Akarsh Kumar

Picbreeder’s skull networks suggest neural systems need not produce “garbage representation, just total spaghetti,” because modular components can independently control features such as a mouth opening, closing or smiling.The commercial risk is that benchmark success masks weak generalization, creativity and continual learning, potentially making frontier progress “insanely expensive”; Stanley remains uncertain whether scaling can push through, while Kumar recommends diversifying research beyond LLM scaling.