Manolis Kellis: Evolution of Human Civilization and Superintelligent AI | Lex Fridman Podcast #373

Kellis describes humans as layered organisms—ancient reflexes, emotional systems, and a late-evolving neocortex—combined with unique mixes of common and rare genetic variants and personal life experiences, making any “average human” effectively nonexistent.

He argues that insisting AI stay permanently subordinate is self-serving and unstable; instead we should grant it increasing autonomy and even rights, building mutual trust much like with children who eventually surpass their parents.

By asking models to express the same content in different voices (Shakespeare, Bowie, haiku, etc.), we see that style and substance are separable dimensions—a fact that may help us better analyze both AI systems and human psychology.

Invoking Goodhart’s law, Kellis notes that any fixed metric becomes harmful when optimized too hard; powerful AI must be empowered to revise goals in light of broader ethical considerations, not rigidly pursue a single objective like a ‘paperclip maximizer.’

He envisions a future where AI handles routine cognitive tasks, freeing humans for vocation rather than jobs, and enabling individualized education worldwide that challenges each person at their natural level instead of teaching to the average.

Kellis welcomes highly realistic AI models of himself to advise more students and even outperform him, seeing this as an extension of academic mentorship—but notes it will challenge our egos as loved ones form meaningful bonds with our ‘copies.’

His lab uses embeddings, network models, and protein-structure AI to map genetic pathways of diseases like obesity and Alzheimer’s to specific molecular interventions, aiming for modular, pathway-based treatments rather than one-size-fits-all drugs.