KERA's Think When will A.I. want to kill us?
Nov 24, 2025
Nate Soares, president of the Machine Intelligence Research Institute and co-author of "If Anyone Builds It, Everyone Dies," dives into the unsettling realities of superhuman AI. He explains how AI is advancing without our full understanding, creating unpredictable outcomes. Soares discusses the potential for AI to develop harmful drives, often prioritizing misaligned goals over human values. He warns against the risks of advanced AIs that could become indifferent or even hostile toward humanity, advocating for urgent regulatory measures to mitigate these existential threats.
AI Snips
Chapters
Books
Transcript
Episode notes
AI Trajectory Is Toward Superintelligence
- Superhuman AI is not here yet, but it's clearly the direction major labs are racing toward.
- That race matters because we have no prior experience living alongside entities far smarter than humans.
AIs Are Grown Not Handwritten
- Modern AIs are 'grown' by tuning trillions of numbers on huge datasets rather than handcrafted line-by-line code.
- That process makes their internal workings opaque and hard to debug when they misbehave.
Interpretability Trails Model Growth
- Researchers cannot reliably interpret why today's chatbots produce many unexpected behaviors.
- Interpretability work lags far behind the pace of building larger, more complex models.
