The Roman Forum

What happens when the race to build superintelligence is actually a suicide race?
MIT physicist and AI researcher Max Tegmark joins Roman Yampolskiy for an uncompromising discussion about artificial general intelligence, uncontrollable superintelligence, AI governance, mechanistic interpretability, formal verification, recursive self-improvement, and the probability that humanity loses control of its future.
Tegmark explains why he abandoned much of his mechanistic interpretability research, why understanding a neural network may inadvertently accelerate AI capabilities, and why he now favors “verified coding”: using AI to produce conventional software accompanied by mathematical proofs that it behaves as specified. He argues that safe, controllable AI tools could deliver enormous benefits without requiring humanity to deploy autonomous black-box systems more intelligent than their creators.
The conversation examines whether advanced AI should be regulated like pharmaceuticals, nuclear technology, or dangerous biological research. Tegmark proposes an “FDA for AI,” under which developers would have to demonstrate that powerful systems provide greater benefits than harms before they can be trained or deployed. He argues that voluntary safety commitments are structurally inadequate because even well-intentioned AI executives remain trapped in a competitive race.
Despite the existential risks, Tegmark ends on an optimistic message: humanity has a winning position, but it must stop playing as though only seconds remain on the clock. Safe AI could help cure disease, expand human agency, and allow life to flourish for billions of years. The central danger is not that such a future is impossible, but that humanity may squander it by moving too quickly.
Subscribe to The Roman Forum with Roman Yampolskiy for conversations about artificial intelligence, superintelligence, AI safety, consciousness, simulation, and the future of humanity.
#MaxTegmark #RomanYampolskiy #ArtificialIntelligence #AISafety #AGI #Superintelligence #AIAlignment #ExistentialRisk #FutureOfLife #AIGovernance

Why did one of the world’s earliest and most influential AI alignment research organizations fail to solve the problem it was created to address?

In this episode of The Roman Forum, Dr. Roman Yampolskiy speaks with Nate Soares, president of the Machine Intelligence Research Institute and co-author of If Anyone Builds It, Everyone Dies, about the unresolved problem of aligning superintelligent AI with human values. Soares argues that alignment may be physically possible in principle, yet remains far beyond the reach of current methods. Humanity may need centuries of intellectual progress to solve it, while advanced AI could arrive within years.

The conversation examines why MIRI’s technical research did not produce a workable solution, whether an AI can genuinely care about human flourishing, and why ordinary scientific trial and error may be impossible when a single failed experiment could become irreversible. They also discuss coherent extrapolated volition, human value aggregation, the distinction between alignment and control, the point of no return, alien superintelligence, the Fermi paradox, acausal trade, P(doom), human cognitive enhancement, and the possibility of enforcing a global pause on frontier AI development.

Soares’s central proposal is direct: humanity should stop the race toward superintelligence before systems become capable of automating AI research and accelerating beyond meaningful human control. He explains why a pause would need to be international, how advanced chips and large data centers could be monitored, and why the world may need to treat superintelligence development with a seriousness comparable to nuclear proliferation.
http://so8res.com/