Noah Hollmann

From Human Intuition to Superintelligence

The session’s core message is that the rise of increasingly capable AI forces a practical boundary-setting exercise: deciding when to trust machine prediction and when human intuition, values, and responsibility must remain decisive. The opening uses a vivid time-machine scenario set in 2040, where AI has surpassed human intelligence but daily life still presents irreducibly human choices. A parent is told there is a 40 percent chance of bullying on a child’s first school day; an AI can extend a dying dog’s life by six months, though the owner senses the dog does not want that; and an AI recommends a New York job opportunity with a “90977% probability” of personal and career success, while another option later appears more fulfilling. These examples frame the panel’s question: where is the boundary between AI, human intelligence, and human intuition? The panel is introduced as Noah Hollmann of Prior Labs, Matthias Niessner of SpAItial AI, Tim Rocktessner of Recursive, and Andre Retterath of Earlybird VC. The discussion then turns to recent claims attributed to Anthropic researchers that self-improving superintelligence could arrive this decade and that there is a nontrivial chance it could ultimately kill humanity. Tim Rocktessner responds skeptically, calling a claim such as a 10 percent chance of humanity being dead by the end of the decade a “massive overreaction.” He argues that extraordinary claims require extraordinary evidence, and that while AI-related cybersecurity and biosecurity risks are sensible concerns, the evidence cited so far does not point to a short- or mid-term future in which everyone dies. He also questions probability estimates from figures such as Jeff Hinton when they appear to rest on gut feeling rather than transparent evidence. Matthias Niessner similarly says the extinction framing is overstated from a technical perspective, while emphasizing that AI can still be dangerous depending on human use. He distinguishes between an intrinsic will in AI systems and humans deploying AI for harmful purposes, including defense applications where the key issue is who controls the trigger. Andre Retterath adds that even small percentages of humanity-level or severe downside risk deserve discussion because the technology is evolving rapidly and can create both good and bad outcomes. The panel agrees that safeguards, pressure on companies, and safer system setup can reduce the chances of things going wrong. In the available excerpt, the conversation closes by defining superintelligence: Prior Labs’ work on structured and tabular data is described as already superhuman in a narrow domain, while broader superintelligence is framed as a system able to solve all, or the vast majority, of meaningful human tasks over time.