“I Do Not Save for Retirement” Is Not a Forecast

Nate Soares, president of the Machine Intelligence Research Institute, told The Atlantic’s podcast “The Most Interesting Thing in AI” that he does not currently save for retirement, Futurism reports. He said he still has a 401(k) from Google but has not contributed for a decade. Futurism connects that decision to his view that AI is more likely than not to wipe out humanity within the next ten years.

Granted, arranging personal finances around a warning can demonstrate conviction. Soares is not merely describing a future that would inconvenience somebody else. Yet the reported decade without contributions also matters: this is not a newly emptied account responding to a newly documented breakthrough. The account supplies evidence about his behavior, not a dated technical finding.

Still, sincerity, expertise, and forecast calibration are different credentials. Relevant expertise can help someone identify a dangerous mechanism. Calibration concerns whether a forecaster’s confidence tracks outcomes over time. A willingness to bear costs establishes neither the mechanism nor the accuracy. The retirement account is not a peer reviewer.

The strongest general catastrophic-risk argument deserves better than a joke about savings: increasingly capable systems could gain the ability to act beyond reliable human control, and intervention could fail before institutions understand the danger. That possibility warrants investigation even without certainty. But this summary of the risk argument is not a reconstruction of Soares’s technical case. Futurism’s account does not provide enough of that case to validate his probability or ten-year deadline.

Soares also says he works to prevent the outcome and does not consider it inevitable. That qualification makes the forecast conditional on human decisions, not merely technical progress. It also creates an accountability requirement: a forecaster should explain which interventions would reduce the danger. Otherwise, both catastrophe and its successful prevention can be made to sound like confirmation of the same warning.

Several futures remain possible without assigning them invented odds. Capabilities could advance rapidly while safeguards fail. Progress could slow, leaving the extinction timetable unsupported. Capabilities and effective controls could advance together. Serious harms short of extinction could also occur. One researcher’s financial decision cannot distinguish these paths, and the account does not establish a field-wide consensus behind his preferred forecast.

Evidence that systems can independently sustain dangerous actions, defeat meaningful oversight, and resist shutdown would strengthen the near-term catastrophic case. Reproducible limits on those abilities and independently verified controls would weaken it. For Soares’s ten-year forecast, the useful next document is an explicit chain from measured capabilities to catastrophe, with milestones and conditions for revision. Readers need a forecast that can change when the evidence changes, not a retirement decision to imitate.