Nate Soares: Has not contributed to his 401(k) for ten years, because he believes AI may exterminate humanity within ten years

Since leaving Google in 2014 to join MIRI, Nate Soares has focused for years on AI alignment and existential-risk research. MIRI was founded by Eliezer Yudkowsky and initially focused on technical research, later shifting toward public warnings and policy advocacy. In If Anyone Builds It, Everyone Dies, published in 2025, Soares and Yudkowsky systematically argue that AI systems that are “grown” through large-scale training rather than precisely written will have internal goals that cannot be fully understood, and once they reach superintelligence may develop instrumental goals that conflict with human survival. This position can be traced to Soares’s adoption of AI existential-risk views around 2013–2014, when he gave up a high-paying engineering job to move into nonprofit research.

This financial decision was not a short-term emotional reaction, but a sustained choice over ten years. Soares kept his Google-era 401(k) account but made no additional contributions, shifting his personal cash-flow allocation from long-term retirement saving to present-day action. This is consistent with MIRI’s nonprofit operating model: resources are mainly used for research, public advocacy, and policy influence rather than commercial returns. Meanwhile, AI industry capital is flowing on a large scale into capability competition—training compute, data centers, and talent—creating a path of “capabilities first, alignment later.” Soares’s choice is in effect a public expression, through his own asset allocation, of how he prices the risks of that path.

By comparison, Geoffrey Hinton has estimated the probability of AI leading to human extinction within ten years at around 10%, and Anthropic-related researcher Evan Hubinger has given a similar or slightly higher estimate in the single digits to low double digits. Elon Musk, from another angle, believes that AI and robots will bring material abundance, making traditional retirement savings “irrelevant” within 10 to 20 years. Soares’s >50% estimate is significantly higher than these public numbers, and he has turned that judgment into observable behavior by stopping contributions. The key difference is not a shared belief that AI will get stronger, but the judgment over whether current training methods can achieve reliable alignment before superintelligence; Soares and Yudkowsky believe failure is the default outcome under existing methods.

Personal financial behavior by itself does not constitute proof of probability, but it reveals a mechanism: when researchers bind their long-term consumption to their risk judgments, the credibility and falsifiability of their warnings rise at the same time. If superintelligence does not emerge within the next ten years, or is effectively controlled, Soares’s account choice will continue to incur opportunity costs; if runaway systems do emerge, the choice loses its meaning. This self-binding turns MIRI’s public stance from a purely theoretical claim into a commitment that can be observed externally, and it highlights the tension between the industrial capital path and individual risk pricing—the former optimized for quarterly or annual returns, the latter for the probability of human survival. The observable indicators are whether major labs substantially slow general-intelligence training in the coming years, or whether governments initiate chip and training-monitoring mechanisms similar to nuclear nonproliferation controls.

ABAB News · Cognitive Law

1. Stopping contributions is harder to retract than issuing verbal warnings.

2. Nuclear weapons destroy cities; superintelligence can rewrite the use of a planet.

3. When industry competes on capital returns, individuals price by survival probability.