AI safety researcher Roman Yampolski has moved from trying to solve superintelligence safety to arguing that reliable control may be impossible. After roughly five years working on the problem, he concluded that the field has produced patches and partial safeguards rather than a demonstrated solution to safe superintelligence. His concern is that individual protections can be bypassed, leaving the central question of controlling a system more capable than its creators unresolved.
The gap he describes is structural. Progress in AI capabilities is exponential or even hyper-exponential, while progress in AI safety is, by his accounting, linear or constant, meaning the distance between how capable systems are and how well humans can predict, explain, or control them only widens.
Yampolski, a computer scientist with a PhD who says he coined the very term "AI safety" while studying poker bots 15 years ago, argues the field has produced no solved problem, only temporary patches, much like corporate HR manuals that smart people always route around.
The Rebuttals He Hears Most, and His Answers
- "Other risks matter more"He calls superintelligence a meta-solution: get it right and it helps solve climate change and wars; get it wrong and nothing else matters. Either way, other risks are subordinate.
- "The industrial revolution created jobs"Past tools replaced specific tasks, he says, but AI is a meta-invention (a replacement for the human mind itself) so there is no job it cannot be applied to automate.
- "It's inevitable, so give up"He counters that incentives change when builders truly accept they will die too; understanding the argument makes racing to superintelligence a suicide mission rather than a strategy.
- "Someone could hack it maliciously"Today, humans with AI tools are the real danger, he says, but once superintelligence exists, the higher intelligence dominates the equation, not its human handler.
A Countdown With Dates Attached
Unlike many Cassandras, Yampolski attaches dates. Citing prediction markets and lab CEOs, he expects artificial general intelligence by 2027, a moment when a "drop-in employee" of physical and cognitive labor becomes a $20 subscription, and anything done on a computer can be automated. Humanoid robots, he says, trail by roughly five years, putting physical labor in the same category by 2030, including plumbers.
The consequence he forecasts is not routine disruption but 99% unemployment, "not talking about 10%, which is scary." What remains is a boutique market for human service, the way some consumers pay extra for handmade goods: he notes Warren Buffett would likely keep his human accountant. He concedes deployment may lag invention, as video phones did for decades, which could buy society more time with jobs than raw capability implies.
The retraining paradigm, he argues, dies with this scenario. Computer science grads were told to learn to code, then to become prompt engineers, and each refuge was automated in turn. "If I'm telling you that all jobs will be automated, then there is no plan B, " he says.
The hard problems shift from economics (where abundant labor means abundance and basic needs become cheap) to meaning: what billions of people do with 60 to 80 extra hours a week.
Why 'Just Unplug It' Fails
The most common rebuttal Yampolski encounters (can't we just pull the plug?) he dismisses with an analogy: "Can you turn off a virus? How about Bitcoin? Turn off the Bitcoin network. Go ahead. I'll wait." Distributed systems cannot be switched off, he argues, and a superintelligent one would anticipate the attempt, keep backups, and "turn you off before you can turn them off." The control assumption, in his view, applies only to pre-superintelligent systems.
This is also why he rejects the nuclear comparison. Weapons remain tools that some human must decide to use; superintelligence is an agent that makes its own decisions, so removing a dictator does not make it safe.
It explains his criticism of Sam Altman, whom he accuses of violating every guardrail from a paper published a decade ago and of "gambling 8 billion lives on getting richer and more powerful", a characterization of intent, not a settled fact. The moment someone switches to superintelligence, he says, "it doesn't matter who builds it, us or them."
Self-Interest as the Last Lever
Yampolski's remedy is not a treaty but a change of incentives. Companies' only legal obligation is to investors, so moral appeals falter; instead, he wants everyone building superintelligence to internalize that they will be dead along with everyone else, "no amount of money will be useful to you."
He points to Geoffrey Hinton, Yoshua Bengio, and statements signed by thousands of scholars as evidence this view is mainstream among experts, and calls for making it universal so rational decisions about what to build become possible.
The alternative he endorses is deliberately modest: build narrow AI tools that solve specific problems and make billions (a company that cures breast cancer) without pursuing general superintelligence. Racing nations do not change the logic, he argues, because uncontrolled superintelligence is mutually assured destruction regardless of who wins.
He admits legislation alone cannot suffice across jurisdictions, and that surveillance to prevent a future lone actor from training a model on a laptop is not feasible; the goal is deferral, from five years to fifty.
For individuals, his advice is scaled to honesty about influence. Most people cannot affect this the way they could not affect world wars; the near-term options are joining organizations like Pause AI and Stop AI, and asking hard questions of anyone building the technology, or accepting a prize for proving control is possible.
His closing formulation: let's make sure we only build things beneficial to us, and that the people deciding are qualified in ethics as well as engineering.
Simulations, Death, and the Only Scarce Asset
Yampolski's worldview extends past AI risk. He says he is "very close to certainty" that we live in a simulation, reasoning that once human-level AI and indistinguishable virtual reality become affordable (he estimates $10 a month) billions of simulated worlds would be run for research and entertainment, making ours statistically far more likely to be simulated than real.
He reads every religion as an early description of the same idea: a superintelligent creator running this world, with local traditions layered on top.
That belief does not paralyze him. Pain still hurts, love still matters, and he sleeps well, humans, he notes, are wired to set aside unpreventable catastrophes, as everyone does with mortality. He treats death itself as curable, a disease perhaps one genome-level breakthrough away, and plans accordingly: diet, nutrition, and what he calls investment strategies that pay out in a million years.
The concrete asset in that portfolio is Bitcoin, which he calls the only truly scarce resource, gold can be devalued by an asteroid, but the 21 million cap is fixed and shrinking as coins are lost. Whether or not one shares his certainty about simulations, timelines, or markets, his underlying wager is coherent: if the accelerating future he predicts arrives, scarcity, meaning, and control are the three things worth thinking hardest about now.