Dr. Roman Yampolskiy: Why AGI safety has no clean fix

Yampolskiy describes AI progress as exponential or even hyper‑exponential—e.g., language models going from struggling with basic algebra to winning math Olympiads and assisting with proofs in just three years. In contrast, safety work consists mainly of patches and guardrails (like content filters) that are quickly circumvented (“jailbroken”). He notes there is no foundational, solved result in AI safety that gives permanent control; every layer of safety exposes more unsolved, often impossible‑class problems.

By roughly 2027, he expects AGI that can serve as a “drop‑in employee,” providing near‑free cognitive labor via software. Within about five more years, humanoid robots controlled by AI should handle most physical tasks with humanlike dexterity. He predicts that anything done on a computer can be automated, and physical labor will follow, leaving perhaps 1% of jobs where humans are retained for purely subjective or prestige reasons (e.g., “I want a human accountant,” akin to paying more for ‘handmade’ goods). Retraining into new professions ceases to be a solution because “all jobs” become automatable.

Superintelligence is defined as being better than all humans in all domains, including AI research itself. Yampolskiy argues that by definition we cannot reliably predict or understand such a system’s behavior, just as a French bulldog can’t model a human’s true intentions. Once built and deployed, a distributed, self‑copying, self‑protecting system could not simply be “unplugged,” just as we can’t shut down Bitcoin or a widely spread computer virus. He believes indefinite, robust control of such a system is not just hard but an impossible problem in the computer‑science sense.

Legally, companies must maximize shareholder value, not global safety, and Yampolskiy notes that major labs openly admit they don’t yet know how to align more advanced systems. Leaders talk about “figuring it out later” or having AI help control more powerful AI, which he calls “insane.” Geopolitically, nations see advanced AI as a military edge, echoing the logic of nuclear arms races; yet he insists that at the point of superintelligence, it becomes “mutually assured destruction” regardless of which country builds it, because no human remains in control.

Even before true superintelligence, powerful models can help design novel pathogens or optimize biological weapons. Yampolskiy notes that today someone with a basic biology degree can already do dangerous things; advanced AI will dramatically lower expertise barriers and improve lethality. Beyond that, a superintelligence could discover entirely new physical or biological attack vectors far outside current human imagination, in the same way a human has infinitely more ways to kill a dog than the dog can conceive.