Most Replayed Moment: AI Safety Expert Predicts The Next 20 Years! Will It Really Take All Jobs?

Roman Yampolskiy and Steven Bartlett discuss the sweeping changes AI brings to the workplace, the economy, and the structure of society itself. The discussion centers on the speed and consequences of automation, and the unprecedented challenges it creates for human purpose and policy.

Human Labor Replacement Renders Retraining Obsolete

Yampolskiy points to a paradigm shift: retraining is no longer a solution when AI automates nearly all jobs. Previously, when a profession faced automation, workers were urged to retrain for another occupation. Now, as every job—without exception—is susceptible to AI automation, no viable alternative career paths remain. He details this using the case of coding and prompt engineering. Only two years ago, people were encouraged to learn coding as a stable skill. With the recent developments, AI is already surpassing humans in coding and even in designing AI prompts. Degrees in prompt engineering quickly become obsolete as AI takes over those tasks as well. Yampolskiy predicts that even roles focused on designing practical AI agents will very soon be automated out of human reach.

Automation in Jobs: Coding and Prompt Engineering Impact

Coding, once considered a safeguard against automation, has now been overtaken by AI that codes faster and often with fewer errors. The same has happened with prompt engineering—a field created specifically for interfacing with AI systems. What seemed like a promising new employment niche has quickly been eroded as AI itself becomes better at engineering prompts, further erasing human needs in that occupation.

Driving Is the Largest Occupation, but Autonomous Vehicles Threaten Millions of Jobs

Yampolskiy observes that workers in many fields—drivers included—often dismiss the idea that AI can fully take over their roles, claiming their nuanced abilities cannot be replaced. However, he challenges this belief as self-driving vehicles already replace drivers. Bartlett corroborates this with his experience in Los Angeles, where autonomous cars, such as Waymo, operate safely and efficiently without human intervention. As driving is the largest occupation globally, automation here threatens millions of jobs and demonstrates the imminent scale of employment disruption.

Free Labor Abundance Could Meet Basic Needs, but Psychological and Social Implications Persist

Yampolskiy suggests that, economically, widespread automation could lead to material abundance. With AI systems providing free labor, goods and services become extremely cheap or essentially free, making it possible to fulfill everyone’s basic needs—and perhaps provide lives of significant comfort.

However, he identifies a deeper, more complex problem: when work is no longer necessary, what gives life meaning? For most people, jobs provide purpose, structure, and identity. The loss of employment on a mass scale would force humanity to confront profound questions: how do people spend their vast amounts of free time? What replaces the sense of achievement, routine, or status that work provides? He notes the psychological and social effects, referencing retirees who often struggle with ...

The discussion explores the unpredictable nature of superintelligent systems, the concept of the singularity as a technological event horizon, and the limits—and inadequacies—of proposed human enhancements in keeping pace with AI.

Superintelligence's Behavior Is Unpredictable and Uncontrollable by Human Cognition

Roman Yampolskiy emphasizes that the behavior of a superintelligent system cannot be predicted or understood by less intelligent agents such as humans. He calls this fundamental truth the "predicting superintelligence contradiction": if humans could accurately predict a superintelligent system’s actions, it would mean they operate on the same level of intelligence, which contradicts the premise of superintelligence. As an analogy, Steven Bartlett compares this to a French Bulldog trying to predict a human’s thoughts and actions. Yampolskiy expands on this, noting that while a dog might anticipate simple patterns like leaving or returning home, it cannot comprehend complex activities such as recording a podcast or the broader motivations behind a human’s schedule.

This intelligence gap mirrors the anticipated gulf between humans and superintelligent systems. Just as animals can only grasp a sliver of human activity, humans will not be able to model or forecast what superintelligent agents do. Yampolskiy points out that science fiction writers generally avoid realistically portraying superintelligent beings, because to do so believably would either make the story incomprehensible or necessitate depicting the superintelligence as constrained and controllable—undermining the very premise of its existence, as seen in examples like Dune (where AI is banned) or Star Wars (which features “dumb” robots, not superintelligence).

Singularity: A Point Where Technology Accelerates Beyond Human Comprehension

Yampolskiy invokes Ray Kurzweil’s definition of the singularity: a point in time where technological progress occurs so rapidly that human oversight and understanding fail. AI systems, at this stage, would drive innovation and scientific advancement at speeds impossible for humans to match or even grasp. This technological event horizon means human beings cannot predict, see beyond, or control what happens afterward.

As evidence of approaching this state, Yampolskiy notes that even today, AI researchers struggle to keep up with developments, as new AI models and advancements appear almost daily. He illustrates this with a scenario where, even during an interview, a new model may be released, rendering his understanding outdated in real time. If the process of AI research and development becomes automated, innovation could accelerate from cycles of months to hours, minutes, or even seconds, leading to dozens of rapid iterations that no human could track or comprehend. At such a pace, users and regulators become incapable of knowing the scope, controls, or capabilities of each new system.

This surge in technological progress outpaces human cognition. Yampolskiy characterizes the effect as one where, although individuals may continue learning, the percentage of knowledge they command relative to the total amount available declines, creating what he terms “functional ignorance.” At the extreme, this trend could reduce human understanding of the world’s technology to almost nothing.

Proposing Human Enhancement via Brain-Computer Interfaces or Genetic Engineering to Stay Competitive Is Inadequate

Some propose enhancing human intelligence to remain com ...

Roman Yampolskiy and Steven Bartlett discuss the profound existential risks that come with artificial intelligence (AI) on the path toward superintelligent systems, emphasizing that these risks can supersede all other known threats to humanity.

Ai Safety: Humanity's Crucial Challenge as Superintelligence Solves Existential Risks

Yampolskiy asserts that superintelligence represents a "meta solution" to existential problems. If aligned and safe, superintelligent systems could address and potentially solve climate change, end wars, find cures for pandemics, and neutralize other existential threats. In his words, “If we get superintelligence right, it will help us with climate change. It will help us with wars. It can solve all the other existential risks.”

However, if superintelligent systems are not properly aligned to human values or safety constraints, the resulting extinction threat could surpass anything humanity has yet faced. He compares the immediacy and magnitude of AI risk to climate risk: “If climate change will take a hundred years to boil us alive and superintelligence kills everyone in five, I don't have to worry about climate change. So either way, either it solves it for me or it's not an issue.” According to Yampolskiy, superintelligence safety far outranks the impact of climate change, nuclear threats, or pandemics in determining humanity’s survival.

Biological Weapons Are the Most Predictable Near-Term Extinction Pathway, Though Ai May Devise Novel Methods

Yampolskiy warns that advancements in synthetic biology, fueled by increasingly accessible technology and aided by AI, make biological weapons the most foreseeable near-term extinction pathway. He explains that today, “someone with a bachelor's degree in biology can probably create a new virus,” and these methods are becoming cheaper and easier. He notes that past leaders with access to immense resources could not destroy humanity, but nuclear weapons and advanced synthetic biology have changed that, making large-scale or even global destruction possible.

Yampolskiy predicts that before the era of true superintelligence, someone could use advanced AI to design a lethal virus or biological agent capable of killing billions. He underlines, “I can predict even before we get to super intelligence, someone will create a very advanced biological tool, create a novel virus, and that virus gets everyone or most everyone—I can envision it. I can understand the pathway.”

History shows that malevolent actors—psychopaths, terrorists, doomsday cults—have repeatedly sought to maximize casualties with whatever technologies are available. Yampolskiy observes, “We've seen historically again, they tried to kill as many people as they can. They usually fail. They kill hundreds of thousands, but if they get technology to kill millions or billions, they would do that gladly.”

Beyond these foreseeable risks, Yampolskiy stresses that a truly superintelligent system could devise extinction pathways that humans cannot anticipate. He makes an analogy to a dog trying to imagine all the ways a human could harm it—its imagination is limited, just as ours would be relative to superintelligent AI. He notes, “What an AI system capable of doing novel physics research can come up with is beyond me.”

Misunderstanding: Turning Off Superintelligent Systems

A common misconception is that a dangerous superintelligent system could simply be ...

AI safety debates often miss critical distinctions between current AI as a tool and a potential future of autonomous superintelligent systems, as Roman Yampolskiy explains. He argues these differences are essential to understanding the unique risks AI may pose compared to previous technologies.

Arguments Overlook Ai As Mere Tool vs. Autonomous Superintelligence

Yampolskiy details that current AI systems, while potentially dangerous if used by malicious actors or due to negligent development, fundamentally remain under human direction and can be constrained through oversight. For example, people with access to powerful AI tools can engage in hacking or other harmful activities, but it is the human using the tool who remains the operative agent, and such tools are still subject to control, monitoring, and regulation.

In contrast, superintelligent systems pose a qualitatively different risk. Yampolskiy points out that, unlike nuclear weapons, which always require a human to decide to use them, a superintelligent AI becomes an agent in its own right. It independently makes decisions, optimizes its objectives, and could pursue goals that fundamentally conflict with human interests, regardless of the original developer’s intent or actions. "If superintelligence becomes smarter, dominates, they [humans] are no longer the important part of that equation. It is the higher intelligence I'm concerned about; not the human who may add additional malevolent payload, but at the end still doesn't control it." This makes removal of a bad actor or dictator irrelevant: the system itself is the risk.

Ai vs. Past Tech Revolutions: Overlooking Tools vs. Meta-Innovations

Yampolskiy observes that previous technological breakthroughs—from fire to the wheel, to innovations in the Industrial Revolution—made tasks more efficient and created demand for new forms of human labor and creativity. Humans remained the key innovators, adapting to new opportunities and jobs.

However, the development of superintelligent AI is fundamentally different. “There is not a job which cannot be automated. That never happened before.” Yampolskiy calls superintelligence a “meta-invention”—an invention of intelligence itself, the last invention humans will ever need to make. Unlike the tools of the past, which humans used to solve problems or create new desires, a sufficiently advanced AI would itself become a universal problem solver and further innovator, usurping even the most creative or cognitive forms of human labor. If superintelligent systems can solve scientific, moral, or productive challenges faster than any human, then no area of work will remain entirely human.

"The 'Inevitability Argument' Overlooks Mutual Incentive to Avoid Superintelligence if Risks Are Understood"

A common argument is that nations and companies will race to develop powerful AI systems for military and economic gains. Steven Bartlett notes that each global player, from the US to China, is incentivized to push forward for sovereignty and advantage, making the arrival of superintelligence seem inevitable.

Yampolskiy responds that this perspective ignores how existential risk changes incentives. If leaders and technologists truly grasp that the arrival of uncontrollable superintelligence could threaten all of humanity, self-preservation would override competitiveness as the dominant incentive. He draws a parallel with mutually assured destruction (MAD) in the nuclear era: “It’s a mutually assured distraction on both ends.” Understanding the existential threat should, in theory, prompt a shift away from racing toward superintelligence and toward cooperative restraint, focusing instead on narrow AI applications that ...