THE DIARY OF A CEO · EXTRACTED

The AI Safety Expert: Only 5 Jobs Will Survive by 2030 ft. Dr. Roman Yampolskiy

Why controllable superintelligence is mathematically impossible, what 99% unemployment actually looks like, and how to think about the next 20 years when the people building it don't know what's inside it.

▶20.6M views on YouTube

"While progress in AI capabilities is exponential or maybe even hyper exponential, progress in AI safety is linear or constant. The gap is increasing." — Dr. Roman Yampolskiy
Dr. Roman Yampolskiy, on the episode

Dr. Roman Yampolskiy is a professor of computer science and engineering who coined the term "AI safety" in 2010, roughly a decade before it entered mainstream conversation. He began his work studying poker bots, noticed that narrow systems were improving on a trajectory that pointed somewhere alarming, and spent the next fifteen years trying to solve the control problem before concluding it is almost certainly unsolvable. The popular framing of AI risk is that a few bad actors might misuse a powerful tool. Yampolskiy's actual model is sharper: the tool-versus-agent distinction means that once a system is smarter than its creators, no one is holding the leash, regardless of intent. This protocol pulls from his core technical argument about the capability-safety gap, his timeline predictions for AGI and humanoid robotics, his analysis of why retraining advice is now structurally broken, his view on what individuals can actually do, and his parallel thinking on simulation theory, longevity, and Bitcoin as a scarce-resource hedge.

Tactic 01

Use The Capability-Safety Gap As Your Primary Decision Signal

Yampolskiy's central technical argument is not that AI is dangerous in the abstract. It is that progress in AI capability is exponential, possibly hyper-exponential, while progress in AI safety is linear or flat. The gap between what these systems can do and what we can predict, explain, or control about them is widening every year. That gap is the signal. He traces this to how modern AI is built. The systems are not programmed with explicit rules in the way software was for the first fifty years of computing. They are trained on massive datasets and then studied like an artifact from the outside. "We are creating this artifact, growing it," he says. "It's like an alien plant. And then we study it to see what it's doing." Even the engineers building these systems run experiments on their own products to discover what those products are capable of. New capabilities are still being found in old models. Nobody can tell you precisely what a system will do given a set of inputs. The practical implication is that any safety claim made by an AI company should be evaluated against this gap, not taken at face value. "The state-of-the-art answers are, we'll figure it out when we get there, or AI will help us control more advanced AI," Yampolskiy says. "That's insane." The gap also means that patches work only briefly. Jailbreaks appear. Restricted behaviors migrate to unrestricted subdomains. The HR-manual analogy he uses is precise: a smart enough employee always finds a workaround to any written policy.

The play

When you encounter a safety claim from an AI company, ask one question: where is the peer-reviewed paper explaining the specific mechanism? Not a press release, not a Senate testimony, not a blog post. A published, falsifiable, scientific explanation of how controllability is achieved. If none exists, treat the safety claim as unverified and factor that into any decision you are making about the technology, whether as a user, investor, employee, or voter.

Tactic 02

Plan Around AGI By 2027 And Physical Labor Automation By 2030

Yampolskiy is not speculating about distant futures. He is working from prediction markets, the public statements of lab CEOs, and his own fifteen years of watching narrow systems improve on a curve that now points at general capability within a two-to-three year window. His specific call: artificial general intelligence by approximately 2027, with humanoid robots capable of competing with humans in physical domains roughly five years behind that, placing full physical labor automation around 2030. The 2027 date matters because AGI, in his framing, means a drop-in replacement for most cognitive work. "If I can just get a twenty-dollar subscription or free model to do what an employee does, first, anything on a computer will be automated." The 2030 date matters because that is when the combination of cognitive and physical capability removes the last practical refuge. "The combination of intelligence and physical ability really doesn't leave much, does it, for us human beings. Not much." He is careful to note that capability and deployment are not the same thing. Video phones were invented in the 1970s and did not reach widespread use until the iPhone era. The capability to replace most humans in most jobs may arrive on these timelines while actual displacement lags. But the planning horizon is set by when the capability exists, not when it propagates.

The play

Use 2027 as a planning horizon for cognitive work and 2030 for physical work. For any career, education, or investment decision with a payoff window longer than two years, ask explicitly whether that payoff depends on a human doing something an AGI system could do for twenty dollars a month. If it does, either the timeline is compressed or you need a non-automation-dependent version of the plan.

Tactic 03

Stop Advising Retraining Into Specific Occupations

The standard response to automation anxiety has always been retraining. Learn to code when manufacturing jobs disappear. Become a prompt engineer when creative work is automated. Yampolskiy's argument is that this advice is now structurally broken, not because the new jobs are bad jobs, but because the replacement career can be automated before the retraining is complete. He walks through it explicitly. Two years ago the advice was to learn to code, because artists couldn't make money. Then AI became a capable coder. The advice shifted to prompt engineering, framed as a four-year degree worth of skill. Then AI became better at designing prompts for other AIs than any human. The current hot recommendation is designing AI agents for practical applications. He says to expect that to be gone within a year or two as well. "I can't really tell you right now, the hottest thing is design AI agents for practical applications. I guarantee you in a year or two it's going to be gone just as well." The paradigm shift underneath this is that all previous automation created tools. A tool does one thing and stops. The industrial revolution did not invent a new worker; it invented a more efficient way to do existing work, freeing humans to do new work. AGI invents a worker. A worker who can be applied to whatever new job the retraining was supposed to produce. "There is not a job which cannot be automated. That never happened before."

The play

If you are giving career advice to someone facing displacement, or receiving it yourself, stop at the point where the new career is named and ask: can an AGI system do this job in two years? If the honest answer is probably yes, the advice is directing someone into a temporary shelter, not a durable position. The more useful frame for that person is not which job to pivot to, but how to think about the economic and meaning questions that arrive when the pivot options close.

Tactic 04

Treat AI As Science, Not Engineering, When Evaluating It

For the first fifty years of computing, building software meant knowing what you were building. A knowledge engineer specified the rules, programmed the expert system, and the outputs followed from the inputs in a traceable way. Modern AI does not work like this, and Yampolskiy argues that the biggest source of false confidence in current AI development is the failure to internalize that shift. Today, even the people building these systems have to run experiments on their own products to learn what those products can do. "Oh, do you speak French? Can you do mathematics? Are you lying to me now?" He describes a model that takes a year to train and then six months of experiments to develop a basic map of its capabilities, with new capabilities still being discovered after the fact. Ask a question in a different way and the system becomes smarter. The same model, different prompt, different capability profile. That is not engineering. That is science conducted on an artifact that was grown rather than designed. "It's like an alien plant," he says. "And then we study it to see what it's doing. And just like with plants, we don't have 100% accurate knowledge of biology. We don't have full knowledge here." The practical consequence is that anyone claiming to know what a given AI system will do in a novel situation is overclaiming. The correct assumption for evaluation purposes is a black box with a partial and growing capability map, not a deterministic machine.

The play

When evaluating any AI system, whether for adoption, investment, or policy, assume it is a black box with an incomplete capability map. Run experiments rather than accepting stated capabilities. Ask what the system does when the question is framed differently, what it does at the edges of its stated domain, and what mechanisms exist for discovering capabilities that have not yet been found. Treat any vendor description of what the system does as a hypothesis to test, not a specification.

Tactic 05

Issue A Public Controllability Challenge To AI Developers

Yampolskiy's argument is that AI companies are making safety claims they cannot substantiate with published science. His proposed mechanism for exposing this is a public challenge with stakes. Not a debate, not a regulatory hearing, but an open offer: whoever can demonstrate, in scientific terms, that they have solved the controllability problem gets the prize, the platform, the attention. The logic is simple. These companies are valued at hundreds of billions of dollars partly on the basis of safety claims. If those claims are real, someone should be able to claim the prize. If no one claims it after a reasonable window, the absence of claimants is itself evidence. "We have companies valued again at billions and billions of dollars working on safe superintelligence. We haven't seen their output yet." He applies the same logic to his own work on impossibility. His goal is not to demoralize researchers but to establish that controllability of indefinitely self-improving superintelligence is an impossible problem in a specific technical sense, not a difficult one. Once something is established as impossible, fewer people waste resources claiming they can do it for a fee. The challenge mechanism is the public version of that: put the claim to a test with real stakes and see who shows up.

The play

If you have a platform, issue a standing open challenge: any AI developer who can present a peer-reviewed, scientifically specific explanation of how they will maintain control of a superintelligent system as it self-improves is invited to make that case publicly. Set a window of one to two years. Document who engages and who does not. The pattern of non-engagement is informative regardless of whether anyone claims the prize.

Tactic 06

Concentrate AI Development On Narrow-Domain Tools, Not General Agents

Yampolskiy is not anti-technology. He uses AI constantly and describes himself as a scientist and engineer who loves the field. His argument is not that AI development should stop. It is that the specific path of building general agents capable of recursive self-improvement is the dangerous one, and that enormous economic value can be extracted from narrow AI tools without taking that path. "Okay, my company cures breast cancer. That's all. We make billions of dollars. Everyone's happy. Everyone benefits. It's a win." The narrow-tool path preserves human control because narrow systems do not generalize. A protein-folding model is superhuman at protein folding. It does not parlay that into unanticipated capabilities across other domains. The controllability problem is tractable at the narrow level in a way it is not at the general level. He adds a scale argument. Current models, even without any further development, represent roughly 1% of their deployed economic potential. The infrastructure, workflows, and integrations needed to extract value from what already exists have not been built. "We make AI so quickly, it doesn't have time to propagate through the industry, through technology." He estimates that 60% of existing jobs could be replaced by models that already exist. The race to superintelligence is not necessary for economic growth. It is being run for other reasons.

The play

If you are building in AI, evaluate your project against one question: does this require a general agent capable of operating across arbitrary domains, or can it be accomplished with a system constrained to a specific problem? If the narrow version solves your problem, build the narrow version. If you are investing in AI companies, apply the same filter: narrow-domain tools with a clear capability ceiling present a different risk profile than general-agent development, both financially and in terms of the externalities you are funding.

Tactic 07

Target Longevity Escape Velocity As A Practical Goal

Yampolskiy treats aging as a disease with a cure, not a fixed constraint. His position is that nothing in biology makes death at a certain age inevitable. What exists is a rejuvenation loop somewhere in the genome that appears to be set to a maximum of roughly 120 years, and that reset point can in principle be changed. He thinks the breakthrough is one discovery away and that AI will accelerate the timeline significantly because it can process genomic data at a scale no human research team can match. The practical concept he points to is longevity escape velocity: the point at which medical advances are adding more than one year of life expectancy per year lived. Once you cross that threshold, you do not need to solve aging completely. You just need to survive long enough for the next advance. He describes it as a relay race where each breakthrough carries you to the next one. Brian Johnson's framing of "don't die now" is the colloquial version of the same idea. He connects this to investment strategy explicitly. If you believe you might live significantly longer than standard life tables suggest, the time horizons relevant to your financial decisions change. Plans that pay out in thirty years look different if your planning horizon is eighty years. He describes himself as actively thinking about investment strategies that pay out over long time periods, precisely because he takes the longevity trajectory seriously.

The play

Identify the three most controllable inputs to your health span: diet, sleep, and one domain-specific intervention relevant to your current biology. Manage them actively rather than as background variables. Separately, if you are making financial or career decisions with fixed time horizons, run a version of those decisions where your planning horizon is twenty to thirty years longer than your current assumption. The goal is to stay biologically viable long enough for AI-accelerated genomics research to reach the escape velocity threshold.

Tactic 08

Hold Bitcoin As The Only Provably Scarce Asset In An Abundance Economy

Yampolskiy's investment thesis for Bitcoin follows directly from his broader economic model. If AGI produces effectively unlimited free labor, then the price of most things that require labor to produce collapses. Abundance becomes the default. In an abundance economy, the things that retain value are the things that cannot be reproduced regardless of price. Gold fails this test: given a high enough gold price, mining operations scale up, and theoretically an asteroid made of gold could arrive and destroy the value entirely. Bitcoin is different because the supply is capped at 21 million coins by protocol, and that cap cannot be changed by economic incentives. He adds a scarcity-within-the-cap point. The practical supply is already below 21 million because coins are lost permanently when passwords are forgotten or wallets are abandoned. The lost coins include a significant portion held by the pseudonymous creator. The effective supply shrinks over time while demand, on a long enough horizon, can only increase as more of the global economy shifts to digital assets. He acknowledges the quantum computing risk directly. A sufficiently powerful quantum computer could break the cryptographic algorithm that secures Bitcoin holdings. His answer is that quantum-resistant cryptography migration strategies exist and that quantum computers remain weak enough that this risk is not yet urgent, but it is worth monitoring.

The play

If you hold Bitcoin or are evaluating it, monitor the development of quantum computing capability on a timeline of five to ten years and track the progress of quantum-resistant cryptography standards. The migration path exists and the threat is not immediate, but it is real enough to warrant a planned response before quantum computers mature. Separately, when thinking about Bitcoin's role in a portfolio, evaluate it specifically as a scarce-resource hedge in an AI-driven abundance economy rather than as a speculative technology bet.

Tactic 09

Filter Global Catastrophe News As A Deliberate Mental Health Practice

Yampolskiy makes a point that sits outside AI safety but follows from everything else in the conversation. Humans evolved to process bad news at a tribal scale. A dozen people around you, a handful of bad events per year, all of them relevant to your immediate survival. The internet has replaced that with thousands of reported deaths, disasters, and catastrophes per day, most of them entirely outside your ability to influence. The result is not heightened awareness. It is a system processing inputs it was never designed to handle, producing anxiety without agency. "If I go on the internet, somebody gets killed everywhere all the time. Somehow, thousands of people are reported to me every day. I don't even have time to notice. It's just too much. So, I have to put filters in place." His own practice is to concentrate on what he can change and not attempt to hold the full weight of global catastrophe as a personal emotional burden. This is not denial. He spends his professional life working on what he believes is an existential threat. But he separates the work of trying to influence the outcome from the consumption of distressing information about outcomes he cannot influence. The two are not the same activity and treating them as the same produces the worst of both: paralysis without impact.

The play

Set a deliberate daily limit on consumption of global catastrophe news. One specific check of credible sources per day is sufficient to stay informed. Above that threshold, you are no longer gathering information that changes your decisions. You are processing tribal-scale emotional inputs at global scale, which is a mismatch your nervous system was not built for. Apply the recovered attention to whatever you can actually influence in your immediate environment.

Tactic 10

Extract The Common Core Across Religions Rather Than Their Local Traditions

Yampolskiy arrives at a specific and operational reading of religious traditions through the simulation hypothesis. His argument is that all major religions, stripped of their local customs and dietary rules and scheduling conventions, describe the same underlying structure: a superintelligent being who created this world, the idea that this world is not the primary one, and some form of consequence that extends beyond this life. He uses a thought experiment to make the point. Take the simulation hypothesis paper, translate it into a local language, and explain it to a tribal community with no exposure to contemporary science. Return two generations later. You find a religion. "That's basically what the story is." The local traditions that differentiate religions from each other, which day not to work, which animals not to eat, which rituals to follow, are local flavors layered on top of a theory that is structurally identical across all of them. The practical implication is that if you dismiss religion wholesale because of its local traditions, you are discarding a consistent cross-cultural signal about the nature of reality that has persisted across thousands of years of human history. The more robust move is to separate the layer that varies by geography and era from the layer that all traditions share, and to take the shared layer seriously as a data point.

The play

When evaluating any religious tradition, run a two-layer filter. Identify the local traditions: dietary rules, specific holidays, institution-specific practices. Set those aside as cultural artifacts. Then identify what that tradition shares with every other major tradition: a higher intelligence, the non-primacy of this world, consequences beyond this life. Weight the shared layer more heavily than the local layer. If you find meaningful overlap in the shared layer across traditions you previously treated as contradictory, update your prior about what those traditions are pointing at.

Your action plan

the plays, one checklist.

01
Use The Capability-Safety Gap As Your Primary Decision Signal
Demand a published, peer-reviewed safety mechanism before accepting any AI company's safety claims as verified.
02
Plan Around AGI By 2027 And Physical Labor Automation By 2030
Treat 2027 as the cognitive automation horizon and 2030 as the physical one when evaluating any career or investment with a multi-year payoff.
03
Stop Advising Retraining Into Specific Occupations
Before recommending any retraining path, ask whether an AGI system can do that job within two years, and be honest when the answer is probably yes.
04
Treat AI As Science, Not Engineering, When Evaluating It
Treat any AI system as a black box with an incomplete capability map and run experiments rather than accepting stated capabilities.
05
Issue A Public Controllability Challenge To AI Developers
Issue a public, standing challenge inviting any AI developer to demonstrate controllability with published science, and treat non-engagement as evidence.
06
Concentrate AI Development On Narrow-Domain Tools, Not General Agents
Build or back narrow-domain AI tools with a clear capability ceiling rather than general agents, and recognize that most economic value can be extracted without crossing into general intelligence.
07
Target Longevity Escape Velocity As A Practical Goal
Actively manage the controllable inputs to your health span and extend your financial planning horizon on the assumption that medical breakthroughs will add more than one year of life per year lived.
08
Hold Bitcoin As The Only Provably Scarce Asset In An Abundance Economy
Hold Bitcoin as a scarce-resource hedge and monitor quantum computing development so you can migrate to quantum-resistant cryptography before the threat becomes urgent.
09
Filter Global Catastrophe News As A Deliberate Mental Health Practice
Cap global catastrophe news consumption to one credible daily check and redirect the recovered attention to whatever you can actually influence.
10
Extract The Common Core Across Religions Rather Than Their Local Traditions
Separate local religious traditions from the cross-tradition shared core, and weight the shared core, higher intelligence, non-primacy of this world, consequence beyond life, more heavily as evidence.