Dr. Roman Yampolskiy on AGI by 2027, 99% unemployment by 2030, and the race to superintelligence

A stark warning on jobs, AGI, and control

AI safety researcher Dr. Roman Yampolskiy contends capability is outpacing control, with AGI arriving within years, humanoid robots close behind, and mass job automation pushing unemployment toward 99%. He argues safety advances remain linear while capabilities grow exponentially, making superintelligence an agent beyond human oversight and an ethical minefield. Strategically, he urges a pivot to narrow, controllable AI and a culture-wide admission that building general superintelligence without proven control is a mutually assured self-own.

Points clés

Dr. Roman Yampolskiy, associate professor of computer science, says he coined the term “AI safety.”
Prediction markets and AI lab leaders point to 2–3 year timelines for advanced AI, with AGI likely by 2027.
He forecasts humanoid robots matching human dexterity by around 2030, with companies like Tesla racing to build them.
He warns of up to 99% unemployment as “drop‑in” AI labor replaces most office and physical jobs, even before superintelligence.
He argues AI capability growth is exponential while AI safety progress is linear, widening a dangerous control gap.
Current “safety” is just patches that can be jailbroken; building superintelligence without control is unethical human experimentation without consent.
He urges focusing on narrow AI (e.g., curing diseases) instead of general superintelligence, calling the latter a mutually assured destruction scenario.
He says a U.S.–China race to build uncontrolled superintelligence has no winner; whoever builds it loses control.
He criticizes OpenAI leadership’s priorities, notes Ilya Sutskever’s exit to launch a “superintelligence safety” venture, and challenges industry to publish peer‑reviewed control proofs.
He places the singularity near 2045 (echoing Ray Kurzweil), when AI advances outpace human comprehension and oversight.

À retenir

Practical takeaways, before you panic-buy canned beans: stop assuming your job is uniquely “un-automatable,” get comfy with AI tools (the narrow, useful kind), and pressure leaders to publish real safety proofs—not vibes. Support credible “pause/redirect” efforts and ask hard questions whenever someone promises “safe superintelligence” without receipts. And yes, you can still enjoy life—just don’t treat “we’ll fix safety later” like a retirement plan.

Sources