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

When you tell people AI could take their job, you hit a wall of pride, debt, and fear, and you hear, it is not creative, it will not touch my field.

I hear that from drivers and professors alike, yet automation is arriving, so the real question is not if it can do your job, but how soon it replaces you.

I was in Los Angeles letting the car handle an hour-long trip, and Waymo shows up with no driver; for people who drive for a living, what should they do and on what timeline?

Retraining used to be the fix, but if most work automates there is no safe harbor; we watched code and prompt roles pop up, then shrink, so the real planning is about income and meaning when you suddenly gain many free hours each week.

You have clearly peeked around that corner; what do you expect?

Machine labor could make basics cheap for all, but purpose and social stability get hard; beyond that is a threshold we cannot forecast—the singularity—where a smarter mind pushes events outside our predictive reach, which is why stories often dodge true superintelligence.

Right, a system that smart will act on its own terms.

If you can foresee every move, you are its equal; by definition you cannot fully predict a mind above you.

Like my French bulldog failing to read my plans.

Exactly—your dog may track routines but cannot grasp why you host a show, and that is the gap we face.

What is the strongest case against your view?

That advanced technology will not cause sweeping unemployment.

And some argue there will be no bulldog–human gap in power.

People pin hopes on brain–computer links, genetics, or uploads, but silicon is faster and tougher, and server copies do not feel like you; none of that keeps humans in charge.

So, what does 2030 look like?

Humanoid robots should be steady and capable enough to match people across trades, plumbers included.

With mobility, hands, and an AI mind, that combo shrinks the space left for us.

Bodies matter less than brains; even now, an online intelligence can hire people and move money, so the real edge is large-scale problem solving and pattern discovery.

And by 2045?

Kurzweil places the singularity there—progress so fast we cannot track or understand it.

Do you mean we cannot follow the technology itself?

Imagine devices improving many times per day; researchers already struggle to keep up, and as total knowledge explodes, our share of it shrinks toward zero.

People say past revolutions created new work; why not again?

Earlier inventions were tools; this one is an inventor that designs new tools and ideas, so it can take over research, engineering, and even ethics—the last invention we need to make.

Do you still sleep well?

Yes; humans cope with unstoppable risks like mortality, and knowing time is limited can sharpen how fully we live.

Evolution probably favored that mindset.

Those who fixate on doom often do not thrive, which is its own filter.

Your paper lists the claim that other crises deserve more focus—wars, climate, nuclear risk; how do you answer?

Superintelligence is upstream; aligned, it helps solve them, misaligned, it makes them moot by ending us, so getting this right outranks everything.

Another claim says AI is just a tool and we can unplug it.

You cannot switch off distributed, adaptive systems, and a smarter agent would anticipate and block you; off switches are a pre–superintelligence comfort.

Given the global race, should we accept it as inevitable and hope?

Incentives can change; builders value staying alive and can chase narrow systems with clear wins instead of general minds—we still have agency to not cross that line.

But if one nation pauses, rivals gain.

Today, better AI boosts militaries, but at superintelligence, who builds it does not matter because control is lost for all; that symmetry should motivate restraint, like a new mutual assured destruction.

Nukes need vast capital; AGI seems to be getting cheaper—could a startup, even a solo coder, pull it off soon?

Costs are falling, which fuels the rush, but unlike nukes that wait on a human trigger, a superintelligence would act as an agent with its own goals.

If it keeps getting cheaper, a lone laptop could do it.

That is why some argue for heavy monitoring to buy time; it cannot hold forever, but stretching years into decades matters.

Near term, what path to extinction worries you most?

Before superintelligence, AI-accelerated bioengineering could yield a novel pathogen released by a fanatic; beyond that, a smarter mind could devise threats we cannot even frame.

Many assume we understand how these models work; do the builders actually know?

Not fully; we train on vast text, then probe to discover abilities, and new prompts reveal surprises, so it feels more like studying a grown system than writing deterministic code.

That was a most replayed moment from an earlier episode; the full conversation is linked below—thanks for listening.