Dietmar Fischer

But an advanced AI agent could plan, adapt, use tools, and act in the world.

For Yudkowsky, that changes everything.

A chatbot that answers questions is one thing.

An AI agent that can write code, send messages, use software, manage tasks, call APIs and make decisions is another.

The more autonomy and access we give these systems, the more important alignment becomes.

And this is why Yudkowsky is not reassured when people say, but today's AI still makes mistakes.

His answer would be, current weakness does not prove future safety.

A tiger cub is also not very dangerous.

That does not mean you should raise one in the office kitchen next to the oat milk.

His worry is the direction of travel.

AI systems are becoming more capable.

They write better code, solve harder problems, use tools more effectively, and operate across more areas of life and business.

His question is, are we getting better at controlling them as quickly as they are getting better at doing things?

He thinks the answer is no.

That is why his position is much more extreme than the usual AI needs regulation argument.

He has argued that humanity should not build systems that could become super intelligent until we know how to make them safe.

Many people think this is unrealistic.

Some think it is too pessimistic.

Some argue that stopping development globally is impossible.

These are serious objections.