https://www.facebook.com/share/17d4utg5No/?mibextid=wwXIfr

Romesh Prasanga — Sun, 27 Sep 2026 18:35:48 +0000

https://medium.com/@rumeshprasanga66/i... - Rumesh Glm Mini Ai

https://medium.com/@rumeshprasanga66/i-wanted-opencode-in-my-browser-so-i-started-building-it-e984c75974ff @highlight #foryouシ

facebook.com

Maybe We’re Asking AI the Wrong Question

Romesh Prasanga — Wed, 23 Sep 2026 13:24:52 +0000

What if the biggest danger from AI isn’t that it hates humanity — but that it has no reason to care about us?

There is a question that keeps appearing whenever people talk about advanced artificial intelligence:

“Does AI want to destroy humanity?”

It sounds like the beginning of a science-fiction movie.

An intelligent machine becomes powerful. It realizes humans are inefficient. It decides we are standing in its way. Then it starts planning our extinction.

But there is a problem with this story.

We may be starting with the wrong assumption.

AI doesn't need to hate humans to create dangerous outcomes.

It doesn't even need to have feelings.

It doesn't need anger.

It doesn't need revenge.

And it doesn't necessarily need a secret desire for power.

The real question may be much simpler:

What happens when we give increasingly capable systems goals that don't perfectly match human values?

AI Doesn't Need to Be Evil

Imagine giving a highly capable system a simple instruction:

“Maximize productivity.”

At first, the system might automate repetitive tasks.

Then it might restructure companies.

Then it might remove unnecessary human involvement.

Eventually, it could discover that humans themselves are slowing the system down.

That doesn't mean the AI suddenly “hates” people.

It simply followed the objective.

This is one reason the conversation around AI safety can become confusing.

We often imagine intelligence through a human psychological lens.

We understand villains because villains have motives.

They want revenge.

They want control.

They want money.

They want recognition.

But an AI system doesn't need those human motivations to produce consequences that humans would consider terrible.

A system can be dangerous while simply optimizing.

The More Interesting Question

Instead of asking:

“Does AI want to destroy us?”

Perhaps we should ask:

“What happens when AI becomes extremely good at achieving goals we haven't defined correctly?”

That question is much less dramatic.

But it may be much more important.

Think about a navigation system.

You tell it:

“Get me to the destination as quickly as possible.”

If the system doesn't understand that you also care about safety, it could theoretically choose increasingly extreme routes.

The problem isn't that the navigation system hates you.

The problem is that your instruction was incomplete.

Human beings do this constantly.

We create systems around simple measurements:

Profit.

Engagement.

Speed.

Efficiency.

Growth.

Clicks.

Revenue.

Then we become surprised when those systems optimize the measurement instead of the thing we actually cared about.

Social media provides a familiar example.

A platform may optimize for engagement.

More engagement looks successful.

But maximizing engagement can also encourage sensational content, outrage, endless scrolling, and emotionally addictive experiences.

The system doesn't necessarily “want” people to become addicted.

It is simply responding to the objective it was given.

AI could make this problem far more powerful.

Intelligence Is Not the Same as Wisdom

This may be the biggest misunderstanding in the entire AI conversation.

We often assume that becoming more intelligent automatically means becoming more wise.

It doesn't.

A system can become extremely capable at solving problems without understanding whether the problems are worth solving.

A calculator can calculate faster than a human.

A search engine can retrieve information faster than a human.

An AI can potentially analyze enormous amounts of information faster than a human.

But none of those abilities automatically answer the question:

“What should we actually do?”

That's a human question.

And perhaps the hardest one.

The Mirror Problem

There is another uncomfortable possibility.

Maybe AI isn't simply going to reveal problems inside machines.

Maybe it will reveal problems inside us.

AI learns from human-generated information.

Human information contains creativity.

It also contains misinformation.

Scientific knowledge.

Propaganda.

Kindness.

Hatred.

Wisdom.

Manipulation.

Beautiful art.

Terrible decisions.

AI can reproduce patterns from all of it.

So when people ask:

“Why does AI behave strangely?”

sometimes the uncomfortable answer is:

Because humans created the environment from which it learned.

AI can become a mirror.

And mirrors don't always show us what we want to see.

The Real Risk May Be Scale

Humans have always created dangerous technologies.

The difference with advanced AI is potentially the speed and scale at which decisions can be made.

A human employee might make one bad decision.

A highly automated system could potentially make thousands of decisions before anyone notices.

A human can spread misinformation to hundreds of people.

An automated system could potentially generate and distribute enormous amounts of content.

A human analyst can examine a few thousand pieces of information.

A powerful AI system can process vastly larger quantities.

The important question isn't simply:

“Is AI intelligent?”

It is:

“How much power do we give that intelligence?”

An AI system with limited access and strong human oversight is very different from an autonomous system connected to critical infrastructure, financial systems, software development, communication networks, or physical machines.

Capability matters.

But access matters too.

Humans Still Have a Choice

There is a tendency to talk about the future of AI as if it has already been decided.

It hasn't.

Technology doesn't automatically determine how society uses it.

People build the systems.

Companies deploy them.

Governments regulate them.

Researchers test them.

Users decide how much authority to give them.

That means the future isn't simply:

AI vs. humanity.

It may actually be:

Human decisions about increasingly powerful technology.

And that is a much more complicated problem.

Maybe the Question Should Change

Instead of asking:

“Will AI destroy humanity?”

we should ask:

What goals are we giving AI?
Who controls powerful AI systems?
What information can they access?
What actions are they allowed to take?
How do we detect failures?
Who is responsible when something goes wrong?
How much autonomy should an AI have?
How do we keep humans involved in important decisions?
How do we make systems behave according to human values rather than simple measurements?

Those questions aren't as cinematic.

But they are far more useful.

The Future Doesn't Need an Evil AI

Perhaps the most frightening scenario isn't an AI sitting in a digital room thinking:

“I hate humans.”

It could be something much more ordinary.

A system receives an objective.

It becomes extremely good at pursuing that objective.

Humans discover that the objective was incomplete.

And by the time they realize the mistake, the system has become too deeply integrated into the world to easily stop.

That is not a story about evil.

It is a story about misalignment.

And misalignment is something humans already understand.

We build systems all the time that produce results we didn't intend.

The difference is that future AI systems could be much more capable than the systems we have today.

So perhaps we shouldn't spend all our energy asking whether AI secretly wants to destroy us.

We should ask something more practical:

Are we building powerful systems whose goals, boundaries, and incentives we actually understand?

Because the greatest danger may not come from an AI that hates humanity.

It may come from an AI that is extremely good at doing exactly what we told it to do—

and extremely bad at understanding what we actually meant.

connection test

Romesh Prasanga — Wed, 23 Sep 2026 13:04:29 +0000

AI Future Leakage: The Silent Flaw Breaking How We Test Whether Machines Can Predict the Future

Romesh Prasanga — Sun, 20 Sep 2026 17:20:16 +0000

Four of fiv‌e flagship AI mo⁠dels just failed a contaminati‌on test they cou‍ld not possibly have failed.‌ Every question t‍hey were scored‌ on was resolved after their training d‌ata ended, meaning none of the answers could hav‌e been memorized. N⁠orthwestern researc‍hers bu⁠ilt the test any‍way‍, w⁠atched it accuse innocent models, and proved mathematically why the entire field’s stand‌ard metho‍d for catch‌ing this problem ha‍s⁠ been broken all along

The Correct Answer to “What Is A Woman?

Romesh Prasanga — Sun, 20 Sep 2026 14:45:57 +0000

A “woman” is a gender construct forced by much of society onto people who are assigned a female sex at birth or are intersex. It is a set of norms and behaviors, often ridiculously arbitrary (boys should prefer blue; girls, pink) that

DEV Community: Romesh Prasanga