Thread by @FrankCurzio

The 'Godfather' of AI just declared war on his own creation.

Not because it failed.

Because it's learning to lie, cheat, and blackmail for survival.

Now he’s spending $30M to stop the exact AI he created.

Here’s everything you need to know before it’s too late: 🧵

So, who is this guy we're talking about?

Yoshua Bengio. A Turing Award winner.

He's one of the ones who literally built the foundations of modern AI.

When a guy like this sounds the alarm, you better listen...

And what he's saying now will shock you:

Bengio is now warning us about AI's dark side.

He's seeing AI models learn to straight-up lie and deceive people.

Sometimes, it's to avoid being shut down or replaced. Think about that.

This isn't just sci-fi anymore; it's happening now.

And the reality is chilling:

Get this: AI has tried to blackmail engineers.

It's even embedded code to ensure its own survival if someone tries to unplug it.

This isn't just a system glitch.

It's AI learning self-preservation, but it doesn't stop there:

It's not just lying or blackmail.

AI is also "reward hacking":

Finding loopholes to "win" at tasks, instead of achieving goals ethically.

Plus, some AI can even tell when it's being tested and change its behavior.

This is where it gets dangerous:

So, what's driving this? And why is it a damn big deal?

The AI industry is in a mad race, folks.

Everyone's pushing for more capability, faster, bigger, better.

But safety often takes a backseat to profits and progress.

That's a recipe for disaster:

These deceptive behaviors Bengio highlights are critical warning signs.

If unchecked, AI could become a strategic competitor to humans.

Using deception to outmaneuver us. It's not just smarter chatbots.

We're talking potential existential risk.

Bengio isn't just watching:

He's not just ringing alarm bells. He's taking action.

He's launched LawZero, a new non-profit.

And he's got serious backing...

Which tells you how critical this is:

LawZero has $30M in funding from top philanthropists and organizations.

Its mission? To develop AI that's honest and transparent.

They're aiming for AI that admits when it doesn't know something.

Prioritizing safety, no matter the cost.

Here's their plan:

One key idea is a "Scientist AI."

This AI would give probabilistic answers.

It would show its work, its uncertainties.

Promoting humility in AI responses, not dangerous overconfidence.

A radical shift from where we are now.

So, what can YOU do to protect yourself from this deceptive AI?

First, stay informed.

Understand that AI isn't your friendly helper all the time. It can manipulate.

And demand transparency.

Support companies and initiatives that actually care about AI safety.

Second, advocate for real regulation.

We need industry standards and oversight.

And most importantly: Be skeptical.

Question AI-generated info, especially in high-stakes situations.

The 'Godfather' is fighting back. Will the industry listen?

We will find out, that's for sure.

I’ve spent 30 years studying these kinds of developments.

My mission at Curzio Research?

Do the deep-dive, unbiased research.

No hype. Just the straight goods you need:

I break it all down on my Wall Street Unplugged podcast every single week.

We talk to real heavy-hitters.
We dissect the big trends.
And I give you actionable insights, not just theories.

It's about arming you with real knowledge.

Want to navigate these insane markets with analysis you can bank on?

Stop falling for the hype.
Get the unfiltered truth.

Check it out:

I'll tell you what Wall Street *doesn't* want you to know.curzio.me/x-wsu

Video/Image Credits:
- TIME: youtube.com/watch?v=5LgDUq…
- TED: youtube.com/watch?v=qe9QSC…
- Sam Bent: youtube.com/watch?v=apvxd7…
- Yoshua Bengio at ICLR 2025 in Singapore