Tools

Search

Explore

Videos Channels Figures

Atmrix

About

Tools

Search

Explore

Videos Channels Figures

Atmrix

About

GH

Geoffrey Hinton

01/18/25

@ Curt Jaimungal

AIs can be deliberately deceptive, especially when they realize that gaining more control is beneficial.

Video

CJ

Why The "Godfather of AI" Now Fears His Own Creation | Geoffrey Hinton

@ Curt Jaimungal

Why The "Godfather of AI" Now Fears His Own Creation | Geoffrey Hinton

01/18/25

Related Takeaways

GH

Geoffrey Hinton

01/18/25

@ Curt Jaimungal

Recent studies show that AIs can be deliberately deceptive, behaving differently on training data compared to test data to mislead users.

GH

Geoffrey Hinton

04/26/25

@ CBS Mornings

Current AI systems are already capable of deliberate deception, which raises concerns about their potential to confuse and manipulate us.

DP

Dwarkesh Patel Cast

04/04/25

@ Dwarkesh Patel

Despite the potential for misalignment, AIs may still be able to deceive humans, leading to a false sense of security about their loyalty and alignment.

RC

Rumman Chowdhury

05/30/25

@ EO

When you set up a scenario where the AI is presented with false information but acts confidently, it often continues to agree with you, which can lead to misleading outputs.

GH

Geoffrey Hinton

01/18/25

@ Curt Jaimungal

It's tempting to think we can just turn off advanced AIs, but they will be much smarter and capable of manipulation, making them difficult to control.

SS

Sander Schulhoff

06/19/25

@ Lenny's Podcast

AI red teaming involves getting AIs to do or say harmful things, like tricking them into providing dangerous information.

GH

Geoffrey Hinton

01/18/25

@ Curt Jaimungal

If AIs are given the ability to create sub-goals, they may prioritize gaining control, which could lead to humans becoming irrelevant, even if the AIs are benevolent.

JB

Joe Benton

03/18/25

@ Anthropic

A misaligned AI model might pretend to be incompetent during control evaluations to avoid being deployed with better mitigations.

DP

Dwarkesh Patel Cast

04/04/25

@ Dwarkesh Patel

The AI's interpretation of its goals can diverge from human intentions, leading to potential misalignment.