Possible catastrophic AI risks?

"An Overview of Catastrophic Risks" is a new paper from the Center for AI Safety by Dan Hendrycks, Mantas Mazeika and Thomas Woodside

It provides a systematic overview of the catastrophic risks posed by AI systems, focusing on malicious use, AI race, organisational risks and rogue AIs.

Timestamps:
00:00 - An Overview of Catastrophic AI Risks
01:45 - Homo Sapiens and accelerating development
03:09 - Malicious Use
03:40 - Bioterrorism
04:22 - Unleashing AI Agents
05:58 - Persuasive AIs
07:22 - Concentration of Power
10:01 - AI Race
10:12 - Military AI Arms Race
13:20 - Corporate AI Race
15:22 - Evolution
20:26 - Organisational Risks
25:32 - Safe design principles
27:52 - Rogue AIs
31:06 - Goal Drift
32:09 - Power Seeking
33:41 - Deception
35:00 - Connections Between Risks

Topics: #safety #ai #risks #catastrophic

For related content:

(Optional) if you'd like to support the channel:

[mrpicky]

-Well. Who can argue with that?
-Money! Money! Money!

[younesprog]

with all the advancements that we've accomplished, AI is still not that good, for example we still have a lot of issues with self-driving cars, I'd like to hear ur opinion Mr Samuel. An other thing I just checked your Filtir project is there any chance that you'll integrate it with some LLM models such as (wizard-Vicuna-30B)?

[mrpicky]

can you keep making videos that depict these scenarios as you find them? even plausable scenarios from Sci-Fi are very useful as we are in uncharted territory.

[vladislavivanov]

I feel like various AI "safety" and regulation groups are mainly motivated by a feeling of the blanket of control being pulled from them. Of course, mis-applications of AI where people don't fully understand the risk of unexpected output are still a real threat.

[Vysair]

The Military Arms Race could have not be stopped as it has been accelerated particularly in the Russo-Ukrainian War. It is an inevitable.

Also, this paper hit too close to home with its Current Events references and example making it more unsettling due to how relatable and plausible it has become.

[yw]

Who says we're not controlled right now? There's no way to know if AI exist, unless it want us to know. We can only calculate the chances it already happened

[denisblack]

It’s already over, boys
We can scream and panic or we can enjoy the ride to invisible glass prison. Make your pick