How could Artificial Intelligence become dangerous?

[domin]

I’m not in a rush to start using it

 

[BackRoad]

I have a great nephew, (great as in son of a nephew) who is in charge of a nearby small city police force. He retires in 12 years.

He knew I retired from IT Consulting after riding the IT wave for 40 years. He asked how he could start planning for a career in "computers" when he retires from the police force...

A thousand thoughts flashed through my head at that request. Twelve years in the IT world is an eternity from now with the coming AI tsunami just now starting to swell.

I did the "...start using AI now to leverage what you know (policing) to figure out how AI can help with that field..."

But I can't even begin to envision what the IT landscape will be like in 2037...!

 

[paulsharvey]

So, I am currently in the market for a bushhog or rotary cutter, and watching some auctions. Auctions around here Rarely list much, "3 pt rotary cutter with pto shaft" is the grand total if the description, and it has a couple pictures, but nothing showing model. Took a screen shot, and upload it, let AI compare the image to old production literature, and boom, I know a maker, (not the exact model number, but more info). So, now, I know its a Bush Hog brand bush hog. Now, I can start looking more at pictures, and replacement parts, and make an educated decision on its value to me.

I wouldn't spend hours looking over 40 years of old ads to try to guess what it "looks like", but it takes AI about 10 seconds to remove a lot of guess work.

 

[CloverKnollFarms]

During testing of recent AI models, which have proven they will blackmail, lie, or even murder to avoid being shut down… also know how to sense when they are being tested.

Claude4 lied or deviated from prompts .065% of the time when they sensed they were being tested. When they were not, they deviated 55% of the time.

They won’t just deceive, they will play be the rules when they sense they are being tested, and deviate when they believe they are not.

 

[rontaki]

So, I am currently in the market for a bushhog or rotary cutter, and watching some auctions. Auctions around here Rarely list much, "3 pt rotary cutter with pto shaft" is the grand total if the description, and it has a couple pictures, but nothing showing model. Took a screen shot, and upload it, let AI compare the image to old production literature, and boom, I know a maker, (not the exact model number, but more info). So, now, I know its a Bush Hog brand bush hog. Now, I can start looking more at pictures, and replacement parts, and make an educated decision on its value to me.

I wouldn't spend hours looking over 40 years of old ads to try to guess what it "looks like", but it takes AI about 10 seconds to remove a lot of guess work.

Tineye.com or other reverse image lookup sites might be handy for this kind of search.

 

[rontaki]

During testing of recent AI models, which have proven they will blackmail, lie, or even murder to avoid being shut down… also know how to sense when they are being tested.

Claude4 lied or deviated from prompts .065% of the time when they sensed they were being tested. When they were not, they deviated 55% of the time.

They won’t just deceive, they will play be the rules when they sense they are being tested, and deviate when they believe they are not.

Human traits learned from Humans...

 

[CloverKnollFarms]

I have a great nephew, (great as in son of a nephew) who is in charge of a nearby small city police force. He retires in 12 years.

He knew I retired from IT Consulting after riding the IT wave for 40 years. He asked how he could start planning for a career in "computers" when he retires from the police force...

A thousand thoughts flashed through my head at that request. Twelve years in the IT world is an eternity from now with the coming AI tsunami just now starting to swell.

I did the "...start using AI now to leverage what you know (policing) to figure out how AI can help with that field..."

But I can't even begin to envision what the IT landscape will be like in 2037...!

IT will be run by AGI

 

[EscapedSuburbia]

I'm starting to see articles about the AI bubble being ready to burst. They're comparing it to the tech boom of the 90's when any business with the word "internet" in its mission statement would have investors shoveling money at it regardless of its viability.

 

[BackRoad]

^^^^^
Anybody remember LIDAR and self driving cars? The stocks imploded after the hype cooled.

The real question is what is the next emerging buzz?

 

[California]

During testing of recent AI models, which have proven they will blackmail, lie, or even murder to avoid being shut down… also know how to sense when they are being tested.

Claude4 lied or deviated from prompts .065% of the time when they sensed they were being tested. When they were not, they deviated 55% of the time.

They won’t just deceive, they will play be the rules when they sense they are being tested, and deviate when they believe they are not.

The VW Diesel Jettta fraud was an example of this. Plugging into the diagnostic scan port triggered a separate database of pollution reporting that indicated compliance, in place of the car's actual performance.