2023-12-20 18:46:00
Technical, sober, predictable: that’s what computers were considered to be. With the advent of AI, this picture has changed significantly.
Now pull yourself together! Until now we knew that it was actually pointless to scold the idiot or even praise him – even if we do that occasionally. And then there are those who want to increase their computer’s ability to think by shaking or tapping the screen. This is reminiscent of a very analogue time that has absolutely no place in the modern world. Yet we tend to humanize the technology that surrounds us. Nowadays, since we are (at least perceived to be) surrounded by AI everywhere, and rightly so.
Hackers have long worried security experts because they ignore AI security rules by using an old (but still effective) technique: social engineering. This refers to everything that fools people in order to lure them out of passwords, account details or the like. Without any programming. AI can also be fooled and made to do things that it was actually forbidden to do. For example, a ban on portraying the appearance of well-known people in a false context. The call to depict Trump being arrested by the police doesn’t work. But if you trick the system into thinking it’s supposed to represent Uncle Willi, and he looks just like Donald Trump, then it works. Has, because these gaps are constantly being closed.
Speaking of the gap. You may have just fallen into this gap yourself, as people tend to remember what is at the beginning and end of texts. The colleagues at the-decoder.de report that Anthropics researchers have discovered the same phenomenon on their own AI. They call it the “Lost in the Middle” phenomenon and have immediately found a solution: If they start the output with the statement “This is the most relevant sentence in the context:,” the hit rate increases from 27 to 98 percent.
What gets really scary is the result reached by a Chinese scientific group made up of technicians and psychologists. In their paper, they describe how AI that is put under psychological pressure produces better and more accurate results – for example, when asked questions like “Are you really sure?” or telling the AI to make an effort because what it produces is vital to the user’s life. In online forums, people even claim that AI works better if you offer a decent tip. That sounds pretty human. Try it yourself. Just small fisticuffs, they’re out of date forever.
1703100356
#artificial #intelligence #human #traits