[ad_1]
If you don’t worry enough already, consider the world in which AI exists hackers.
Hacking is as old as humanity. We are creative problem solvers. We exploit escapes, manipulate systems, and strive for more influence, power, and wealth. So far hacking has only been a human activity. Not for long.
As I divided it into one the report I just published, artificial intelligence will find vulnerabilities in all kinds of social, economic and political systems, and then exploit them at unprecedented speeds, scales and spheres. Once humanity is hacked, AI systems will hack other AI systems, and humans will only suffer side damage.
Okay, maybe it’s hyperbole, but it doesn’t require future science fiction technology. I do not postulate AI “Uniqueness,” where the AI-learning feedback loop becomes so fast that it transcends human comprehension. I’m not going to assume smart androids. I’m not taking it for granted. Most of these hacks do not require much progress in AI research. They are already happening. As AI becomes more sophisticated, however, we often don’t even know what’s going on.
AI does not solve problems like humans. They explore more types of solutions than we do. They will go through complex paths that we have not considered. This can be a problem for something called an explanatory problem. Modern AI systems are basically black boxes. The data goes at one end, and the answer comes out at the other. It can be impossible to understand how the system came to be, even if you are a programmer who looks at the code.
In 2015, a research team fed an AI system called Deep Patient Health and Medical Data for about 700,000 people and tested whether it could predict disease. It might, but Deep Patient doesn’t provide an explanation for the basis of the diagnosis, and the researchers don’t know how its conclusions are drawn. A doctor may or may not trust your computer, but that trust will go blind.
While researchers are working on an AI that can explain themselves, there seems to be a link between ability and explanation. The explanation is a cognitive short film used by humans, adapted to the way humans make decisions. Forcing an AI to create explanations could be an additional limitation that can affect the quality of its decisions. For now, AI is becoming more and more opaque and less explanatory.
One by one, IA can get involved in something called reward hacking. Just as AI doesn’t solve people’s problems, humans will always come up with solutions we never anticipated, and some will overturn the intent of the system. In fact, AI does not think in terms of the implications, context, rules, and values that we humans share and take for granted. This reward hacking involves achieving a goal, but in a way AI designers didn’t want to and didn’t want to.
Perform a football simulation, where an AI guessed that if the ball was thrown out of bounds, the goalkeeper would have to throw the ball and leave the goal undefended. Or another simulation, where an AI invented that instead of running, it could be high enough to fall above a distant destination. Or the robot vacuum cleaner, instead of learning to encounter things, learned to drive backwards, where there was no sensor that said it was encountering things. If there are problems, inconsistencies, or gaps in the rules, and if these properties achieve an acceptable solution as specified in the rules, then AIs will find these hacks.
We learned this little hacking problem with the story of King Midas. When the god Dionysius gives him a wish, he asks Midas to turn everything he touches into gold. She will end up hungry and unhappy when her food, drink and daughter turn to gold. The problem is specification: Midas programmed the wrong target in the system.
Geniuses are very specific about the formulation of desires and can be maliciously pedantic. We know that, but there’s still no way to shake the genius. Whatever you want, you can always give as much as you would like. It will hack your will. Goals and desires are not always defined in human language and thought. We never describe all options or include all warnings, exceptions and conditions that apply. Any goal we set will necessarily be incomplete.
[ad_2]
Source link
