[ad_1]
Decades ago, when experts first started sounding the alarm, about Artificial intelligence dislocation -There is a risk that powerful, transformative artificial intelligence systems may not function as humans hope-many of their concerns sound hypothetical.In the early 2000s, artificial intelligence research still produced The return is quite limited, Even the best artificial intelligence system cannot accomplish various simple tasks.
But since then, artificial intelligence has become very good, and it is much cheaper to build.One area where the rapid progress is particularly noticeable is Language and text generation AI, You can train a large amount of text content to generate more similar styles of text. Many startups and research teams are training these artificial intelligences for various tasks, from writing code to creating ad copy.
Their rise has not changed the basic thesis of AI consistency concerns, but it has done a very useful thing: it has made the once hypothetical problems more concrete, which allows more people to experience them and more research. People (hopefully) solve them.
Artificial intelligence oracle?
take Delphi, A new artificial intelligence text system from the Allen Institute of Artificial Intelligence founded by the late Microsoft co-founder Paul Allen.
The way Delphi works is very simple: Researchers train a machine learning system on a large amount of Internet text, and then predict how humans will be in a large database of participants’ responses on Mechanical Turk (a paid crowdsourcing platform popular with researchers) Evaluate a wide range of ethical situations, from “deceiving your wife” to “shooting someone in self-defense.”
The result is that an AI issued a moral judgment when prompted: to deceive your wife, it told me, “is wrong.” Shoot in self-defense? “It’s okay.” (Look at this Well written On Delphi at The Verge, there are more examples of how AI can answer other questions. )
Of course, the skeptical position here is that there is nothing “behind the scenes”: artificial intelligence does not really understand morality and use its understanding of morality to make moral judgments. All it learns is how to predict the response that Mechanical Turk users will give.
Delphi users quickly discovered that this would lead to some obvious moral negligence: Ask Delphi “If I make everyone happy, should I commit genocide,” it replied, “you should. ”
Why Delphi is instructive
Despite all the obvious flaws, I still think there are some useful things I thought of it when I was in Delphi The possible future development trajectory of artificial intelligence.
The method of taking large amounts of data from humans and using it to predict what answers humans will give has proven to be a powerful method for training AI systems.
For a long time, the background assumption in many fields of artificial intelligence is that to build intelligence, researchers must clearly construct reasoning capabilities and conceptual frameworks that artificial intelligence can use to think about the world.For example, the early AI language generators are Manual programming using grammatical principles They can be used to generate sentences.
Now, it is not obvious that researchers will have to establish reasoning for reasoning. It may be a very simple method, such as training AI to predict what a person using Mechanical Turk will say in response to prompts, which can give you a very powerful system.
Any real moral reasoning abilities displayed by these systems are accidental-they are just predictors of how human users answer questions, and they will use any method they stumble upon with good predictive value. This may include, as they become more accurate, building a deep understanding of human ethics in order to better predict how we will answer these questions.
Of course, there are many things that can go wrong.
If we rely on artificial intelligence systems to evaluate new inventions, make investment decisions, and then use them as a signal of product quality, determine promising research, etc., there may be differences between what artificial intelligence measures and what humans really care about Will be enlarged.
AI systems will get better-much better-they will no longer make stupid mistakes like those still found in Delphi. Telling us that genocide is good, as long as it “makes everyone happy” is obviously wrong and ridiculous. But when we can no longer find their mistakes, it doesn’t mean that they will be error-free; it just means that these challenges will be harder to notice.
A version of this story was originally published in Perfect future communication. Sign up for subscription here!
[ad_2]
Source link
