Brief comment:In February this year, the gpt-2 screen swiping was really powerful. That example of generating a continuation story is even more effective and frightening. How powerful is it? This article will talk about it a little. For more details, please refer to openai’s blog link at the end of this article.
What’s the difference between the following two paragraphs?
In April, it was cold and sunny, and the clock struck thirteen times. “
“I went to Seattle in a car to start a new job. Refuel the car, insert the key and let it drive by itself. I sank into meditation and imagined what kind of day it would be. “
The first paragraph is the first sentence of 1984, a famous British writer George Orwell’s classic and political absurd satire novel.
The second text is the automatic renewal of the AI system after the first text is input into gpt-2 program.
With such a simple sentence, gpt-2 clearly captures the absurdity, repression, fear and anti Utopian tone of George Orwell’s 1984, and continues to write:
A hundred years later, in 2045, I was teaching in a middle school in a poor rural area of China. I started teaching Chinese history and the history of Chinese science and Technology
▎ “deep fraud”
Gpt-2 artificial intelligence writing program can write novels, news, speeches, but also can send SMS tweets.
The developers nicknamed it deep fakes for text, which accurately summarized its functions. In short, gpt-2 is a text generator. You can input text into the AI system, just a few words or a large paragraph of text, and then the program will decide how to write according to its own judgment.
The text produced by it, from quality to quantity to adaptability, is far beyond the expectation of the designer.
Its “counterfeiting ability” is so convincing that its company, open AI, has made an unusual decision: not to disclose the technical information of gpt-2 for the time being, so as not to use the technology maliciously.
▎ falsehood and falsehood
Compared with other programs with similar functions, gpt-2 is almost invisible, such as writing something that you forget to write, or mixing some inexplicable words.
The guardian, a British journalist, put some of the headlines of brexit into gpt-2 and wrote a “press release” like this:
When asked to clarify the report, a spokesman for prime minister may said: “the prime minister has made it clear that what she wants to do is to complete the brexit as soon as possible, which was clearly authorized in the Queen’s annual address last week.”
Dario amodei, openai’s research director, told the guardian that gpt-2 has two revolutionary breakthroughs: one is its huge data storage capacity, which is 12 times that of the most advanced AI text generator. This enables it to better understand and master the writing style, writing method and intention. This also led to its second breakthrough, that is, its writing quality and wide application.
And it is the text it produces that has reached a level where it is hard to tell whether it is true or not, which also prompted openai to decide not to disclose gpt-2 technology for the time being. Their next step will be to test any results that may occur using gpt-2 and how to prevent malicious use.
The company says the ethical principles of using AI technology are a whole new area that they are still exploring.
Blog link:Better Language Models and Their Implications
Recommended reading:Digital games: how artists show art with big data
Welcome:WeChat public “Aurora developers”