https://mp.weixin.qq.com/s/ZunCFquj73XhMBIo97gesg
By 超神经
我们就快到了「宁愿相信世上有鬼，也不相信 AI 的破嘴」的时代，人工智能又在 NLP 领域进化到了新的巅峰。AI that is better than others is here
Give a beginning and let the other person write the following story, which may stump some people. Then if it is given to AI, how good can they be?
Today, OpenAI announced an automatic text generation model that can write "realistic" articles.
Given the beginning artificially, this AI model can be quickly supplemented into a complete manuscript. As for the readability and fluency of the text, if you don't tell in advance, you may not be able to guess that this is the work of AI.
For example, give him the beginning: Scientists have made a shocking discovery that a group of unicorns live in a remote and undeveloped valley in the Andes. Even more surprising is that these unicorns speak perfect English.
The articles generated by this AI model are as follows (partial):
These creatures have unique horns, and scientists named them Ovid's Unicorn. The silver-white creature with four horns was not known to the scientific community before.
……
Although the origin of these creatures is not yet clear, some people believe that they were born from the intersection of a person and a unicorn, when human civilization did not exist. Professor Pérez said: "In South America, this phenomenon is very common."
...
If you want to confirm that they are descendants of the disappearing race, DNA testing may be the only way.
In addition to being able to write fake manuscripts, it also has the ability to read comprehension, question and answer, generate abstracts, and translate texts.
Translation: From French to English
Data set: WMT-14 Fr-En
Original sentence
One man explained that the free operation he had to treat a hernia would allow him to work again.
Artificial
One man explained that the free hernia surgery he’d received will allow him to work again.
AI translation
A man told me that the operation gratuity he had been promised would not allow him to travel.
AI 模型翻译实例
This AI is a bit strong
This AI model is called GPT-2, and it is an "upgraded version" of GPT. The cruel thing about it is that it uses more training data this time, which is the same as the principle of the previous version, but GPT-2 is a direct amplification of the GPT model. It trains on more than 10 times the amount of data. The amount of parameters is also 10 times more.
By analyzing the input text, GPT-2 can perform basic text processing functions. It is good at language modeling tasks. The task is to let the program predict the ability to give the next word in the sentence. Just give it a title, and AI can write the rest of the article perfectly, even with fake quotes and statistics.
Some people say it like this, "Want a short story? Just give it the first line, and you can get an unexpected and wonderful story. If you have the right hints, it can even write a long story."
The goal of training GPT-2 is simple: given the previous words in the text, to predict the next words and sentences. The diversity of training data sets makes it possible to generate a large number of texts in different fields.
Although there is no new place in technology, people have mineral-level training, which is why new monster-level tools have been created.
OpenAI researchers said that GPT-2 has achieved excellent scores in language modeling tests on various domain-specific data sets. As a model that has not been specifically trained in any field data, its performance is better than those specially built models.
The era of the rise of NLP?
The language model BERT launched by Google a few months ago has aroused widespread attention in the industry. It has been constantly refreshing the screen in a period of time, and its 300 million parameter volume refreshed 11 records. It is full of praise. But GPT-2 launched by OpenAI this time is even more terrible. It has reached 1.5 billion parameters.
Compared with the most advanced artificial intelligence models before, the GPT2 model is “12 times larger, the data set is 15 times larger, and the scope is wider.” It was trained on a data set of approximately 10 million articles, selected through news links with more than 3 votes on Reddit. The trained text data is up to 40GB!
Before the BERT bloodbath NLP (Natural Language Processing) top indicators, OpenAI's GTP stood among the first-class masters, and the new GPT-2 also directly took this field to new heights through massive training data. .
With BERT and GPT-2, the road of NLP will definitely be booming. As for how to better benefit mankind, this is still a prudent topic.
Ani Kembhavi, a researcher at the Allen Institute for Artificial Intelligence, said that one of the reasons for being excited about GPT-2 is that predictive text can be considered a computer's "super task". Once this challenge is solved, the door to wisdom will be opened.
Will it be Pandora's Box?
Unfortunately, such a powerful tool cannot be announced yet. The consideration behind it is the hidden dangers it may bring, such as generating fake news, malicious comments, spam and so on. Such weapons are used in illegal ways, and the consequences are also catastrophic.
For this aspect, developers are also worried. The OpenAI researchers said that they cannot predict what it will bring. They are still exploring. For various reasons, they are very cautious about the content shared by the project, and currently do not disclose the main basic code and training data.
They pointed out that another reason for caution is that if someone provides GPT-2 texts about racism, violence, misogyny or abusiveness, it will create a very dangerous situation. After all, it relies on Internet training.
There is no denying that this technology will bring about tremendous changes, but any tool in the hands of bad actors will bring catastrophic consequences.
Moreover, since the text written by GPT-2 is newly generated, there is no copy and paste problem, and it is more difficult to find and troubleshoot with previous detection methods, which will be a potential threat.
So, here comes the key question, is this article written by AI?
