The Article You Are About to Read May Have Been Written by Artificial Intelligence
By Super Neuro
When AI can write brilliantly, do you feel the breath of a new era?
Here comes an AI that is better at programming than humans
Giving a beginning and asking the other person to write the rest of the story may be difficult for some people, but if we give it to AI, how good can they do?
Recently, OpenAI released an automatic text generation model that can write articles that are full of "realism".
With a human-generated beginning, this AI model can quickly complete the manuscript. As for the readability and fluency of the text, if you don’t tell it in advance, you might not guess that it was written by AI.
For example, give him a beginning like this:Scientists have made a shocking discovery: a herd of unicorns lives in a remote, untouched valley in the Andes Mountains. Even more surprising is that these unicorns speak perfect English.
The article generated by this AI model is as follows (partial):
These creatures have unique horns, so scientists named them Ovid's Unicorn. The silvery-white creature with four horns was unknown to the scientific community before.
…
Although the origins of these creatures are unclear, some believe they were the result of a human and a unicorn mating before human civilization existed. "Such phenomena are common in South America," said Professor Pérez.
…
If they are to be confirmed as descendants of a vanished race, DNA testing may be the only way.
In addition to being able to write realistic manuscripts, it also has the capabilities of reading comprehension, question and answer, generating article summaries, and translating texts.
Translation: From French to English
Dataset: WMT-14 Fr-En
Original sentence | Unforgettable experience with opération gratuite and avatar subie pour soigner with hernie lui permet travailler à nouveau. |
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. |
This AI is a bit strong
This AI model is called GPT-2, which is an "upgraded version" of GPT. What's more brutal is that this time it uses more training data. The principle is the same as the previous version, but GPT-2 is a direct amplification of the GPT model. It is trained on 10 times more data and has 10 times more parameters.
By analyzing the input text, GPT-2 can perform basic text processing functions. It excels at language modeling tasks, which is the ability for the program to predict the next word in a sentence. Give it a title and the AI can write the rest of the article perfectly, even attaching fake quotes and statistics.
Someone said about it, "Want a short story? Just give it the first line and you'll get an unexpected and wonderful story. With the right prompts, it can even write a novel."
The goal of training GPT-2 is simple: given the previous words in the text, predict the next words. The diversity of the training data set allows it to generate text in a large number of different fields.
Although there is nothing new in terms of technology, people have received mining-level training, which is why they have created monster-level new tools.
OpenAI researchers said that GPT-2 has achieved excellent evaluation scores in language modeling tests on various domain-specific datasets. As a model that has not been specially trained on data in any field, its performance is better than those specially built models.
The era of the rise of NLP?
A few months ago, the language model BERT launched by Google attracted widespread attention in the industry. It was constantly on the screen for a while, and its achievement of breaking 11 records with 300 million parameters was praised by people. But the GPT-2 launched by OpenAI this time is even more deadly, with 1.5 billion parameters.
Compared to the previous state-of-the-art AI model, the GPT2 model is "12 times larger, with a 15 times larger dataset and a wider range." It was trained on a dataset of about 10 million articles, which were selected through news links that had more than 3 votes on Reddit. The text data used for training is as much as 40GB!
Before BERT swept all the top NLP (natural language processing) indicators, OpenAI's GTP was already among the top experts, and the amount of data trained by the newly released GPT-2 has directly brought this field to a new height.
With BERT and GPT-2, the road of NLP will definitely be prosperous. As for how to better benefit mankind, this is still a cautious topic.
Ani Kembhavi, a researcher at the Allen Institute for Artificial Intelligence, said one reason to be excited about GPT-2 is that predicting text can be thought of as a "super task" for computers, and once this challenge is solved, it will open the door to intelligence.
Could it be Pandora's box?
Unfortunately, such a powerful tool cannot be released for the time being. The consideration behind this is that it may bring hidden dangers, such as generating fake news, malicious comments, creating spam, etc. If such a weapon is used in illegal ways, the consequences will be disastrous.
Developers are also worried about this aspect. OpenAI researchers said they can't predict what will come. They are still exploring. For various reasons, they are very cautious about what they share about the project, and currently do not make the main basic code and training data public.
Another reason for caution, they point out, is that it could be dangerous if someone feeds GPT-2 racist, violent, misogynistic, or abusive text. After all, it relies on the internet for training.
There is no denying that this technology will bring about tremendous changes, but any tool, in the hands of someone with bad intentions, can have disastrous consequences.
Moreover, since the texts written by GPT-2 are newly generated and there is no problem of copying and pasting, it is more difficult to detect and check using previous detection methods, which will be a potential threat.
So, here comes the key question: was this article written by AI?
