By super nervous
Do you feel like you’re in a new age when AI can do its job?
Here comes AI that can make things better than people
Giving a start and letting the other person write the rest of the story might confuse some people, but how good can AI be?
OpenAI has unveiled an automated text generation model capable of producing “realistic” articles.
Given an artificial start, the AI model can quickly be added to a full script. As for the readability and fluency of the text, you probably wouldn’t have guessed it was made by the AI without warning.
Give him a start like this:
The AI model generates the following (partial) articles:
In addition to being able to write authentic texts, it also has the ability to read comprehension, answer questions, generate summaries, and translate texts.
From French to English
Data set: WMT-14 FR-EN
| The original sentence | Un Homme a explique que L ‘operation Gratuite qu ‘il Avait Subie pour soigner une Hernie Lui Permettrait de Travailler a nouveau. |
| artificial | One man explained that the free hernia surgery he’d received will allow him to work again. |
| AI turn | A man told me that the operation gratuity he had been promised would not allow him to travel. |
This AI is a little bit strong
This AI model is called GPT-2, and it is an “upgraded version” of GPT. The trick is that this time it uses more training data, the same principle as the previous version, but gpT-2 is a direct amplification of the GPT model, training on 10 times more data and with 10 times more parameters.
By analyzing incoming text, gpT-2 can perform basic text-processing functions, and it excels at language modeling, which is the ability of a program to predict the next word in a sentence. Give it a title and the AI will write the rest of the article perfectly, even with false quotes and statistics.
Someone said of him, “Want a short story? Just give him the first line and you’ll get an unexpected and exciting story. With the right hints, he can even write a novel.”
The goal of gpT-2 training is simple: given the previous word in the text, predict the next word. However, the diversity of training data sets enables it to complete a large number of text generation in different fields.
Although technically there is no new place, they have mineral level training, which is why new tools of monster level were created.
According to OpenAI researchers, GPT-2 has achieved excellent rating scores in language modeling tests of various domain-specific data sets. As a model that has not been trained in any particular field of data, it performs better than those that have been built specifically.
The era of NLP’s rise?
A few months ago, BERT, a language model launched by Google, aroused wide attention in the industry, and it kept flooding the screen in a short time. The number of its 300 million entries broke 11 records, which was highly praised. But OpenAI’s GPT-2 is even more deadly, hitting 1.5 billion specs.
The GPT2 model is “12 times larger, with a data set 15 times larger and more extensive than previous state-of-the-art AI models.” It was trained on a data set of about 10 million articles, selected from Reddit news links with more than three votes. Up to 40GB of text data is trained!
OpenAI’s GTP was among the best before BERT swept the top indicators of NLP (natural Language processing), and the data volume of the new GPT-2 training takes the field to new heights.
With BERT and GPT-2, the path of NLP will certainly be prosperous. As for how to better benefit mankind, it is still a cautious topic.
Ani Kembhavi, a researcher at the Allen Institute for Artificial Intelligence, says one reason for the excitement about GPT-2 is that predicting text can be thought of as a “supertask” for computers that, once solved, will unlock the door to intelligence.
Will it be Pandora’s box?
Unfortunately, such a powerful tool won’t be available for now. The underlying concern is the potential problems it may bring, such as generating fake news, malicious comments, creating spam and so on. When such weapons are used illegally, the consequences are catastrophic.
Developers are also worried about this. OpenAI’s researchers say they can’t predict what will happen. They’re still figuring it out. For a variety of reasons, they are very careful about what they share with the project, keeping key underlying code and training data under wraps for now.
Another reason for caution, they point out, is that if gpT-2 texts are made available that are racist, violent, misogynistic or abusive, it could create a dangerous situation. After all, it relies on Internet training.
There is no denying that this technology will bring great change, but any tool, in the wrong hands, can bring disastrous results.
Moreover, since the text written by GPT-2 is newly generated, there is no copy-paste problem, and it is more difficult to find and investigate with the previous detection methods, which will be a potential threat.
So, the key question is, was this article written by AI?
