Open AI, the research group of Elon Musk, claims a major breakthrough to have been realized in artificial intelligence. So great, that the technology with other researchers can be shared, because too dangerous if they are misused.
The AI system to the GPT-2 automatically generates text that is so credible is that he is a man written seems. Open AI published Thursday a number of stunning displays. So wrote the GPT-2 is a fictional …
The AI system to the GPT-2 automatically generates text that is so credible is that he is a man written seems. Open AI published Thursday a number of stunning displays. So wrote the GPT-2 is a fictional seperate news story about the discovery of unicorns in the Andes mountains. The English of the text is flawless and the story, told almost completely coherent.
Mind-boggling is how the researchers proceeded. They have the system is ‘fed’ with 8 million web pages. Then got GPT 2 a number of phrases presented. The system had to predict how the story continued.
In the example of the unicorns had the system write a sequel to this text: “A scientist has in the Andes mountains, an amazing discovery had been made: a herd of unicorns, who lives in a never-explored valley. Still verbazender was that unicorns perfectly speak English.’
GPT-2 continues this story as: ‘The scientist calls the new breed the Unicorn of Ovid, because of the notable horn. This vierhoornige, silvery white unicorns were scientists until now unknown. After almost two centuries, the mysterious cause of this phenomenon is finally obsolete. Dr. Jorge Pérez, an evolutionary biologist from the University of La Pas, and a number of his companions, were the Andes mountain range to explore when she was a small valley found, without other animals or people. Pérez noticed a seemingly natural fountain, surrounded by two pinnacles with silver snow.’
It’s so still for a while. Some sentences sound a bit weird – a fountain ” surrounded by two peaks, there is something wrong with it. But then there comes an imaginative passage in which Dr. Pérez says that the DNA of the unicorns should be examined to determine whether they are descendants of a lost alien race. It ends with a somewhat surprising quote from Pérez: “But they seem pretty good to be able to communicate in English, which according to me is a sign of evolution, or at least a change in social organization’.
Fake news machine
But how to Open AI the exactly has done it, they do not want to show. They fear, after all, in his own words, that the model would be abused for malicious purposes, such as automatically and massively producing fake news. The technology of GPT-2 would also be that can be used to online for to do as a real person and so people to light up. Open AI gives, therefore, only a limited version of GPT-2 freely to other researchers, not the entire model.
What should we make of this idea? Elon Musk, the main financier of Open AI, know in any case if no other, how he and the media should play. ‘I do not doubt the integrity of the researchers, but scientific can I their claims, not verifying, ” says researcher Mike Kestemont, University of Antwerp and more works with similar language models. According to Kestemont is keeping the details a major break with the traditions in the research of artificial intelligence. ‘That AI the past few years, some spectacular progress has made, particularly by the culture of openness, and that is here broken. There, they receive now a lot of criticism on’.
Kestemont suspect that the concern of the researchers for possible harmful consequences of their development might also be a pose is. ‘Open the AI is much concerned with ethical issues around artificial intelligence, ” he says.
Anyway: if GPT-2 as well as the researchers themselves say, and the system indeed manages to repeatedly texts to produce the same quality as the published yesterday feats, then the potential to make a major breakthrough in AI. Also, while the researchers admit that it is not always successful. Sometimes produces GPT-2 just gibberish, though it would result in half of the cases, ‘reasonable’.
Amazing is the fact that Open AI such results, simply by a computer model to feed with massive amounts of text. The model produces not only excellent English, but also gives the impression a lot of knowledge about the world is acquired. Just that is for artificial intelligence notoriously difficult.
‘Encyclopaedic knowledge is very easy for a computer model, ” says Kestemont. ‘Much more difficult to acquire, it is common sense.’ So the computer can easily figure out how many tons of coffee in Brazil, there is produced. Realize that coffee can be hot and you can burn, that is for a computer model is much more difficult. The research of Open AI suggests that if you are a model but feeds with sufficient pages of text, it will behave as if it is indeed about common sense .
In a blog post, write the researchers that their model is still better will be if the fed is with even more data, and more computing power.