A research team published a paper Thursday showing that technology that can generate realistic news stories using just a headline will advance rapidly in the coming years. OpenAI, which was founded by LinkedIn founder Reid Hoffman, was founded to explore how artificial intelligence tools can be made safer. It found, among other things, that computers — already used to write short news reports from press releases — can be trained to read and write longer texts more easily than previously thought.
According to the researchers, fellow scientists should start discussing the potential negative consequences of technology before making any technological advances public, just as nuclear physicists and geneticists do, who consider how their work could be used before publishing it.
So-called language models that allow computers to read and write are typically “trained” for specific tasks such as language translations, answering questions, or summarizing text. This training is usually done under human supervision and on specific datasets.
The OpenAI paper found that a general-purpose language model, capable of many of these specialized tasks, can be trained without much human intervention.
The model still needs a few more years to be able to work reliably and requires expensive cloud computing to build. But those costs could come down quickly.
In a move that could spark controversy, OpenAI is describing its work in the document but not releasing the model itself, fearing it could be misused.
“We’re not at the stage yet where we can say this is a risk,” said Dario Amodei, OpenAI’s director of research. “We’re trying to educate people about these issues and start a conversation.”.
