AI Can Write in English. Learn Other Languages Now
“What’s amazing about so many language models is what they know about how the world works from reading all the things they find,” he said. Chris Manning, a Stanford professor who specializes in AI and language.
But GPT and other factors have a talent for static parrots. They know how to recreate the patterns of words and grammar that appear in speech. That means they can blush at fools, wild incorrect facts, ug hate speech scraped from the darker corners of the web.
Amnon Shashua, a professor of computer science at the Hebrew University in Jerusalem, is the founder of yet another startup to build an AI model based on this approach. He knew a thing or two about AI marketing, selling his last company, Mobileye, which pioneered using AI to help cars see things on the road, in Intel in 2017 at $ 15.3 billion.
New company in Shashua, AI21 Labs, which came out secretly last week, has created an AI algorithm, called Jurassic-1, that demonstrates surprising English and Hebrew language skills.
In demos, Jurassic-1 can generate paragraphs of text on a given topic, dream up compelling titles for blog posts, write simple pieces of computer code, and more. Shashua said the model is more sophisticated than the GPT-3, and he believes future Jurassic versions could be able to make a kind of understanding of the common sense of the world from the information it gathers.
Other efforts to recreate GPT-3 are appearing in different languages around the world-and on the internet. In April, researchers at Huawei, China’s tech giant, published the details in a GPT-like language model called PanGu-alpha (written as PanGu-α). In May, Naver, a South Korean search giant, says it has created its own language model, called HyperCLOVA, that “speaks” Korean.
Jie Tang, a professor at Tsinghua University, led a team of Beijing Academy of Artificial Intelligence developing another model of Chinese language called Wudao (which means “enlightenment” ’) with help from government and industry.
Wudao’s model is more numerous than any other, meaning that its simulated neural network spreads across multiple cloud computers. Increasing the size of the neural network is key to making GPT-2 and -3 more capable. Wudao could also work with both images and text, and Tang organized a company to commercialize it. “We believe it can be a cornerstone of all AI,” Tang said.
Such enthusiasm seems to stem from the capabilities of these new AI programs, but the race to commercialize such speech models may also be quicker to move than efforts to add guardrails or limit. the misuse.
Perhaps the most urgent concern about AI language models is how they can be misused. Because models can alter compelling text on a topic, some people worry that they can easily be used to create bogus reviews, spam, or fake news.
“I wonder if disinformation operators won’t be less likely to invest serious energy experimenting with these models,” he said. Micah Musser, a Georgetown University research fellow studying the potential for language models to spread misinformation.
Musser said the research suggests that AI cannot be used to capture the disinformation created by AI. There may not be enough information in a tweet for a machine to judge if it was written on a machine.
More and more problems other than bias may be hiding within the larger language models as well. Research has shown that language models trained in the Chinese internet show censorship molded that thread. Programs also inevitably capture and copy subtle and obvious biases about race, gender, and age in the language they use, including gruesome statements and ideas.
Similarly, these large speech models may fail in surprising or unexpected ways, moreover. Percy Liang, another professor of computer science at Stanford and lead researcher at a new center dedicated to studying the potential of powerful, general purpose AI models such as GPT-3.