[ad_1]
“It’s amazing that these great language models know how the world works, by reading all the things they can find,” he says. Chris Manning, A Stanford professor specializing in AI and language.
But GPT and the like are basically statistically talented parrots. They learn to recreate patterns of words and grammar found in language. This means that they can throw away nonsense, events of inaccurate accuracy, and hateful language taken from the dark corners of the web.
Amnon Shashua, A professor of computer science at the Hebrew University of Jerusalem, is the founder of another startup that builds an AI model based on this approach. He knows a thing or two about AI marketing, selling his latest company, Furniture, AI was a pioneer in helping cars detect things on the road Intel In 2017, $ 15.3 trillion.
New Shashua company, AI21 laboratories, which came out of hiding last week, has developed an IA algorithm called Jurassic-1, which demonstrates striking language skills in English and Hebrew.
In demos, Jurassic-1 can create text paragraphs on a specific topic, dream up attractive titles for blog posts, write weird computer codes, and more. Shashua said the model is more sophisticated than the GPT-3, and believes that future versions of the Jurassic can build a way to understand the world sensibly with the information it gathers.
Other efforts to recreate GPT-3 reflect the linguistic diversity of the world — and of the Internet. In April, researchers Huawei, China’s technology giant, published details Of a Chinese GPT-like model called PanGu-alpha (written under the name PanGu-α). In May, NaverThe South Korean search giant has said it has developed its own language model, called HyperCLOVA, which “speaks” Korean.
Jie Tang, A professor at Tsinghua University, leads a team Beijing Academy of Artificial Intelligence which developed another Chinese model called Wudao (“illustration”) with the support of the government and industry.
The Wudao model is much larger than any other, which means that its simulated neural network extends to more cloud computers. Increasing the size of the neural network was key to increasing GPT-2 and -3 capacity. Wudao can also work with images and texts, and Tang has set up a marketing company. “We think that can be the basis of all AI,” Tang says.
This excitement seems to be guaranteed by the capabilities of these new AI programs, but the race to commercialize these language models can also move faster than efforts to add railings or limit misuse.
Perhaps the most serious concern about AI language models is how they can be misused. Because models can create compelling text on a topic, some people worry that they can be easily used to create fake opinions, spam, or fake news.
“I would be surprised if misinformation operators don’t at least invest the serious energy they experiment with with these models,” he says. Micah Musser, a research analyst at Georgetown University who studied the potential of language models to spread misinformation.
Musser says the study suggests that AI cannot be used to capture the misinformation generated by AI. It’s hard to find enough information in a tweet to judge whether a machine has been written by a machine.
It is likely that even more problematic types of bias will lurk within these huge language models. Research has shown that language models are trained in Chinese content on the Internet will reflect censorship which shaped that content. Programs inevitably capture and reproduce subtle and overt biases about race, gender, and age in the language they consume, including hate speech and ideas.
Also, these large language models can surprisingly or unexpectedly fail, he adds Percy Liang, Stanford computer science professor and principal investigator new center Dedicated to exploring the potential of powerful general purpose AI models like GPT-3.
[ad_2]
Source link
