[ad_1]
That’s what Gebru, Mitchell, and five other scientists in their paper warned about, among other things, as LLMs call them “stochastic flowers”. “Language technology can be very, very useful when it’s properly positioned and positioned,” says Emily Bender, a professor of linguistics at the University of Washington and one of the authors of the paper. But the overall goals of LLMs — and the credibility of their mimicry — encourage companies to use them in areas that are not necessarily equipped.
In a recent speech at one of the largest AI conferences, Gebru linked this urgent expansion of the LLM to the consequences of his life. Gebru was born and raised in Ethiopia war intensifies it has destroyed the northern Tigray region. Ethiopia is a country where 86 languages are spoken, almost all of which are not found in ordinary technological languages.
When the Tigray war first began in November, Gebru saw the platform move in a wave of disinformation. This is indicative of the ongoing model of content-observed by researchers, an area in which Facebook is based on LLMs. Communities that speak languages Silicon Valley does not prioritize hostile digital environments.
Gebru has stated that the damage does not end there either. When fake news, hate speech, and even death threats are not moderated, they come out as training data to create next-generation LLMs. And the models, by pushing back those who are being trained, are undermining these toxic linguistic models on the internet.
In many cases, researchers have not researched well enough to find out how this toxicity can appear in declining applications. But there are some scholarships. In the 2018 book Oppression algorithms, Safiya Noble, an associate professor of information and African American studies at the University of California, Los Angeles, has documented how biases embedded in a Google search perpetuate racism and, in extreme cases, also motivate racial violence.
“The consequences are pretty serious and significant,” he says. Google is not only the main portal for the knowledge of the average citizen. It also provides information, infrastructure to institutions, universities, and state and federal governments.
Google already uses an LLM to optimize some of its search results. With the latest LaMDA announcement and a recent proposal Published in a preprint paper, the company has made it clear that it will increase confidence in the technology. The problems that this noble concern revealed could be even more serious: “They released Google’s AI team for raising very important questions about racist and sexist models of discrimination embedded in large language models.”
BigScience
The BigScience project was given a direct response to the growing need for scientific study in LLM. With the rapid proliferation of technology and Google’s censorship of Gebru and Mitchell, Wolf and several colleagues realized that it was time for the research community to take matters into their own hands.
Inspired by open scientific collaborations such as CERN in particle physics, they came up with an idea for an open source LLM that could be used for critical research that is independent of any company. In April this year, the group received a grant to build using a French government supercomputer.
In technology companies, LLMs are only performed by half a dozen people with mostly technical specialization. BigScience wanted to bring in hundreds of researchers from across countries and disciplines to participate in the process of building truly collaborative models. The wolf, a Frenchman, first approached the French PNL community. From there, the initiative became a worldwide operation that brought together more than 500 people.
The collaboration is now freely organized in a dozen working groups and they are counting, each addressing different aspects of model development and research. One team will measure the environmental impact of the model by considering, among other things, the carbon footprint of LLM training and performance and the costs of the supercomputer life cycle. Another will focus on developing responsible ways to obtain data – looking for alternatives to simply extracting data from the network, such as transcribing historical radio archives or podcasts. The goal here is to prevent toxic language and non-consensual collection of private information.
[ad_2]
Source link
