Elon Musk’s Xai explores how to make AI like Donald Trump
Researcher affiliated with Elon Musk startup xai We found new ways to measure and manipulate established preferences and values artificial intelligence Models – Includes their political views.
The work was guided and Hendrickdirector of a nonprofit organization AI Safety Center Xai’s advisor. He suggests that using this technique can better reflect voter will by popular AI models. “In the future, (models) may be consistent with a particular user,” Hendrycks told Wired. But in the meantime, he says, a good default is to use election results to guide the views of the AI model. He isn’t saying that the model should always be “trump all the time,” but he argues that it should be slightly biased towards Trump “because he won the popularity vote.” .
Published by Xai New AI Risk Framework On February 10, he said that GROK can be evaluated using Hendrycks’ utility engineering approach.
Hendrycks leads a team at the AI Safety Center for AI Safety, Berkeley, California, and the University of Pennsylvania, and uses technology borrowed from economics to measure the preferences of various products to create AI models. I analyzed it. By testing the model in a wide range of virtual scenarios, researchers were able to calculate what is known as utility functions. This is a measure of satisfaction that people come from good or service. This allowed us to measure preferences expressed in different AI models. Researchers often found that they were consistent rather than accidental, and showed that these preferences become more ingrained as the models become larger and more powerful.
Some Research We found that AI tools such as ChatGpt are biased towards views expressed by environmental environments, left-wing, and libertarian ideologies. In February 2024, Google faced criticism from Musk and others after the Gemini tool was predisposed to generating images branded as “critics.”woke up“Black Vikings and the Nazis, etc.
Techniques developed by Hendrycks and his collaborators provide a new way to determine how the perspective of an AI model differs from the user. Ultimately, some experts have hypothesized, and this type of divergence can be potentially dangerous for very intelligent and capable models. For example, studies show that certain models consistently assess the presence of AI above AI in certain nonhuman animals. Researchers also say they find that models seem to value some people more than others, raising their own ethical questions.
Some researchers, including Hendrycks, believe that current methods for adjusting models such as output manipulation and blocks are not sufficient when unnecessary targets lurk underneath the surface within the model itself. “We have to stand up to this,” says Hendrycks. “You can’t pretend you’re not there.”
Dylan Hadfield MennellA Hendrycks paper says that a professor at MIT, who studies how to match AI to human values, suggests a promising direction for AI research. “They find some interesting results,” he says. “The main thing that stands out is that the utility representation becomes more complete and consistent as the model scale increases.”