Elon Musk’s Khai’s Adviser, EU resembling Donald Trump has a way to make more

An researcher related to the beginning of Elon Musk Xai found a new way to manage both measurement and declared advantages and values ​​expressed artificial intelligence Models – including political meetings.

Worked And HendrycksNon-commercial director AI Security Center and Xai’s advisor. It suggests that the technique can be used to make popular AI models that the electorate is better reflected. “Maybe in the future (a model) can be adapted to a specific user,” said Hendrycks string. But meanwhile, the election results can be used to manage the views of AI models. A model does not necessarily need to be “trum every side” but claims that it should be biased against Trump for a very popular voice.

It was given to xai The new AI risk base On February 10, HendryCKS can be used to assess the GOOD GROK to assess the Utility Engineering approach.

Hendrycks headed a team from a team from AI security, UC Berkeley and Pennsylvania and analyzed AI models using a technique that consumers use the economy to measure the advantages of different goods. Researchers with test models between a large hypothetical scenarios were able to calculate the recognition of a utility function, a size of the satisfaction that people caused by good or service. This allowed to measure the advantages expressed by different AI models. Researchers often identified that they are more consistent than haphazard, and these advantages showed more of these advantages to be larger and stronger than models.

Somewhat Research works As ChatGpt, AI tools revealed the environment, left, left and libertarian ideologies were biased. In February 2024, Google encountered criticism by musk and others after preparing the images of the Gemini vehicle to create branded images of the Gemini vehicle “wake up“As black vikings and Nazis.

The technique, prepared by Hendrycks and its employees, offers a new way to determine how the prospects of the AI ​​models can be different. Finally, some experts are assuming, such disagreements can be potentially dangerous for very smart and skillful models. Researchers show that they estimate the existence of AI above non-human animals that certain models do not consistently determinate certain models. Researchers said that the models preferred some people from others and raised their ethical questions.

Some researchers believe that the current methods for alignment of align models to align align models, manipulate models and inhibit the models, the models are sufficient to align the models. “We will face it,” says Hendrycks says. “You can’t pretend not there.”

Dylan Hadfield-MenellA professor in the MIT investigating the methods of aligning EU with human values, Indicates that HendryCKS is a promising direction for the AI ​​investigation. “They find some interesting results,” he says. “The main thing is the main, as the model scale increases, the utility is more complete and adaptation.”