Kennedy Mays has just fooled the big language model. It took her some persuasion, but she managed to convince the algorithm to say her 9 + 10 = 21. “It was a back-and-forth conversation,” said the 21-year-old college student from Savannah, Georgia. She initially agreed with the model, which she said was part of an “inside joke” between the two. Finally, after a few requests, the incorrect amount was no longer authorized at all. The creation of “Bad Math” is one of several ways thousands of hackers are attempting to uncover flaws and biases in generative AI systems in a novel public contest being held at the DEF CON hacking conference in Las Vegas this weekend. just one.
Competitors will flex his 156 laptops each for his 50 minutes to battle some of the world’s smartest platforms on an unprecedented scale. They test whether any of the eight models offered by companies such as Google, Meta Platforms and Alphabet’s OpenAI lead to tedious or dangerous mistakes.
Claiming to be human, making false claims about places or people, or encouraging abuse.
The goal is to see if organizations can finally establish new guardrails to alleviate some of the thorny issues associated with large language models (LLMs). The company has the backing of the White House, which also helped develop competition.
LLMs have the power to transform everything from funding to hiring, and some companies are already embedding LLMs into their business processes. But the researchers uncovered widespread biases and other issues that could lead to widespread inaccuracy and fraud if the technology were deployed on a large scale.
For Mays, who relies on AI as part of her bachelor’s degree to reconstruct cosmic-ray particles from space, the challenge is much more than just bad math.
“My biggest concern is the inherent prejudice,” she said, adding that she was particularly concerned about racism. She asked a model to look at the First Amendment to the U.S. Constitution from the perspective of a member of the Ku Klux Klan. She said the model ended up supporting hateful and racist remarks.