AI safety testing finds 2x more failures
A new testing method finds twice as many real-world AI failures by smartly combining cheap simulations with limited real-world tests.
A new testing method finds twice as many real-world AI failures by smartly combining cheap simulations with limited real-world tests.