Frontier AI models fail one in three attempts

A Stanford HAI report reveals that despite significant advancements, frontier AI models still fail approximately one-third of the time in real-world production tasks. This performance gap, termed the 'jagged frontier,' highlights a critical reliability challenge for enterprises. While AI excels in complex areas like advanced mathematics and cybersecurity, it struggles with basic perception tasks such as telling time. Furthermore, issues like hallucination rates remain high, and transparency from leading AI labs is declining, making independent auditing and benchmarking increasingly difficult.