THE AMAZON'S INTERNAL EXPOSURE COSTS ARE OUT OF CONTROL

On 30 July, the Financial Times cited sources who claimed that there had been significant overspending on AI projects as reported internally in the Amazon. The most exaggerating one is the matching of author information with Claude Sonnet on the commodity page. The project eventually failed, but spent $1.8 million, 86 per cent more than the budget, and was discovered five months later. Another financial audit tool project had an unexpected additional cost of approximately $541 million. A project to optimize the speed of logistics distribution also cost $134,000, which was discovered more than two weeks later. Amazon engineers attributed the problem to deployment errors and lack of cost constraints. A small error in the traditional system, with little to pay, could turn into a "dangerous and expensive" bill in the AI project. The company is developing automation measures to prevent similar projects from continuing out of control. According to the Amazon response, these cases involve only a small number of teams and do not represent the overall day-to-day use of AI by companies. Prior to this, the Amazon also closed the internal AI use list because of the additional consumption of Token by its employees in order to get Agent to perform meaningless tasks。