{"text":[[{"start":8.72,"text":"Amazon staff have identified cases of what they called “catastrophically expensive” cost overruns caused by mistakes in how the company has deployed AI and a lack of spending controls."}],[{"start":20.38,"text":"Senior engineers told colleagues in a staff meeting on Tuesday that attempts to switch tasks from conventional programming to using AI models had led to “unplanned” spending, said multiple people familiar with the matter."}],[{"start":32.88,"text":"Engineers were working to create “automated” guardrails for future projects, among other measures to keep a lid on costs, the people said."}],[{"start":39.72,"text":"The episode highlights the difficulty that even the biggest tech companies are having integrating AI into their daily operations without overspending."}],[{"start":47.6,"text":"“It’s difficult to figure out how much anything [AI related] costs,” one senior Amazon employee told the FT."}],[{"start":54.52,"text":"During a presentation this week, employees were told about one instance in which the company spent $1.8mn on matching author details with listings on the group’s ecommerce website using Anthropic’s Claude Sonnet AI model despite the deployment failing, the people said."}],[{"start":70.72,"text":"The spending represented an 860 per cent cost overrun compared to the project’s budget and took five months to detect, they added."}],[{"start":78.84,"text":"Amazon said: “As with any new technology, we’re experimenting, learning and improving how we use it, including how we drive cost efficiencies.”"}],[{"start":87.68,"text":"Staff were told coding mistakes that were “trivially cheap” in traditional systems were proving “catastrophically expensive” as teams started to deploy AI models to complete tasks."}],[{"start":97.86,"text":"Senior engineers told staff that the overspending was not an “isolated” incident. In a separate episode, the company incurred roughly $541,000 in unexpected costs related to building financial auditing tools."}],[{"start":110.56,"text":"A third incident, in which AI was being used to improve delivery speeds across its logistics network, incurred $134,000 in accidental spending and took more than a fortnight to detect."}],[{"start":122.72,"text":"The company added: “Cherry-picking small, isolated examples where teams are learning from one another and portraying them as business as usual doesn’t reflect how teams across Amazon are using AI.”"}],[{"start":132.96,"text":"The presentation noted these examples spanned only a handful of teams at Amazon, which employs roughly 300,000 corporate staff and makes about $180bn in revenue each quarter."}],[{"start":144.32,"text":"The $2.5tn group is undertaking sweeping lay-offs in an attempt to reduce its costs and help finance its vast AI investment. Amazon this year is expected to spend $200bn in capital expenditure, the vast majority of which will go towards AI and data centre infrastructure."}],[{"start":163.12,"text":"Companies have been grappling with an increase in the cost of deploying AI tools as providers shift their pricing models. AI labs including Anthropic and OpenAI have transitioned some services from flat subscriptions to token-based billing, which tracks the units of data processed by models."}],[{"start":180.76,"text":"Changes in AI pricing models have led companies to shift their AI usage away from the most advanced models to cheaper mid-tier versions and alternatives, including open-weight models that only require users to pay for the computing power needed to run them."}],[{"start":195.32,"text":"Amazon earlier this year shuttered an internal leaderboard that ranked employees’ use of its Kiro developer platform after it led to so-called “tokenmaxxing”, with employees inflating their consumption of AI tokens."}],[{"start":207.64,"text":"Executives at the group told employees the service had been created with “good intentions” but ultimately ran up costs, the FT previously reported."}],[{"start":220.64,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1785413726_9510.mp3"}