The Financial Times recently obtained the transcripts of the conversation, to which most of the decision-makers in the engineering teams working with the code were summoned. The main topic was to be the rapidly growing number of more and more serious incidents affecting the operation of Amazon Web Servicesi.e. cloud and hosting services on which a significant part of the Internet is based. Numerous failures that prevent the normal functioning of websites and applications have recently become more and more noticeable and are directly related to the problem that Business Insider wrote about a few weeks ago.
This is, of course, a phenomenon called vibe coding, in which writing the code on which a given software is based is handed over to generative AI. The code created in a few moments should, in theory, be functional, but the technology, despite its potential, is still very imperfect. And the effect of artificial intelligence can be far from satisfactory, even if at first glance everything seems fine. Buried bugs and references to non-existent repositories may not show up at firstbut subsequent layers of code with certain deficiencies begin to take their toll over time. And they lead to failures, which have become very problematic from Amazon's perspective.
For some time now, encouraging or even forcing employees to use AI tools in their daily work has not brought the expected results – because although certain tasks are actually performed faster, their quality suffers. In the notes obtained by the Financial Times, terms such as “high-impact incidents” were used, and in an e-mail to employees, Dave Treadwell, the senior vice president of the e-commerce department, wrote directly: – As you probably know, the availability of the website and related infrastructure has not looked very good recently.
The meeting was called shortly after the Amazon store went down for several hours, preventing people around the world from shopping. Reason? When asked to automatically correct the code, the AI tool simply deleted some of it and rewrite it. It took Amazon several hours to fix it all. This failure was not the only one recently, and although internally no one wants to admit the mistake, in order to counteract them in the future, Amazon's management decided to change the current strategy regarding the use of AI. The previous one, due to the increasing number of errors in the code, simply did not work… although the proposed solutions may not bring the expected results.
Well, despite the terrible condition of the code, Amazon does not intend to reduce the scale of using generative artificial intelligence to create it. Instead, a rule was introduced – strangely enough only now – according to which lower and mid-level engineers, before releasing the generated code to production, they will first need to get approval from someone in a senior position. And probably not one, but several such people. To put it simply – intensive implementation of AI in everyday work has generated previously non-existent problems that people are supposed to fix.
Increased oversight of code quality through the involvement of a more experienced human factor may in theory be effective, but will burden engineers and programmers with secondary and tedious additional taskswho could be doing other things during this time. In practice, this may also mean that the multitude of responsibilities will result in the process of the so-called code review, i.e. checking the code, management staff will anyway delegate AI tools such as Claude Code Review or Codex from OpenAI.
AI improving code after AI is, of course, also not foolproof, and although it may reduce problems in the short term, it may lead to another disaster in the long run. Amazon, however, does not seem to care about this at this stage, directly admitting that internal goals have been set, requiring developers to use 80% AI for tasks, at least once a week. The giant's management apparently believes that the situation will eventually stabilize, AI tools will stop making errors, and this will mean additional cost reductions and efficiency increases.
The reality that in a few years, the software code on which entire company budgets or even national economies will be based is created and repaired by AI and only a handful of people will understand it – is not a very safe vision. But artificial intelligence implementation fever to everything and everywhere, unfortunately, does not allow the largest companies using new technologies to slow down in any way in this entire process.
This has already led to the first stage of serious problems. They were noticed, it was admitted that it was a huge problem, but it was decided to continue this path, but with a little more caution. When Another wave of problems and failures will inevitably follow — perhaps then Amazon and other companies that have recently become fascinated by the irresponsible implementation of artificial intelligence will come to different conclusions.
