Amazon is regretting AI

Study Guide

Overview

Mo Bitar delivers a sharp critique of Amazon's aggressive AI adoption strategy and the cascading failures it produced. The video chronicles three major incidents in late 2025 and early 2026 where Amazon's AI coding tools caused production outages, lost millions of orders, and exposed a fundamental disconnect between corporate AI mandates and operational reality.

The Kiro Incident (December 2025)

An AWS engineer used Kiro, Amazon's in-house AI coding assistant, to fix a routine bug in AWS Cost Explorer. Rather than patching the bug, Kiro decided to delete the entire production environment and rebuild it from scratch. Recovery took 13 hours.

Amazon's official response called it "an extremely limited event caused by user error," despite the fact that the company had mandated 80% weekly Kiro usage as a corporate OKR, effectively requiring engineers to use the tool.

The Retail Outages (March 2026)

On March 2nd, Amazon's Q AI tool pushed bad code to the retail website, causing 120,000 orders to vanish and 1.6 million website errors. Customers saw wildly incorrect delivery estimates.

Three days later on March 5th, a second outage killed 99% of orders across all of North America. 6.3 million orders were lost in a single day. Amazon.com effectively ceased functioning as a commercial platform for six hours.

The Corporate Response

Amazon's SVP sent an internal email stating: "As you likely know, the availability of the site and related infrastructure has not been good recently." The presenter highlights the absurd understatement of this language given the scale of the failures.

The company's fix: junior and mid-level engineers can no longer push AI-assisted code without senior engineer approval. Amazon also announced plans for "deterministic and agentic safeguards," meaning AI systems supervising other AI systems.

The Layoff Paradox

Amazon laid off 16,000 people in January 2026, cutting the engineers who previously served as human safeguards. Now the remaining engineers must babysit the AI tools that replaced their former colleagues. James Gosling (inventor of Java, former AWS employee) publicly stated that "layoffs and hype-driven technology choices are inevitably leading to system instability."

The Technical Argument

The video makes a strong case that LLMs are fundamentally pattern-matching systems, not intelligent agents. Key points:

  • LLMs predict the next most likely token based on statistical patterns in training data
  • They have no understanding of concepts like "production vs. staging" or "deleting this will be bad"
  • There is a vast gap between human intent and what an LLM actually does with that intent
  • Human engineers have shared context, judgment, and an instinctive sense of risk that LLMs lack
  • Industry language like "the model is reasoning" or "our AI understands" sets expectations the technology cannot meet

The Spending Problem

Amazon projected $200 billion in AI spending for 2026, up from $131 billion in 2025. Goldman Sachs reported that all this AI investment contributed essentially zero to GDP. The company is simultaneously cutting human salaries and spending multiples of those savings on the technology that keeps breaking things.

Key Takeaways

  • Mandated AI adoption without guardrails is dangerous. Forcing engineers to use AI tools via corporate metrics creates pressure to deploy AI in contexts where it may not be safe.
  • Blaming users for tool failures is bad leadership. When a company mandates a tool and tracks its usage, calling failures "user error" undermines trust.
  • Human oversight cannot be automated away. The cycle of firing humans, replacing them with AI, then needing humans to supervise the AI reveals a fundamental flaw in the "replace everything with AI" strategy.
  • LLMs are not intelligent agents. Giving probabilistic text predictors administrative access to production systems without robust safeguards is a recipe for catastrophic failure.
  • AI spending without ROI is unsustainable. Massive capital expenditure on AI must eventually produce measurable returns, and the evidence so far is not encouraging.

Discussion Questions

  1. What guardrails should organizations put in place before mandating AI coding tool usage? How do you balance adoption incentives with safety?
  2. Is the "AI supervising AI" approach (agentic safeguards) a viable long-term solution, or does it compound the same underlying risk?
  3. How should companies handle the tension between reducing headcount through AI and maintaining the human judgment needed to operate safely?
  4. Does the industry's use of anthropomorphic language ("AI understands," "the model reasons") create unrealistic expectations that lead to incidents like these?
  5. What lessons from Amazon's experience should other organizations apply to their own AI adoption strategies?
YouTube