Claude code's descent: amd exec sounds alarm on ai code assistant

Anthropic’s Claude Code, once heralded as a potential disruptor to OpenAI's coding dominance, is facing a serious credibility crisis. Internal assessments from AMD, one of the world’s largest chip manufacturers, reveal a worrying decline in the AI’s accuracy and reliability, prompting a stark warning from AMD’s AI Director, Stella Laurenzo.

The root of the problem: a growing pattern of 'laziness'

Laurenzo, who first flagged concerns internally in February, has now publicly detailed months of analysis showing a sharp increase in errors and inefficiencies. What began as anecdotal reports from senior engineers quickly escalated into a statistically significant trend. A deep dive into 6,852 sessions of Claude Code revealed a concerning shift: a surge in “stop-hook” violations, interruptions in thought processes, and what Laurenzo bluntly describes as “laziness” within the model.

The numbers paint a grim picture. Before March 8th, such interruptions were virtually nonexistent. Now, the average sits at a staggering 10 per day. But the issues go beyond mere interruptions. Laurenzo’s team has observed a significant decrease – nearly 70% – in the number of code fragments Claude reads before making changes, plummeting from an average of 6.6 to just 2 by the end of March. The AI is increasingly opting to rewrite entire files instead of making targeted corrections, a telltale sign of superficial analysis.

“When the thinking is shallow,” Laurenzo explained in a GitHub post, “the model defaults to the cheapest available action: editing without reading, stopping without finishing, shirking responsibility for failures, opting for the simplest solution instead of the correct one.” This raises serious questions about Anthropic’s recent update (version 2.1.69) and its impact on Claude Code’s core reasoning capabilities.

Token usage, security concerns, and a flawed subscription model

The issues with Claude Code don't end with its declining accuracy. Users are reporting a sudden and unexplained spike in token usage, quickly exhausting their allocated limits. Adding to the growing unease, a recent leak of Claude Code’s source code has further eroded trust in the platform’s security. Laurenzo highlights a fundamental flaw in Anthropic’s subscription model: “The current model doesn't differentiate between users who need 200 tokens of thought per response and those who require 20,000,” effectively penalizing those who rely on Claude Code for complex workflows.

The rapid deterioration of Claude Code’s performance represents a significant setback for Anthropic and a cautionary tale for the wider AI industry. While the promise of AI-powered coding assistants remains alluring, as AMD's experience demonstrates, shortcuts and superficial solutions can quickly undermine even the most promising innovations. The question now is whether Anthropic can address these critical flaws before Claude Code’s reputation is irrevocably damaged.