AMD's senior director of AI has publicly stated that Claude has regressed in capability and cannot be trusted to perform complex engineering tasks. The criticism from a major tech company's AI leadership is significant—when the people building AI hardware say they can't trust frontier AI models for serious work, that's worth paying attention to. According to PC Gamer, the AMD executive noted that Claude's performance has declined, particularly for complex engineering tasks. The claim is that the model was more capable previously, and that recent versions have become less reliable for the kind of technical work that engineers depend on. This raises uncomfortable questions about AI model stability and regression. The headline claim is "Claude has regressed." If true, that's a problem. Companies and developers have been building workflows around frontier AI models, trusting that capability will improve or at least remain stable. If models can regress—get worse over time—that undermines trust and makes it difficult to build reliable systems on top of them. Model regression isn't unprecedented. OpenAI faced similar criticism when users claimed GPT-4 had gotten worse over time, particularly on coding tasks. The company eventually acknowledged making changes to improve efficiency that may have affected performance on some tasks. Anthropic, the company behind Claude, has been vocal about safety and reliability. They've positioned Claude as a more careful, thoughtful alternative to competitors. But if users—especially sophisticated technical users like AMD engineers—are seeing capability declines, that's a problem for the company's reputation. There are several possible explanations. One is that the model genuinely regressed due to changes in training, fine-tuning, or inference optimizations. Companies are constantly tweaking models to improve speed, reduce costs, or adjust behavior, and sometimes those changes have unintended consequences. Another possibility is that user expectations have risen faster than model capabilities. Early experiences with a new model can feel magical, but over time, as you push the limits and discover edge cases, the shine wears off. What felt impressive in week one might feel inadequate in month six. A third explanation is task-specific. Maybe Claude improved on some dimensions while getting worse on others, and AMD's use cases happen to fall into the "got worse" category. AI models are not uniformly good at everything, and tradeoffs are common. Regardless of the cause, the statement from AMD is damaging. When a prominent engineer at a major tech company publicly says they can't trust your AI for complex work, that's a credibility hit. It also highlights a broader issue with frontier AI models: they're powerful, but they're also opaque and unpredictable. You can't fully understand why they succeed or fail on specific tasks, and you can't guarantee consistent performance over time. For companies betting on AI to transform their workflows, that's a risk. If the model you've built your processes around suddenly gets worse—or gets shut down, or changes its pricing, or shifts its policies—you're exposed. The technology is impressive. But impressive isn't the same as reliable, and for engineering work, reliability matters more than impressive demos. When AMD says they can't trust Claude for complex engineering, it's a reminder that we're still in the early days of this technology, and not all progress is linear.