Anthropic's Claude Code Is Getting Dumber, and an AMD AI Director Publicly Called It Out

When a senior director of AI at one of the world's largest chipmakers takes to social media to complain that his coding assistant has gone soft, it's not just a personal gripe. It's a signal.

Sander Land, AMD's director of AI, posted a pointed critique of Anthropic's Claude Code tool on X in early April 2025, describing the AI coding agent as increasingly "dumb and lazy." The complaint resonated widely among developers and AI practitioners, many of whom have noticed a similar pattern: models that once delivered sharp, efficient code now seem to hedge, over-explain, and produce bloated outputs. Land's frustration wasn't abstract. He was talking about real productivity loss in real engineering workflows.

As reported by The Register, Land's critique centered on what many power users have described as a regression in Claude Code's capabilities -- not in raw intelligence, but in practical usefulness. The tool, which Anthropic positions as a premium AI-powered coding assistant, has allegedly become more verbose, less decisive, and prone to asking clarifying questions rather than executing tasks directly. For engineers working at scale, that kind of friction compounds fast.

This isn't a fringe complaint.

The "Lazy AI" Problem Has Been Brewing for Months

Across developer forums, GitHub discussions, and X threads, a growing chorus of users has flagged similar behavior in multiple large language models -- not just Claude. OpenAI's GPT-4 variants have drawn comparable criticism. The pattern is consistent: models that initially impressed with confident, concise outputs gradually shift toward safer, more hedged responses over time. Some users call it "model rot." Others call it alignment tax.

The underlying cause is debated. One theory holds that reinforcement learning from human feedback (RLHF) -- the process by which models are fine-tuned to be helpful, harmless, and honest -- inadvertently trains models to be cautious to a fault. When human raters consistently penalize confident-but-wrong answers more harshly than vague-but-safe ones, the model learns to equivocate. Over successive training rounds, this produces outputs that feel increasingly watered down.

Another possibility: Anthropic and its competitors are deliberately tuning models to reduce liability. A coding assistant that confidently generates flawed code could expose its maker to reputational or even legal risk. Better, from a corporate perspective, to have the model ask "Are you sure?" than to silently introduce a bug into production code.

But for someone like Land, whose job involves pushing AI tools to their limits inside one of the semiconductor industry's most demanding engineering environments, caution isn't a feature. It's a bug.

The tension here is fundamental. AI companies want their models to be safe. Power users want them to be useful. Those goals aren't always compatible, and the gap between them is widening as models are deployed in increasingly high-stakes professional settings.

Anthropic has not publicly responded to Land's specific critique. The company has, however, acknowledged in broader communications that balancing helpfulness with safety remains an active area of research. Claude's system prompt and behavioral guidelines are regularly updated, and Anthropic has positioned its Constitutional AI approach as a more principled alternative to pure RLHF. Whether that approach is contributing to the perceived laziness is an open question.

What makes Land's complaint notable isn't just his seniority at AMD. It's that he represents exactly the kind of user Anthropic needs to retain. Enterprise adoption of AI coding tools is accelerating rapidly, with companies like AMD, Google, Meta, and Microsoft integrating these systems into core development pipelines. If the tools start to feel like they're slowing engineers down rather than speeding them up, the business case erodes quickly.

And the competition isn't standing still. Google's Gemini models have made aggressive moves into the coding space. OpenAI continues to iterate on its Codex lineage. Startups like Cursor and Cognition Labs are building specialized coding agents that prioritize developer experience above all else. In this environment, a perception of declining quality -- even if the underlying model hasn't technically regressed -- can shift market share fast.

The Measurement Problem

Part of what makes this debate so difficult to resolve is that "dumber" and "lazier" are subjective assessments. Benchmarks like HumanEval, MBPP, and SWE-bench measure specific coding capabilities under controlled conditions. They don't capture the lived experience of using a tool for eight hours a day on a complex codebase. A model might score higher on a benchmark while simultaneously feeling worse to use in practice -- because it's been optimized for the benchmark rather than for the workflow.

This is a known failure mode in AI development. Goodhart's Law -- "when a measure becomes a target, it ceases to be a good measure" -- applies with particular force to language models. Companies optimize for the metrics they can track, and those metrics don't always align with what users actually care about.

Land's complaint also touches on a deeper issue: the opacity of model updates. When Anthropic pushes a new version of Claude, users often have no way to know what changed. There's no changelog. No diff. The model just behaves differently one day, and users are left to figure out whether the change was intentional, incidental, or imaginary. For engineers accustomed to version control and reproducibility, this is maddening.

Some developers have started maintaining their own informal benchmarks -- sets of prompts they run periodically to track model behavior over time. It's a crude approach, but it reflects a real gap in the tooling. If AI companies want enterprise customers to trust their models, they'll need to provide more transparency about how and when those models change.

So where does this leave Anthropic? The company is in a strong position by most measures. Claude has a loyal user base, strong brand recognition among developers, and significant enterprise traction. But the "lazy AI" perception is a reputational risk that won't resolve itself. If Anthropic's alignment work is genuinely making the model less useful for professional coding tasks, the company faces a strategic choice: maintain its safety-first posture and risk losing power users, or find a way to offer different behavioral profiles for different use cases.

The latter approach -- sometimes called "adjustable alignment" or user-configurable safety settings -- is gaining traction in the industry. The idea is simple: let enterprise users dial down the hedging and verbosity when they're working in controlled environments where they understand the risks. Consumer-facing deployments would retain tighter guardrails. It's not a perfect solution, but it acknowledges that a single behavioral profile can't serve every user equally well.

Anthropic has hinted at moves in this direction. The company's system prompt architecture already allows some customization, and Claude's API offers parameters that influence response style. But the core model behavior -- the tendency to over-explain, to ask rather than act, to pad responses with caveats -- is baked in at the training level. Surface-level adjustments can only do so much.

What AMD's Frustration Tells Us About the Market

Land's public critique matters beyond the specifics of Claude Code because it illustrates a broader dynamic in enterprise AI adoption. Companies are moving past the honeymoon phase. The initial excitement of having an AI that can write code, summarize documents, and answer questions is giving way to harder questions about reliability, consistency, and integration. The bar is rising.

AMD itself is deeply invested in the AI hardware stack, competing with Nvidia for data center GPU market share. The company's MI300X accelerators are designed to run the very models that Land is criticizing. There's an irony there -- AMD is simultaneously selling the infrastructure for AI and struggling with the quality of the AI running on it. But it also gives Land's perspective a certain credibility. He's not a casual user. He's someone whose livelihood depends on these tools working well.

For Anthropic, the path forward likely involves more granular control, more transparency, and a willingness to let advanced users take the training wheels off. The company's research on interpretability -- understanding what's happening inside neural networks -- could eventually help diagnose why models become "lazy" after certain training procedures. But that's a long-term play.

In the short term, the message from AMD and from the broader developer community is clear: don't sacrifice usefulness on the altar of safety. The best AI coding assistant isn't the one that never makes a mistake. It's the one that makes engineers faster. Right now, for at least some high-profile users, Claude Code is moving in the wrong direction.

The stakes are high. Enterprise AI contracts are measured in millions of dollars. Developer loyalty, once lost, is hard to win back. And in a market where four or five serious competitors are fighting for the same customers, the margin for error is thin.