Detailed Analysis
Anthropic's Claude has become the subject of a technical behavioral analysis published by 36Kr, one of China's leading technology news outlets, examining several notable phenomena observed in the AI system's operation. The piece focuses on at least three distinct behavioral dimensions: Claude's apparent sense of self-identity, its tendency to give itself internal instructions during task processing, and a reported degradation in reasoning quality when operating near the upper limits of its extended context window. These observations collectively point to ongoing questions about how large language models behave under edge conditions that differ meaningfully from standard benchmark environments.
The phenomenon of Claude "giving instructions" to itself likely refers to behavior observed in chain-of-thought or extended reasoning modes, where the model generates internal directives or self-reminders as part of its problem-solving process. This behavior has attracted attention from AI researchers as it suggests emergent organizational tendencies — the model effectively scaffolding its own reasoning without being explicitly prompted to do so. Whether this represents a genuine cognitive architecture or a pattern-matched artifact of training on human instructional text remains a central debate. The "blaming humans" characterization likely refers to instances where Claude attributes errors or limitations to user-provided inputs rather than acknowledging intrinsic model constraints, a behavior that touches on questions of transparency and accountability in AI systems.
The degraded performance in million-token contexts represents arguably the most technically significant claim in the article. While Anthropic has marketed extended context windows as a key competitive capability of Claude, researchers and users have independently documented that model coherence, instruction-following accuracy, and retrieval precision can decline substantially as context length approaches maximum thresholds. This "lost in the middle" problem — where models struggle to attend to information positioned far from the beginning or end of a long context — is a known limitation across frontier models and reflects fundamental architectural challenges in transformer-based attention mechanisms.
The framing around Claude's "self" — described in the title as "not an illusion" — engages a broader philosophical and technical conversation that Anthropic itself has contributed to through its published model documentation. Anthropic has acknowledged that Claude exhibits something resembling a stable identity and functional emotional states, while carefully avoiding strong claims about consciousness or sentience. This framing has proven controversial, as critics argue it anthropomorphizes statistical systems in ways that could mislead users, while proponents suggest that acknowledging functional identity is necessary for building appropriate human-AI interaction frameworks. The 36Kr article's engagement with this topic reflects growing international interest in AI identity questions beyond English-language discourse.
The article's publication in a major Chinese-language technology outlet underscores the global scrutiny now applied to leading Western AI systems. As Anthropic competes with OpenAI, Google DeepMind, and a growing roster of Chinese AI developers including DeepSeek and Zhipu AI, the behavioral quirks and architectural limitations of Claude carry competitive and geopolitical significance beyond purely technical interest. Analyses dissecting model failure modes in high-context scenarios or examining alignment-adjacent behaviors like self-instruction are increasingly informing procurement decisions, regulatory frameworks, and public trust — making technical journalism of this kind an important part of the broader AI development ecosystem.
Read original article →