Claude變蠢了，新模型發布前的黑暗時刻？

🔑 增強重點摘要

• •Anthropic has been aggressively implementing 'model distillation' and quantization techniques to reduce inference costs, which users often misinterpret as a reduction in reasoning depth or creative capability.
• •The phenomenon of 'model drift' or perceived degradation is frequently observed in the industry during the transition phase between major model generations, as developers shift compute resources toward training the next iteration.
• •Anthropic's recent API updates have prioritized latency reduction and throughput, which has inadvertently altered the 'temperature' and response style of the models, leading to user reports of 'lazier' or less nuanced outputs.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線