馬斯克洩密：Claude Opus 5T 參數，Sonnet 1T

🔑 增強重點摘要

• •The alleged leak occurred during a live-streamed Q&A session on X, where Musk contrasted the efficiency of Grok's sparse architecture against the dense parameter counts he attributed to Anthropic's models.
• •Industry analysts note that the 5T parameter count for Claude Opus suggests a Mixture-of-Experts (MoE) architecture rather than a traditional dense model, as a 5T dense model would be prohibitively expensive to serve at current inference latency standards.
• •Anthropic has not officially confirmed these figures, maintaining their policy of not disclosing specific parameter counts, which complicates the industry's ability to verify the correlation between parameter scale and benchmark performance.

🛠️ 技術深入

• •The 5T parameter count for Claude Opus is widely interpreted by researchers as the 'total' parameter count in an MoE setup, with significantly fewer 'active' parameters per token.
• •Grok 4.2's 0.5T total parameter count utilizes a highly optimized sparse routing mechanism, allowing it to maintain lower compute costs while matching performance of larger models on specific reasoning tasks.
• •The discrepancy in parameter counts highlights a shift in the industry toward 'parameter efficiency' where model performance is increasingly decoupled from raw parameter volume.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線