3000億的智譜和MiniMax，就靠兩個公式？

🔑 增強重點摘要

• •The 'two formulas' critique refers to the industry debate over whether these firms rely excessively on scaling laws (compute-heavy training) versus proprietary data moats, potentially leading to a 'valuation bubble' if model performance plateaus.
• •Zhipu AI has pivoted heavily toward 'Agent-centric' architectures, moving beyond raw LLM performance to focus on autonomous task execution and tool-use ecosystems to differentiate from pure chat-based competitors.
• •MiniMax has aggressively pursued a 'multimodal-first' strategy, integrating native audio and video generation capabilities into their core model architecture earlier than many domestic peers to capture the consumer entertainment market.

🛠️ 技術深入

• •Zhipu GLM-4: Utilizes a General Language Model (GLM) architecture based on a blank-filling objective, optimized for both understanding and generation, with specific enhancements for long-context retrieval and tool-calling.
• •MiniMax abab: Employs a proprietary mixture-of-experts (MoE) architecture designed to handle high-concurrency multimodal inputs, specifically optimized for low-latency voice-to-voice interaction.
• •Both firms have shifted toward 'Data-Centric AI' methodologies, utilizing synthetic data pipelines to augment training sets where high-quality human-annotated data is scarce.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線