The global AI market is currently dominated by a few major corporations, including Microsoft, Google, Meta, and Anthropic. These companies maintain their market position by building massive data centers and controlling access to advanced computing resources, effectively creating a high financial barrier to entry for competitors.
The current AI industry relies on the Transformer architecture, which dates back to 2017. This architecture suffers from quadratic complexity in its attention mechanism, meaning that increasing input data volume leads to a disproportionately higher computational load. Consequently, industry leaders rely on massive energy consumption and expensive hardware to maintain performance.
China has moved to restrict the open-source release of advanced AI models. The Ministry of Commerce of the PRC has introduced export control regulations that effectively prohibit the free publication of open-weight models, classifying such leaks as national security threats. Major Chinese companies like ByteDance, Kuaishou, and Alibaba have already shifted their flagship models to proprietary, closed-source systems.
Independent researchers are developing alternative architectures to challenge the current monopoly. The Mamba architecture, created by Albert Gu and Tri Dao, uses linear complexity to reduce hardware requirements. The RWKV project, led by Bo Peng, enables scalable training without the quadratic costs associated with traditional Transformers.
New approaches to continuous learning are emerging to address the static nature of current AI models. Researchers at Caltech have developed the FIP algorithm, while Tokyo-based Sakana AI is working on Darwin Gödel Machines. These technologies aim to allow AI agents to learn autonomously from experience without needing constant retraining on corporate servers.
