Telegraph

7x24 real-time news feed, auto refreshBeijing Time：

🔄 Auto refresh 🔔 Sound 💬 Notify

All A·Important B·Normal #黑洞 ✕

B

[CCID Consulting: The return on investment for large models with more than 100 billion parameters is only 5%] "Science and Technology Innovation Board Daily" reported on the 6th that CCID Consulting's all-English special issue "Foresight" research shows that the "bigger is better" development paradigm in the generative AI industry is encountering challenges from business reality. In structured business scenarios, the performance of large models reaches a plateau at 40-60 billion parameters. The return on investment for large models with more than 100 billion parameters is only 5%, while the return on investment for models with 7-30 billion parameters can reach 35%. Inference is a major cost black hole, accounting for 70%-90% of the cost of the model's entire life cycle; giant single models also bring compliance governance risks. The report proposes a new paradigm of "precision intelligence", relying on the four major components of domain fine-tuning models, RAG retrieval enhancements, tool agents, and human-in-the-loop verification to match model complexity to business tasks and achieve optimal business results with the least resources. The logic of industry competition has shifted from competing in model parameter scale to competing in the implementation of intelligent architecture. (Kechuangban Daily reporter Li Mingming)

#Generative AI #RAG #large model #fine-tuning #black hole11-06 18:59