LLMs Are Dead. Small Models Are the Future.

Nowadays, LLMs are strong mainly through text prediction and data accumulation. I have to admit that they can handle many cross-domain work, including the hard ones, but their training data is messy and largely outdated, and according to some, only “tutorial-level difficulty”. Their advantages are mostly the same — “coding”, “reasoning”, “long memory” — which, after years of advertising, became plain and not as useful.

Instead, what we are increasingly in need of are efficient small models for deep, narrow areas to ensure real professional performance, not just surface work, which will definitely outperform LLMs in their own field. For example, each one for emotional communication, astrophysics, art, etc. The role of LLMs in the future is just to wire those small models up, to do integration and cross-domain work to produce more grounded achievements.

I'm Richael, and I'm fine-tuning small models like Qwen 14B to improve the performance of Soulor — a multi-companion and simulation app — and VNportal — a music recreation and creation community platform with fair, controllable AI assistance. I hope to redefine AI's role in the modern and future world. If you're interested, welcome to visit my apps, share, and support. Thanks for reading, and have a good day!