Coding abilities are upending the valuation logic of large models

The article discusses how coding capabilities are reshaping the valuation of AI large models. It explores how proficiency in programming tasks influences market expectations and investment in AI development.

Moonshot AI

Prompt, RAG, or fine-tuning: what actually makes your LLM learn

In this article, Sergey Proshchaev, a Tech Lead in FinTech & E-commerce, compares three approaches to teaching an LLM domain knowledge: prompting with context, RAG, and QLoRA fine-tuning. Using one model (Qwen3-8B) and one GPU (RTX 4090), he tested all three on the same 30 questions. The results show that RAG is best for factual questions, QLoRA excels at format and terminology, but fails on updated data, and prompting is a good starting point but has limitations.