Are Large Language Models Really a Silver Bullet? Costs, Limits, and Alternatives

Large Models Treated as a Silver Bullet

After ChatGPT’s popularity, large models are often seen as a panacea for problems such as booking tickets via conversation, generating SQL from natural language, assisting elderly users, or detecting sensitive information.

Are Large Models a Silver Bullet?

Although integrating software functions with large models can improve accuracy through extensive training, real‑world usage faces several drawbacks.

High Cost

Software’s low expansion cost has driven its rapid growth, but large models incur substantial expenses. For example, an application handling 300 QPS on Alibaba Cloud costs about 9 CNY per day for the server, while the same traffic using OpenAI’s GPT‑3.5‑Turbo would cost roughly 1152 CNY per day, over a hundred times more.

Slow Computation Speed

ChatGPT can stream tokens for chat, but generating full responses for tasks like API parameter creation or SQL requires waiting for the entire output, which feels sluggish compared to traditional high‑performance computing expectations.

Unnecessary Features and Legal Risks

Large models can write poetry, tell stories, or even generate software crack codes, but these capabilities add little value for generating application interfaces and may introduce bias or security vulnerabilities, such as the “ChatGPT grandma” exploit.

What NLP Techniques Exist Besides Large Models?

Traditional NLP approaches include:

Rule‑based models (hard‑coded regex patterns, e.g., ChatterBot, Will).

Statistical language models (tf‑idf, PCA, topic and sentiment analysis).

Neural language models (RNN, LSTM, before being surpassed by Transformer‑based large models).

Future Directions

Text‑focused assistants such as writing helpers, speaking coaches, and expert consultants.

Integration platforms (e.g., Zapier, Alfred) that provide a unified natural‑language interface for multiple software tools.

Other simple tasks like text classification, topic detection, and sentiment analysis may still be better served by traditional NLP methods.

The author, not an AI specialist, offers this overview to spark discussion among experts.

https://zhuanlan.zhihu.com/p/643486458

https://zh.wikipedia.org/wiki/大型语言模型

https://github.com/Mooler0410/LLMsPracticalGuide

https://github.com/gunthercox/ChatterBot

https://pypi.org/project/will/

https://zapier.com/

https://www.alfredapp.com/