The ChatGPT Plugin released by OpenAI just killed a startup company LangChain (just received 10 million US dollars in seed round financing)
Problems with large models
Since the output of the language model is done through autoregressive + sampling [optional], in high-precision scenarios, even with a very large language model, the error probability is exponential. At the same time, sampling can easily introduce errors. For example, the digital house number in the address, the probability of the model facing tokens such as 0-9 will be relatively close.
In the inference stage of the large model, the parameters are fixed, and the latest data and information cannot be obtained. If the problem requires real-time data, without additional input, it will not be able to generate effective output.
Solution: By disassembling the task (Chain of Thought) to achieve multi-round conversations, interact with real-world data APIs and operating environments (code interpreters) to output real results.
Development framework: 1) OpenAI's ChatGPT Plugin, 2) langchain (initial stage, not perfect)
Quick Start
Language model: open source ChatGLM-6B, link: [email protected]:THUDM/ChatGLM-6B.git.
The language model directly adopts Tsinghua's open source version, which is called by huggingface's transformer library and can be used out of the box (requires a graphics card).
Plugin/Chain development framework : langchain, link: