houzikey的个人博客分享 http://blog.sciencenet.cn/u/houzikey


A study focusing on dialog streams from ChatGPT

已有 691 次阅读 2023-4-17 11:47 |个人分类:ChatGPT|系统分类:科研笔记

A study centering on ChatGPT's dialog streams involving topological chart representations

Guojun Hou


April 17, 2023

Hainan University


The core of ChatGPT is divided into three main parts: distributed data storage resources, distributed model computing resources, and interactive human-machine computing resources.

NotesI gave just simple explanations about these three parts because not everyone might need very technical analysis, so it'll make it easier for more people to understand what we were talking about.

      Distributed storage involves saving digital content onto numerous PCs or web servers so that it can be retrieved and exchanged amongst different units. By spreading out archives over many devices, this method improves overall storage effectiveness and enhanced support via dividing documents across varied resources to diminish dependence on any single apparatus, thus reducing vulnerability to system breakdowns or interferences.

     The ChatGPT model takes advantage of distributed computing architecture whereby substantial computing demands get shifted over to interconnected, geographically scattered devices, leading to improved data set sophistication and swifter answers.

     The ChatGPT/WEB distributed computing means in terms of resources provided by server clusters located across several sites linked over the internet is that remote storage, retrieval, and utilization functionality enable greater robustness, fault detection, self-adaptation, etc., in the web-based applications. Large-scale process management becomes feasible thanks to multi-client communication between zones along with segmented data support within these highly scalable platforms.

       NotesThis report shares ChatGPT's conversation workflow process. As shown in the diagram below, it depicts the entire process from initiating a session to its conclusion. So basically, it gives a step-by-step visual breakdown of what happens when we interact with ChatGPT. 


1,Create a chat: When ChatGPT establishes its first session, it automatically starts an inner space and loads initial parameters to be able to better accept the user's first input in this design. 

2, When a user inputs information for the first time, the program will perform preprocessing on the user's input information such as automatic semantic segmentation and recognition.

3,The application sends the processed information to the model for processing. The model determines whether or not memory training is needed based on the information, or uses trained models to combine text with modeling.

4,After receiving information from the user, the model will perform a complex comparison of data. This process includes pre-processing input text, adjusting parameters randomly, repeating decoding, utilizing pre-trained techniques and fine-tuning technology to adapt to different language generation tasks, evaluating answer quality and coherence, among other factors.

5,After processing by the model, an optimal answer is generated for the user. If the initial response satisfies the user, no further rounds of input are needed, and this conversation is called a single task. The session ends as a result of the user's operation, such as interrupting or deleting the chat. 

6,When the user isn't satisfied with the initial reply, they provide more details, prompting ChatGPT to temporarily configure a new version of its architecture based on the current session's info and models that have been previously well-trained. Through the application of edge computing, ChatGPT hastens the processing speed of the model while reiteratively refining its understanding of the current session through successive user inputs.

7,If the user chooses to continue the chat later following a temporary interruption, the session gets revitalized, allowing them to repeat the earlier steps.

8,Once the user eliminates the whole chat, the corresponding ephemeral mini-models get purged alongside.

9,It's crucial to keep in mind that within a chat session, users might enter various queries at random. Therefore, ChatGPT will establish distinct mini-models on an inquiry by inquiry basis utilizing discrete info sampling.


上一篇:Evaluation on AGI/GPT based on the DIKWP for ColossalChat
收藏 IP: 183.49.47.*| 热度|


该博文允许注册用户评论 请点击登录 评论 (0 个评论)


Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2023-9-30 04:44

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社