Participated in the SAT (Scholastic Assessment Test) of American universities.
Wrote script dialogues in the style of the characters from “Friends”.
Planned how to destroy the world.
Generated keywords to guide AI in painting.
……
ChatGPT is trending online.
It answers user questions in almost any field with seemingly “human-like” language.
And it has a certain degree of understanding of context, which amazes people.
Today, let’s learn more about ChatGPT.
ChatGPT is an exploratory application of the GPT deep learning model.
▲OpenAI’s official introduction describes ChatGPT as an “optimized dialogue language model”.
GPT stands for Generative Pre-trained Transformer, a text generation deep learning model trained on publicly available internet data. It is used for question answering, text summarization, machine translation, classification, code generation, and conversational AI.
GPT-1 was born in 2018, capable of natural language inference, question answering, commonsense reasoning, and semantic similarity recognition. It further developed into GPT-3, whose conversational abilities are now almost indistinguishable from humans, even surpassing them. The recently popular ChatGPT is an application built upon the GPT-3.5 foundational language model.
How is ChatGPT trained as an AI?
Creating an iterative feedback loop is crucial for AI development. ChatGPT utilizes Reinforcement Learning from Human Feedback (RLHF) and fine-tunes its model using Proximal Policy Optimization (PPO) through multiple iterations of the learning process.
It trains an initial model using a “pre-training-fine-tuning” work paradigm: human AI trainers provide dialogues, playing both roles – the user and the AI assistant. AI trainers can access the dialogue responses generated by the model and help adjust the AI’s replies.
▲ChatGPT Training Method
Once online, ChatGPT uses a large amount of user feedback interactions for AI learning, turning them into a key driver for AI advancement. The more information users input, the more opportunities ChatGPT has for deep learning and continuous improvement, leading to a deeper understanding of human language and “rumination”.
Does AI, including ChatGPT, possess human-like intelligence?
▲Alan Turing, known as the father of computer science and artificial intelligence, proposed a test method to determine whether a machine possesses intelligence – the Turing Test.
One standard for determining whether AI possesses artificial intelligence is whether it can pass the Turing Test. Simply put, it involves putting a machine in a small black room, allowing it to converse with a human, and then discerning whether the respondent in the room is a machine or a human. If it cannot be distinguished, it indicates that the machine has a level of intelligence comparable to that of a human, thus passing the test.
So far, no AI model has truly passed the Turing Test.
Professor Guoqiang Qiang from Tsinghua University
Chief Researcher at the Intelligent Industry Research Institute (AIR) Nie Zaiqing
Provided in-depth analysis of ChatGPT’s innovation, characteristics, and application prospects.
What kind of product is ChatGPT?
Essentially, ChatGPT is a statistical language model. Of course, it is still just a language model at this point, but it should gradually evolve into a multimodal model, such as matching images to corresponding text and providing understanding of images. Conversing with people is just one aspect of its capabilities; its strength lies more in content generation, including potentially more applications in writing in the future.
What distinguishes ChatGPT from existing voice chatbots in China?
Previous chatbots were more task-oriented robots confined to specific domains, executing certain functions through dialogue. The biggest challenge in achieving open-domain dialogue previously was the vast variety of topics input by users, making it difficult to obtain sufficient supervisory signals for training.
ChatGPT achieves open-domain dialogue and generates writing content. It estimates supervisory signals for any question automatically using a reward model learned from human feedback, fully utilizing reinforcement learning’s exploration capabilities to tackle the challenge of a vast open-domain task space.
▲The page notes that there may occasionally be incorrect, harmful, or biased information; limited understanding of the world post-2021, etc.
Currently, ChatGPT is not perfect; it can “seriously misinterpret” and struggles to genuinely understand and answer questions with specific knowledge points in open domains. In dealing with open-domain questions, it is still based on probability, generating the next value based on the previous one. Thus, while ChatGPT seems to answer professionally, users unfamiliar with the subject may find its responses satisfactory, but they are actually incorrect.
Is the content generated by ChatGPT original, or is it merely a copy and paste of information?
It can be considered creative content since it is generated by ChatGPT based on extensive learning and probabilities. Although it currently cannot systematically propose innovative viewpoints, some of its generated content may inspire scientific work.
Will ChatGPT replace some jobs in the future?
ChatGPT can serve as an assistant to humans, significantly improving work efficiency, such as helping users with writing, coding, business proposals, advertisement copywriting, translation, and more. Ultimately, the replacement of jobs may still require human oversight and confirmation.
Do language interaction products like ChatGPT pose privacy risks?
User questions inherently carry information; whether there is a leak depends on whether the product’s developers, operators, and data managers have handled relevant information appropriately, preventing personal privacy from leaking or being used for other purposes.
Source: Guangming Network, China News Service
Contact Number丨 0512-52821036
Editor丨 Youth君
Share, View, and Like
Giving Machines Humanity Through Dialogue
