What is the impact on automotive technology?

Guide: What can ChatGPT do? It can replace Jack to write articles. Through the study of Jack's past articles, AI can summarize the structure and routines of Jack's articles. After that, you can input the core of the article and related data, and he can quickly form Jack’s article; it can replace Jack to do PPT at work, especially the PPT of report summary, which can be quickly formed.

This is true. Obviously, general-purpose AI models such as ChatGPT can subvert many things in the future. Then you should really understand technologies such as ChatGPT.

So this article combines relevant information to summarize:

(1) What is ChatGPT and what can it do?

(2) What is the technology behind ChatGPT?

(3) What are the current limitations of ChatGPT?

(4) What are the similar products of ChatGPT at home and abroad?

(5) What impact does ChatGPT have on automotive technology?

Help self-awareness and learning, and hope to bring some information and inspiration to everyone.

1. What is ChatGPT and what can it do?

ChatGPT, first of all, return to its English meaning Chat is chat, GPT is the abbreviation of Generative Pre-trained Transformer, translated into a generative pre-trained Transformer, Transformer is a deep learning model based entirely on the self-attention mechanism. So ChatGPT is a generative, pre-trained algorithm model that can chat. It belongs to the current hot generative AI artificial intelligence model, that is, AI generates human content.

ChatGPT is an artificial intelligence chat robot for ordinary users, currently mainly based on text chat, it uses advanced natural language processing (NLP) to have realistic conversations with humans. Currently ChatGPT can generate articles, fictional stories, poems and even computer code. ChatGPT can also answer questions, engage in conversations and, in some cases, provide detailed responses to very specific questions and queries.

Actually, chatbots are nothing new. They use keyword search technology and then match answers. This is very common in our daily life, such as Amazon’s Alexa, Apple’s SIRI, Tmall Genie, Baidu Xiaodu, etc. There are still many online customer service, and even the car language control we are familiar with. They are mainly command-based voice assistants based on tasks. However, ChatGPT does use a more refined big data training model, which has a wider range of applications and can be integrated into various applications.

One of the things that sets ChatGPT apart from other chatbots and NLP systems is its hyper-realistic conversational skills, including the ability to ask follow-up questions, admit mistakes, and point out nuances on topic. In many cases, it's basically difficult for a human to detect that they're interacting with a computer-generated robot if they don't tell you. Grammar and grammatical errors are rare, and the written structure is logical and clear.

Some features of ChatGPT include:

Generate human-like representations that mimic the style and structure of the input data;

Generates a response to a given prompt or input text. This may include writing a story or answering questions;

Generate text in multiple languages;

Modify the style of the generated text (e.g., formal or informal);

Ask clarification questions to better understand the intent of the input data;

Reply with text that is consistent with the context of the conversation, such as providing follow-up clarification or understanding references to previous questions;

Other generative AI models can perform similar tasks on images, sound, and video.

In addition, ChatGPT can perform fine-tuning training: the process of adapting LLM to a specific task or field by training on a smaller related data set, which is also the business model that ChatGPT is currently expanding. Professional legal advisors, professional automotive think tank experts.

Second, what is the technology behind ChatGPT?

ChatGPT is OpenAI's latest language model NLP (Natural language processing), which is based on the large language model LLM (Large Language Model) model GPT-3 plus the use of supervised learning and human feedback reinforcement learning RLHF (Reinforcement Learning from Human Feedback ) to fine-tune ChatGPT formation.

Three technical keywords:

NLP(Natural language processing) natural speech processing;

LLM (Large Language Model) large language model;

RLHF (Reinforcement Learning from Human Feedback) reinforcement learning in human feedback;

NLP natural language processing is the interaction between human language and computer, with a relatively large scope. The popular technology in the NLP field is the deep learning model, which mainly relies on the following key technologies: the magically modified LSTM model and a small amount of improved CNN model, RNN as a typical feature extractor; Sequence to Sequence (or encoder-decoder also Yes) + Attention is a typical overall technical framework for various specific tasks.

The LLM large-scale language model, which is currently the model of ChatGPT, is a subset of artificial intelligence. As the name implies, "big" means massive data. is the entire Internet and millions of books (roughly 300 billion words) to produce human-like responses to conversations or other natural language input.

Its main key technical points are:

Word Embedding: An algorithm used in LLM to represent the meaning of a word in numerical form so that it can be fed into and processed by an AI model by mapping words to vectors in a high-dimensional space with Words of similar meaning are placed closer together.

Attention Mechanism: An algorithm used in LLM that enables the AI to focus on specific parts of the input text, such as sentiment-related words in the text, when generating output. This allows the LLM to take into account the context or sentiment of a given input, resulting in a more coherent and accurate response.

Transformers: A neural network architecture popular in LLM research that uses self-attention mechanisms to process input data, allowing them to efficiently capture long-term dependencies in human language.

Transformers are trained to analyze the context of the input data and weight the importance of each part of the data accordingly. Because this type of model learns context, it is often used in natural language processing (NLP) to generate text that resembles human writing. Due to its attention mechanism, which has an extremely long memory than previous deep learning algorithms such as RNN, GRU and LSTM, Transformer can "attend" or "focus" on all tokens generated before. In theory, attention mechanisms, given sufficient computational resources, have an infinite window to refer to, and are thus able to use the entire context of the story when generating text.

Large language models (such as GPT-3) are trained on large amounts of text data from the Internet and are able to generate human-like text. In fact, their objective function is the probability distribution of word sequences (or token sequences), so that they are able to predict what the next word in the sequence will be (more details below), so they may not always produce output that aligns with human expectations or ideals.

In practice, however, LLM models are trained to perform some form of valuable cognitive work, and there are clear differences between how these models are trained and how we want to use them. Although mathematically speaking, a machine-computed statistical distribution of word sequences can be a very efficient choice for modeling language, as humans, we generate language by choosing the sequence of text that best fits a given situation, and using Our background knowledge and common sense guide the process.

The problem with the performance of LLMs when language models are used in applications that require a high degree of trust or reliability, such as dialogue systems or intelligent personal assistants:

Lack of help: not following explicit instructions from the user;

Illusion: Models fabricate facts that do not exist or are false;

Lack of explainability: A human cannot understand how it came to a particular decision or prediction.

Generating biased or harmful output: A language model trained on biased/harmful data may reproduce this result in its output.

But how exactly did the creators of ChatGPT use human feedback to solve the alignment problem? At this time, it is necessary to perform reinforcement learning in RLHF (Reinforcement Learning from Human Feedback) human feedback on LLM. The method generally consists of three distinct steps:

Supervised fine-tuning steps: First, similar to autopilot labeling, the human labeler writes down the expected output response to train LLM. For example, ChatGPT is fine-tuning the GPT-3.5 series. But obviously the cost of manual labeling is very high, and I heard that Open AI has invested a lot of capital in this area, so the scalability cost of the supervised learning step is very high.

The "mimicking human preferences" step: In this phase human labelers are asked to vote on a relatively large number of SFT model outputs, thus creating a new dataset consisting of comparative data. A new model is trained on this dataset. This is called the Reward Model (RM). Sequencing output is much easier for a labeler than producing it from scratch, a process that can be scaled up more efficiently.

Proximal Policy Optimization (PPO) step: It is the step where the reward model is used to further fine-tune and improve the SFT model, the result of this step is the so-called policy model, which is constantly adjusted according to the actions the agent is taking and the rewards received current strategy, and it limits the change of strategy to a certain distance from the previous strategy.

So on the basis of continuous intensive training based on the input of human annotators, they essentially endow ChatGPT with answers. With reinforcement learning in RLHF human feedback, ChatGPT uses human feedback in the training loop to minimize harmful, inauthentic and/or biased output.

Of course there are algorithms, but also a computing center for storing and processing data and computing processing. The computing center behind ChatGPT is Microsoft's Azure cloud computing center.

3. Limitations of ChatGPT

Then the limitations of such a powerful ChatGPT may be understood after familiarizing with the core technology behind it. It is trained based on the accumulation of huge data of human language and characters. It cannot have thinking and innovation, mainly based on the past. The training data is combined.

In addition, in terms of accuracy, there are currently two core related functions that make the language model more accurate:

The ability of the LLM to retrieve information from external sources.

The ability of LLMs to provide references and citations for the information they provide.

So the answer of ChatGPT is limited to the information already stored in its training. For example, the ChatGPT we use now is based on the network data before 2021 and related book knowledge for learning and training. captured in its static weights. (This is why it cannot discuss events occurring after 2021, when the model was trained.) Being able to obtain information from external sources would allow LLM to access the most accurate and up-to-date information available, even when that information changes frequently ( For example, the company's stock price), but obviously not currently.

In addition, the fine-tuning given by the artificial annotator of the artificial intensive training is also the limitation of the current ChatGPT. First of all, ChatGPT cannot get rid of the shadow of the artificial annotator such as position, preference, etc.

4. What are the similar products of ChatGPT at home and abroad?

So is similar to ChatGPT thriving? Does no one else have such a product? In fact, as mentioned above, ChatGPT is a kind of LLM large-scale language model. In fact, many Internet companies have large-scale language models. As for why the new force Open AI released it first and caused a sensation, in fact, traditional forces are most afraid of making mistakes, so I gave it An opportunity for new forces to bravely try and make mistakes.

In fact, in addition to the GPT-3 from OpenAI, the well-known foreign language models also include:

Google's PaLM or LaMDA;

Galactica or OPT for Meta;

Nvidia/Microsoft's Megatron-Turing;

Jurassic-1 from AI21 Labs;

Of course, in addition to the above, Amazon, Microsoft, GitHub, Apple, IBM, etc. abroad have also built natural language processing frameworks that provide different features and functions. These include digital assistants, predictive coding tools, and chatbots, as well as large language models.

Baidu in China has invested in the development of related technologies similar to ChatGPT. The name of the project is determined to be Wenxin Yiyan, and the English name is ERNIE Bot. It will be open to the public after the internal test is completed in March. At present, Wenxin Yiyan is doing the sprint before going online.