Introduction: The great feats that humans think are just the beginning... The technological progress we will make in the next 100 years will far exceed everything we have achieved since controlling fire to inventing the wheel. —By Sam Altman
Explanation: After ChatGPT was released, I experienced its dialogue, translation, programming, and writing effects for the first time. As lines of green words jumped out quickly... a feeling of triggering a mysterious power, I had a premonition that a higher-dimensional behemoth appeared. Such a phenomenal thing may have a huge impact on various industries. Even if the blind feel the elephant, it is worth touching it. So I collected 100+ articles and reports at home and abroad one after another, sorted them out a bit during the Spring Festival holiday, and shared them with everyone. After sorting out, I found that the article is too long. I plan to divide it into the first part "Background + Technology", and the next part will be "Product + Business". So, hurry up and update~
01. Background
**1.1 What are ChatGPT and OpenAI?
**
What is ChatGPT?
ChatGPT is an artificial intelligence chat robot program developed by OpenAI, the top AI laboratory in the United States. It will be launched in November 2022, and it will exceed 1 million users within a week of its launch. The program uses a large language model based on the GPT-3.5 architecture and is trained by reinforcement learning,
What is OpenAI?
OpenAI is an AI laboratory in the United States, a non-profit organization, positioned to promote and develop friendly artificial intelligence to benefit mankind as a whole. OpenAI was founded in late 2015 by Elon Musk and former YC president Sam Altman.
Who is Sam Altman?
Musk does not need to introduce~ Samuel Altman, an American born in 1985, a geek who dropped out of Stanford University's computer department to start a business. The president of YC, a well-known American venture capital institution, and the heir of Paul Graham, the godfather of entrepreneurship in Silicon Valley. If you don’t know YC yet, you may know a famous person: the president of YC China (now renamed Qiji Chuangtan), the overseas station of YC, is the famous Lu Qi.
OpenAI development history (mainly from Wikipedia)
At the end of 2015, OpenAI was established. The goal of the organization is to open patents and research results to the public through "free cooperation" with other institutions and researchers. In 2016, OpenAI announced that it would create "universal" robots, hoping to prevent the catastrophic impact of artificial intelligence and promote artificial intelligence to play a positive role. On March 1, 2019, the OpenAI LP subsidiary was established with the goal of profitability and commercialization. On July 22, 2019, Microsoft invested $1 billion in OpenAI, and the two parties cooperated to develop artificial intelligence technology for Azure (Microsoft's cloud service). The GPT-3 language model was announced on June 11, 2020, and Microsoft obtained an exclusive license on September 22, 2020. On November 30, 2022, OpenAI released a natural language generative model called ChatGPT, which interacts in a conversational manner. January 2023: Microsoft and OpenAI are in talks to invest $10 billion and hope to incorporate OpenAI's artificial intelligence technology into Word, Outlook, Powerpoint and other applications.
02 Technology
2.1 The core competitiveness of ChatGPT
A brief analysis of the three core elements of AI: data, algorithms, and computing power. In addition, in the early stage of a new thing, its founder's original intention and vision are also very worthy of attention, so an additional layer-the analysis of the concept level is added.
Data layer:
Pre-trained a model with 175 billion parameters on a corpus of 300 billion words (60% of the training corpus came from C4 in 2016 - 2019 + 22% from WebText2 + 16% from Books + 3% from Wikipedia).
Algorithm layer:
Powerful Informative Responses from Reinforcement Learning from Human Feedback (RLHF): text-davinci-003 typically takes longer to generate than text-davinci-002 ([29)(]). ChatGPT's responses are more verbose, so that users must explicitly ask "answer me in one sentence" to get a more concise answer. This is a direct product of RLHF. Unbiased Responses: ChatGPT typically gives very balanced answers to events involving the interests of multiple entities, such as political events. This is also a product of RLHF. Rejecting inappropriate issues: This is a combination of the content filter and the model's own capabilities triggered by RLHF, where the filter filters out some and the model rejects some. Rejecting questions outside of its knowledge: for example, rejecting new events that occur after June 2021 (since it was not trained on data after that). This is the most amazing part of RLHF, because it enables the model to implicitly distinguish which problems are within its knowledge and which problems are not. ——By Fu Yao, "Wanzi Dismantling ChatGTP Technology Roadmap"
Computing layer:
Behind ChatGPT is inseparable from big models, big data, and big computing power. Behind ChatGPT becoming a milestone of AIGC is the large model training supported by the development of computing power and the big data formed in the digital age to achieve the current effect. ChatGPT, developed by OpenAI, is a fine-tuned GPT-3.5 series model with as many as 175 billion model parameters, and was trained at the beginning of this year. Model training is inseparable from the support of big data. The public crawler data set mainly used by OpenAI has a human language data set with more than one trillion words. In terms of computing power, GPT-3.5 is trained on the Azure AI supercomputing infrastructure (a high-bandwidth cluster composed of V100GPUs), and the total computing power consumes about 3640 PF-days (that is, one trillion calculations per second, running for 3640 days ).
Idea layer:
- mission and vision. From the official website of OpenAI: OpenAI is an AI research and development and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity. Four main points of the OpenAI charter (dashes are the author's personal understanding): * Widely benefit society - altruism * Focus on long-term security issues - nanny :)> We worry that general artificial intelligence will evolve into a fierce competition in the later stage of development, resulting in a lack of Sufficient time for security precautions. So if a safety-focused project that aligns with human values gets us closer to AGI, we commit to stop the race and assist that project instead. We will design specific cooperation plans for individual situations. However, a typical trigger condition might be "more than half of the probability that this project will successfully develop general artificial intelligence in the next two years." * Leading technology research - frontier * Maintaining willingness to cooperate - open 2. Excerpts from the founder's speech: From Sam Altman Moore's law of all things We need to design a system to embrace this technological future, and then make up most of the value of the future world Assets (companies and land) are taxed so that the resulting wealth is distributed fairly. Doing so can make the future society much less divisive and allow everyone to participate in the distribution of benefits. The coming revolution will revolve around our most extraordinary abilities: thinking, creating, understanding and reasoning. On the basis of the three major technological revolutions (agricultural revolution, industrial revolution and computer revolution), we will enter the fourth stage: the artificial intelligence revolution. If we, as a common society, undertake this revolution responsibly, enough wealth will be generated to give everyone what they need. 3. Technical concept (From Zhang Junlin "The Road to AGI: Essentials of Large Language Model (LLM) Technology") How does OpenAI view LLM? Looking back at the technologies it has continuously introduced, it can be seen that it has basically firmly regardedIt is the only way to AGI. Specifically, in the eyes of OpenAI, the future AGI should look like this: there is a task-independent super-large LLM, which is used to learn various knowledge from massive data, and this LLM solves various problems by generating everything. practical problems, and it should be able to understand human commands so that it can be used by humans. In fact, the understanding of the concept of LLM development, in the first half, is to "build a super-large LLM that has nothing to do with tasks, and let it learn various knowledge from massive data". It's the second half. OpenAI's concept is relatively advanced, and its self-positioning has been set relatively high from the beginning, and it has always been unswervingly exploring whether the above methods can realize AGI. The reason why OpenAI was able to make ChatGPT is that one has a relatively high positioning, and the other is that it is free from external interference and has a firm attitude.
2.2 Evolution of GPT
Model Dimensions (By Fu Yao)
Evolution of Large Model Technology Architecture
Funders who develop large models
Statistical graph of data volume and large model performance
Why is the big model so versatile?
2.3 ChatGPT experience and analysis
Analysis at the experience level: the comprehension ability is almost the same as that of a real person, and the robustness of the model is very good. After moral training, without evaluating people, it is difficult for you to grasp its handle. Without this article, chatGTP would have been ruined long ago, and a bunch of threats and war of words were enough to make it offline. More emphasis on fact, not opinion. You seem to be chatting with a friend who is led by reason rather than emotions. Chinese is slightly inferior to English. If you let it be a sonnet, you will be amazed by the beautiful rhymes. If Xu Yuanchong is alive, this old man who loves to play Chinese, English and French rhymes will probably find a match. Don't understand the world beyond 2022. For example, in the 2022 World Cup in Qatar, it will honestly say that it does not know the world after 2022. This may also be the biggest difference between ChatGTP and search engines. After all, a one-year information gap is enough to discount a lot of knowledge. Finally, if you plant a booby trap in a question, you may find it's serious nonsense.
Technical layer analysis (By Zhang Junlin): The biggest contribution of ChatGPT is that it basically realizes the interface layer of the ideal LLM (Large Language Model), allowing LLM to adapt to people's customary command expressions, instead of allowing people to adapt to LLM. Rack your brains to come up with a command that can work (this is what the prompt technology was doing before the instruct technology came out), and this increases the ease of use and user experience of LLM. It was InstructGPT/ChatGPT who first realized this problem and gave a good solution, which is also its biggest technical contribution. Compared with the previous few shot prompting, it is a human-computer interface technology that is more in line with human expression habits and interacts with LLM. The emergence of large models such as GTP/BERT may lead to the death of some intermediate tasks. Typical intermediate tasks include: Chinese word segmentation, part-of-speech tagging, NER, syntactic analysis, anaphora resolution, semantic parser, etc. These tasks generally do not solve the actual needs of the application, and most of them are used as intermediate stages for those tasks that solve actual needs Or the auxiliary stage exists. Since the emergence of Bert/GPT, there is actually no need to do these intermediate tasks, because through the pre-training of a large amount of data, Bert/GPT has absorbed these intermediate tasks as linguistic features into the parameters of Transformer. At this time, we can completely Those final tasks are directly solved end-to-end without having to specifically model such intermediate processes. There is a similar development process from statistical machine translation to neural network machine translation.
Limitations and Weaknesses Analysis: Here are some limitations of the different channels: Metric flaws: Its reward model is designed around human supervision, which can lead to over-optimization, which affects performance, and it also has the difficulty of how to determine the metrics. Just like the Bleu value of machine translation, it has been complained about, but no better and more convenient evaluation method can be found. Inability to rewrite the model's beliefs in real time: When a model expresses a belief about something, even if the belief is wrong, it is difficult to correct it. This is just like a stubborn old man. Knowledge is not updated in real time: the internal knowledge of the model stays in 2021, and news after 2022 is not included. This is also mentioned at the experience level. Single mode: ChatGPT is currently good at NLP and Code tasks. As an important seed player leading to AGI, it integrates image, video, audio and other images and multimodality into LLM, and even AI for Science, robot control, etc. It is the only way for LLM to lead to AGI. And this direction has just begun, so it has high research value. High cost: Because of the large scale of the model, the training cost of the super large model is too high, so few institutions have the ability to do this.
Conclusion, some unstructured thoughts
- As V God, the founder of Ethereum, said, an era of post-editing has arrived. AI pre-programmed, pre-drafted content, humans to modify. In fact, in the field of translation, this revolution has already begun, and a project (www.languagex.com) I am working on is in this direction. BTW, with LanguageX, you can use 16 mainstream translation engines in the world to translate, including chatGPT, welcome to try (below). If we as a common society undertake this revolution responsibly (the AI revolution), it will generate enough wealth for everyone to get what they need. —Sam forgets that "human needs" are a bottomless pit. However, the inspiration of this sentence is: the AI revolution will greatly enhance social productivity and create huge wealth. If a project that is in line with human values and focuses on safety is ahead of us to achieve general artificial intelligence, we promise to stop the competition and instead assist this project-only good intentions that are pure and determined to serve people can produce good Only with a vision can we be open enough to attract top-level minds to strive for and worthy of top-level capital investment. Diversity, not monopoly. Although diversity or horse racing consumes social resources, it is always the safest way. If Microsoft completely controls OpenAI, I hope there will be an AI organization that can contain and counter it, such as DeepMind, or others.Most of the so-called "unique" problems in a certain field are most likely just an external appearance caused by lack of domain knowledge. As long as there is enough domain knowledge, this so-called field-specific problem can be solved well. In fact, It is not necessary to focus on a specific field of problems and think hard to come up with a dedicated solution. Perhaps the truth of AGI is surprisingly simple: you just need to give more data in this field to LLM, and let it learn more knowledge by itself. * Almost all of ChatGPT's most amazing skills involve creative fields, such as writing, programming, and translation. It now appears that the jobs most likely to be replaced by AI include creative jobs. It turns out that the difficulty in the eyes of AI and the difficulty in our eyes are not the same dimension at all. AI also allows us to know ourselves better, forcing us to think about the nature of some things, such as what is consciousness? What is emotion? What is creation? There is nothing new under the sun. Is our so-called "innovation" largely a kind of inheritance (knowledge learning) and reorganization (content generation)? * AI will also allow us to examine, what is irreplaceable for human beings? What is the lower level of human beings? What are the more precious and unique human qualities? What should be outsourced to AI? What should human beings spend their time and lives on?
(LanguageX's multi-machine translation engine array)
Forecast: If the reminder reaches a certain threshold, we will organize a ChatGPT from product and business perspective~
One more thing, benefits:
1. There are 5 AIGC reports that are worth studying, and you can download them by replying to "chatgpt" in the background of this official account;
Attachment: Explanation of terms
AIGC: AI Generated Content, artificial intelligence automatically generates content NLP: Natural Language Processing, natural language processing LLM: Large language model, large language model AGI: Artificial general intelligence, general artificial intelligence Prompt: prompt word Fine-tuning: model tuning ML : Machine Learning, machine learning DL: Deep Learning, deep learning GPU: Graphics Processing Unit, graphics card for deep learning BERT: Bidirectional Encoder Representations from Transformers", bidirectional encoder representation RLHF: Reinforcement Learning from Human Feedback, reinforcement based on human feedback study
Some more informative further reading:
OpenAI Charter: https://openai.com/charter/ Sequoia Capital: Generative AI, a Creative New World https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/2022 The 32 best AI papers in 2022 https://hub.baai.ac.cn/view/227.98 Dismantling the GTP technology roadmap https://mp.weixin.qq.com/s/7N3HveaIfn2N-zKjBoRL1A Wu Enda’s end of 2022 Inventory: Generative AI, ViT, Large Model https://mp.weixin.qq.com/s/nagtjtYD98OlJlyddt78Aw This article takes you to understand generative AI: https://mp.weixin.qq.com/s/ZE-nyGnCx -bLXwf2rhraTA Chen Wei Tan Xin: ChatGPT features, principles, technical architecture and industry future https://zhuanlan.zhihu.com/p/590655677 OpenAI CEO Sam Altman: AI will be the new basic platform after the mobile Internet https://mp. weixin.qq.com/s/hwfk1j33uLsbiDUA89p9vA's explosive chatGPT, and its past and present: https://m.huxiu.com/article/733716.html About Microsoft and OpenAI, and those things about GPT Generative AI is here: How tools like ChatGPT could change your business ChatGTP-based project: https://www.zhihu.com/question/570189639/answer/2793888150 The road to AGI: the essence of large language model (LLM) technology https://zhuanlan .zhihu.com/p/597586623