Artificial intelligence (AI) is admired by some people passionately, while others fear it like a tiger. The development of AI has entered the high-level fast lane. Technology updates and iterations are changing day by day. Technology is neither sinful nor good or evil. Human-computer communication in the future of sci-fi is now symbiosis and coexistence. We already have the outline consciousness in our minds. Technology is the primary productive force, and no one can be separated from technology.
The latest chat AI-ChatGPT exploded the enthusiasm of technology geeks and AI exploration enthusiasts as soon as it was launched. In less than a week, more than 1 million users were added. They wrote codes, document needs, emotional therapy and weird theory in chatting with AI. Many people fell into the cycle of ChatGPT's crazy answers. Even Elon Musk, one of its founders, couldn't help but sigh:We're not far from powerful and dangerous artificial intelligence。
secondary title
ChatGPT development history
Generative Pre-trained Transformer (GPT), is a deep learning model for text generation trained on data available on the Internet. It is used for question answering, text summarization, machine translation, classification, code generation, and conversational AI.
In 2018, GPT-1 was born, this year is also the first year of pre-training models for NLP (Natural Language Processing). In terms of performance, GPT-1 has a certain generalization ability and can be used in NLP tasks that have nothing to do with supervisory tasks.
Although GPT-1 has some effect on untuned tasks, its generalization ability is much lower than that of fine-tuned supervised tasks, so GPT-1 can only be regarded as a fairly good language understanding tool rather than a conversational one. AI.
2019 GPT-2 releaseHowever, GPT-2 did not carry out too many structural innovations and designs on the original network, but only used more network parameters and larger data sets: the largest model has a total of 48 layers, and the number of parameters reaches 1.5 billion. The goal is to use the unsupervised pre-training model to do supervised tasks.
In terms of performance, in addition to comprehension, GPT-2 has shown a strong talent for the first time in terms of generation: reading summaries, chatting, continuing writing, making up stories, and even generating fake news, phishing emails, or role-playing online No problem. After "getting bigger", GPT-2 did show a general and powerful ability, and achieved the best performance at that time on multiple specific language modeling tasks.
In May 2020, OpenAI released GPT-3, this model contains two orders of magnitude more parameters than GPT-2 (175 billion vs 1.5 billion parameters), which is a huge improvement over GPT-2.
GPT-3 has achieved strong performance on many NLP datasets, including translation, question answering, and cloze tasks, as well as some tasks that require on-the-fly reasoning or domain adaptation, such as using a new word in a sentence or performing 3-digit numbers operation. GPT-3 can generate samples of news articles that are difficult for human raters to distinguish.
In early 2022, OpenAI released InstructGPT, a new version of GPT-3 fine-tuned to minimize harmful, inauthentic, and biased output. InstructGPT changed its name to Chatgpt after launch.
There will be more powerful GPT-4 in the futuresecondary title
Analysis of ChatGPT hot phenomenon
According to OpenAI officials, ChatGPT was created and trained with the help of humans, and human trainers ranked and rated the way early versions of the AI answered queries. That information is then fed back into the system, which adjusts the answers based on the trainer's preferences—a standard way of training artificial intelligence known as reinforcement learning.
To create a reward model for reinforcement learning, OpenAI also needs to collect comparison data, which consists of two or more model responses ranked by quality.
To gather this data, OpenAI collected conversations between the AI trainer and the chatbot, randomly selected a message written by the model, spot-checked several candidate responses, and asked the AI trainer to rank the responses.
Furthermore, using these reward models, the study fine-tuned the model using an approximate policy optimization algorithm and performed multiple iterations of the process.
The popularity of ChatGPT highlights the subversion and convenience of our traditional knowledge acquisition channel. In the past, the dissemination of knowledge was based on centralized passive acquisition. The process was boring and the fluctuations were too large. The learning chain:Database + AI screening + user needs.
secondary title
What are the limitations of ChatGPT
Users who have experienced ChatGPT generally report that the current AI chat feedback content information is inaccurate. Sometimes it will give seemingly correct but absurd answers, fine-tune the way of asking questions will get completely different answers, sometimes use certain sentences repeatedly, and even provide A mess of information, where someone with experience and expertise can see mistakes at a glance.
Specifically, these limitations are:
During the reinforcement learning (RL) phase of training, there is no specific source of truth and canonical answers to your questions, to answer your questions.
Train the model to be more cautious and may refuse to answer (to avoid false positives for hints)
secondary title
Reflections on ChatGPT for Web3
ChatGPT has partially surpassed today's search engines in terms of finding answers and solving problems. ChatGPT may change the way we obtain information and output content in the future. What are the needs of Web3 for ChatGPT?
Shallow needs
Project community search: Point-to-base precise demand screening can improve the selection efficiency of track classification, and focus on hot spots and familiar areas in energy allocation.
Ranking price tracking: rankings and prices have fluctuating values and major changes at different stages of the market. Due to personal needs and usage habits, it is impossible to follow the data ocean of various exchanges and platforms. Accurate content output increases users' stickiness to data.
Knowledge sharing: The hurdles in the knowledge stage generally require personal diligence and enthusiasm to break through. The classification of different communities and content leads to a sense of heaviness for new groups of users. Pioneers cannot get away from imparting experience efficiently. Accurate content output can change traditional knowledge acquisition. channel.
AMA Q&A: The project roadmap, white paper, and the content of key person activities can be quickly grasped in the AMA Q&A format. The development of projects and communities can be aimed at more different user groups, and it can serve as AI customer service in the dissemination of community brand IP Function.
Demand Guidance Forecast: The results of data comparison and screening far exceed personal emotional fluctuations. Data can clarify the content of personal needs, increase or decrease asset books at different stages, predict market environment and future trends, and obviously rich data can better grasp decisions.
deep thinking
Web3 content production: Stable and accurate positive content output is currently urgently needed by Web3. In the grand world of Web3, it is relatively thin and slow to rely only on professional content producers to provide service content. AI will increase the rate and stability of content output. Far beyond the individual, whether it is content quality, screen presentation and communication efficiency.
Smart contract deployment and security audit: AI can provide professional development reference templates and verification checks by using the integrity of the database in the construction of the code. Of course, in the development of smart contracts, the program output of AI cannot be completely relied on. After all, independent projects However, security audit is a process of repeated detection and finding bug fixes. AI can remind and correct errors at any time in the comparison and screening of utilization data.
Virtual machine upgrade: Web3 uses blockchain technology, and the closed nature of blockchain cannot reflect external information in a timely manner. The emergence of virtual machines solves the gap in the transmission of data outside the chain, but the accuracy and generation of data provided by virtual machines This is where AI may change the way it works when its trust is not properly validated.
The flow and accumulation of talents: Web3 is a comprehensive upgrade of the next generation of technology, which requires extensive and professional blockchain talents, but now talents are generally deposited in traditional industries. In terms of recruitment and communication, often due to unclear expressions and traditional thinking The prohibition prevents talents from stepping in easily, and newcomers who turn to enter the industry are troubled by the impact of information in the industry. The rapid iteration of new technologies and project content requires professional teaching, and AI plays the role of professional mentor and book treasure.
Summarize
Summarize
Article Source
"Statement: According to the "Notice on Further Preventing and Dealing with the Risk of Hype in Virtual Currency Transactions" issued by the central bank and other departments, the content of this article is only for information sharing, and does not promote or endorse any business and investment behavior. Readers are requested to strictly abide by the laws of the region Regulations, do not participate in any illegal financial behavior』
IOS
Android