Knowledge related to the big model of the generative artificial intelligence
Generative artificial intelligence large model was the product of the combination of "big data + big computing power + strong algorithms". It was the "hidden knowledge base" that condensed the essence of big data. It contained two meanings of "pre-training" and "large model". That was, the model could directly support various applications without fine-tuning after completing the pre-training on large-scale data sets, or only a small amount of data fine-tuning. From the development process, 1950 - 2005 was the embryonic period, which was the traditional neural network model stage represented by the CPU;2006 - 2019 was the settling period, which was the new neural network model stage represented by the Transformer;2020 - 2023 was the explosive period, which was the pre-training large model stage represented by GPM; From January 2024, its application accelerated. In terms of the filing of large models, according to article 17 of the "Temporary Methods for the Management of Generative Artificial Intelligence Services," the developers of generative artificial intelligence services that provided services to the domestic market and had the ability to mobilize public opinion or society needed to be filed. The relevant information for filing is as follows: - Development background: On December 1,2017, the "Double-New Assessment" proposed in the Regulations on the Administration of Security Assessment of New Technologies and New applications of Internet News Information Service was the predecessor of the big model filing; On November 30,2018, the relevant regulations defined the assessment object as the main body of Internet information service with "public opinion attribute or social mobilization ability"; After the implementation of the "Temporary Methods for the Management of Generative Artificial Intelligence Services" on August 15,2023, relevant developers began to prepare for the filing; On April 2, 2024, the Internet Information Technology Office released a list of 117 large models that had been filed, and the filing of large models entered the normal stage. - Materials required for filing: Large model online filing form (must include basic information about the model, development process, service content, safety precautions, safety assessment results, voluntary commitment, etc.), language annotation rules (must introduce the qualification of the annotation team, annotation rules, annotation process, etc.), list of blocked keywords (at least 10000 keywords, must cover a variety of safety risks and be updated regularly). The filing situation of different companies was different: - The enterprises that need to be filed: the enterprises that are notified by the Internet Information Administration Office to do a large model filing/safety assessment, the enterprises that have reached a certain scale, and the enterprises that have the strength or are willing to do a large model filing. - Enterprise that doesn't need to be registered: Generative artificial intelligence services that don't have public opinion attributes or social mobilization capabilities; those that call the registered large model API interface and serve the domestic public can be registered; enterprises/educational and scientific research institutions/industry organizations/public cultural institutions, etc., whose services are not provided to the domestic public. The big model filing was a branch of the algorithm filing. The algorithm filing was a general concept. In addition to generating and synthesizing, it also included other types of algorithm products such as customized push. The two were different in terms of filing type, materials, approval, and so on. For example, the bean bag model had a strong AI drawing ability. It could generate high-quality AI paintings of various styles, generate pictures according to words, generate videos according to pictures, etc. Its daily token usage exceeded 500 billion, and it was widely used in more than 50 businesses within ByteDance. It was opened to enterprises through the volcano engine. The bean bag APP ranked first among AIGC applications, with more than 26 million monthly active users. There was also the "Xiaoke" big model released by the Taiji of the Electric Technology Department, which was targeted at the party, government, and enterprise users. It had the characteristics of "industry, specialization, domestication, and privatization". It provided a new industry application model of "general intelligent model big cycle + industry intelligent model small cycle" and "model training + test evaluation + scene fine adjustment + credibility enhancement". It had launched intelligent applications such as text assistant, coding assistant, and intelligent plotting. It was used in smart government affairs, smart manufacturing, Special industries and other fields provide special services. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
Generative artificial intelligence
Generative artificial intelligence was an artificial intelligence that could generate text, images, or other media information based on prompts. Its principle was to use machine learning technology to generate new text, program code, images, videos, sounds, and other data based on existing large-scale multi-mode data sets. It had the ability to handle a variety of tasks and scenarios. In the early days of AIGC (artificial intelligence generated content), in 1957, Shearer and Isaac created the first music composition by computer,"Ilyak Suite", by converting the control variables in the computer program into musical notes. Since the 1990s, AIGC has been developing from experimental to practical. In 2007, New York University's Ross Goodwin assembled an artificial intelligence system to create the world's first novel,"1 The Road," written entirely by artificial intelligence. After 2014, AIGC entered a new era with the development of deep learning algorithms, especially the Generative Adversant Network (GAN). The release of works such as DALL-E and ChatGPM showed that AIGC had made significant breakthroughs in generating content. In terms of applications, it would take the lead in media, e-commerce, film and television, entertainment and other industries with high digital levels and rich content requirements. Its functions included text generation (such as generating coherent text passages, continuing stories, or answering questions based on prompt words), machine translation (based on a large amount of language data learning, translation is more natural and smooth), text summary (extracting key information from a large amount of text to generate a concise summary), creative writing (generating stories, poems, advertising copywriting, etc.), and so on. On December 26, 2023, Generative Artificial Intelligence was selected as one of the top ten scientific terms of 2023. On July 3, 2024, the World intellectual property organization released the Generative Artificial Intelligence patent situation report. From 2014 to 2023, China's number of patent applications for Generative A1 was the highest in the world. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
Generative Artificial Intelligence Technology
Generative artificial intelligence was an important branch of artificial intelligence. It could generate text, images, or other media information based on prompts. The principle was to use machine learning technology to generate new data based on existing large-scale multi-mode data sets, such as text, program code, images, videos, and sounds, so that it had the ability to handle a variety of tasks and scenarios. In the early years (1950s-1990s), due to the limitations of science and technology, it was only in the small-scale experimental stage. In 1957, Hiller and Isaac converted the control variables in the computer program into musical notes and created the first music piece composed by a computer,"Ilyak Suit." In 1966, Weizenbaum and Colby developed the world's first human-machine conversation robot,"Eliza." After the 1990s, AIGC evolved from experimental to practical. In 2007, Ross Goodwin's artificial intelligence system created the world's first novel, 1 The Road, written entirely by artificial intelligence. Since 2014, with the development of deep learning algorithms, especially the proposal and repetition of Generative Adversant Network (GAN), AIGC entered a new era. The release of DALL-E and ChatGPM marked a significant breakthrough in generating content. In terms of application, it would take the lead in media, e-commerce, film and television, entertainment and other industries with a high degree of digitizing and rich content demand. It had the production capacity and knowledge level that surpassed humans. It could undertake mechanical labor such as information mining, material transfer, copying, editing, and so on. It could meet large-scale individual needs at low cost and high efficiency. In addition, on December 26, 2023, the Generative Artificial Intelligence was selected as one of the "Top Ten Science and Technology Terminology of 2023". On July 3, 2024, the World Intelligent Property Organization released the "Generative Artificial Intelligence Patents Situation Report", which showed that China's number of patent applications for Generative A1 was the highest in the world from 2014 to 2023. Moreover, on May 24, 2024, the Ministry of Human Resources and social protection announced the new profession of Generative Artificial Intelligence System Practitioner. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
Generative Artificial Intelligence Images
Generative artificial intelligence could generate images. AI image generation systems such as Midjourney and stable dispersion could create extremely realistic images. Through learning and analyzing massive amounts of images, they could piece together new photos that contained unseen elements. Generative artificial intelligence pictures were very realistic and could even fool professionals. For example, some people abroad used artificial intelligence systems to generate photos of former President Trump being arrested and photos of the Pope wearing fashionable white cotton clothes and went viral on the Internet. At present, it is difficult for the human eye to distinguish between images generated by artificial intelligence and real photos, but some details can provide clues, such as the early AI drawing too many fingers when restoring human hands (but the new generation of systems has solved this problem in some photos). At the same time, with the development of science and technology, there were also image detection systems (deep forgery detection) that could determine the authenticity of images by analyzing the subtle differences in photos. In some preliminary studies, image detection AI could reach or even exceed the accuracy of humans. The domestic AI smart bean buns also had the function of generating pictures and paintings. They could generate all kinds of pictures according to the descriptions entered by the user. They could be applied to many aspects of life, work, and learning, such as providing inspiration for designers, generating children's paintings for children, making avatars, generating home improvement renderings, new media accompanying pictures, background pictures, etc., and were completely free. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
Generative artificial intelligence example
Generative artificial intelligence included ChatGPM, DALL-E, Stable Diffusing, Midjourney, and so on. In addition, in 2007, Ross Goodwin of New York University assembled an artificial intelligence system to create the world's first novel," 1 The Road," which was completely written by artificial intelligence. The artificial intelligence system used in the novel was also a type of generative artificial intelligence. " A Short History of the Future: Legends of the Intelligent Era " was equally exciting. Everyone was welcome to click and read it!
Generative Artificial Intelligence Technology
Generative artificial intelligence was an important branch of artificial intelligence. It could generate text, images, or other media information based on prompts. The principle was to use machine learning technology to generate new text, program code, images, videos, and sound data based on existing large-scale multi-mode data sets, so as to have the ability to handle a variety of tasks and scenarios. In the early years (1950 - 1990s), small-scale experiments were carried out due to the limitations of science and technology. In 1957, there were music pieces created by computers. After the 1990s, it evolved from experimental to practical. In 2007, novels created by artificial intelligence appeared. Since 2014, with the development of deep learning algorithms, especially the proposal and repetition of the Generative Adversant Network, it entered a new era. The results such as DALL-E and ChatGPM marked a significant breakthrough in generating content. It had a variety of functions, including text generation (generating coherent text passages, stories, or answers to questions based on hints), machine translation (translating one language into another language more naturally and fluently), text summary (extracting key information from a large amount of text to generate a concise summary), creative writing (generating stories, poems, advertising copywriting, etc. based on hints), and so on. In terms of application, it would take the lead in media, e-commerce, film and television, entertainment and other industries with a high degree of digitizing and rich content demand. In 2024, the profession of a generative artificial intelligence system application worker appeared. However, it also had technical limitations and ethical risks. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
What is Generative Artificial Intelligence?
Generative artificial intelligence was an artificial intelligence that could generate text, images, or other media information based on prompts. Commonly seen AI include ChatGPM, DALL-E, Stable Diffusions, Midjourney, and so on. " A Short History of the Future: Legends of the Intelligent Era " was equally exciting. Everyone was welcome to click and read it!
Generative Artificial Intelligence Technology
Generative artificial intelligence was an important branch of artificial intelligence. It could generate text, images, or other media information based on prompts. Its principle was to use machine learning technology to generate new text, program code, images, videos, sounds, and other data based on existing large-scale multi-mode data sets. It had the ability to handle a variety of tasks and scenarios. In the early stages of AIGC (artificial intelligence generated content), in 1957, Hiller and Isaac created the first music composition created by a computer by converting the control variables in a computer program into musical notes. After the 1990s, AIGC evolved from experimental to practical. In 2007, Ross Goodwin of New York University assembled an artificial intelligence system to create the world's first novel written entirely by artificial intelligence. Since 2014, with the development of deep learning algorithms, especially the proposal and repetition of Generative Adversant Network (GAN), AIGC entered a new era. The release of DALL-E and ChatGPM marked a significant breakthrough in AIGC's generation of content. Its functions included text generation (generating coherent text passages, continuing stories, or answering questions based on prompt words), machine translation (based on a large amount of language data learning, translation is more natural and fluent and can understand the context), text summary (extracting key information from a large amount of text to generate a concise summary), creative writing (generating stories, poems, advertising copywriting, etc. based on prompt words), and so on. In terms of application, it would take the lead in media, e-commerce, film and television, entertainment and other industries with a high degree of digitizing and rich content demand. At the same time, on December 26th, 2023, Generative Artificial Intelligence was selected as one of the top ten scientific and technological terms of 2023. From 2014 to 2023, China ranked first in the world in the number of patent applications for Generative A1. In 2024, a new profession, a generative artificial intelligence system application worker, appeared. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!
What is Generative Artificial Intelligence?
Generative artificial intelligence was an artificial intelligence that could generate text, images, or other media information based on prompts. Its principle was to use machine learning technology to generate new text, program code, images, video, and sound data based on existing large-scale multi-mode data sets. It had the ability to handle a variety of tasks and scenarios. " A Short History of the Future: Legends of the Intelligent Era " was equally exciting. Everyone was welcome to click and read it!
The Rise of Generative Artificial Intelligence
The rise of Generative Artificial Intelligence was a gradual process. In the early stages, in 1957, Hiller and Isaac created the first music composition created by a computer by converting the control variables in a computer program into musical notes. This was an early exploration. After the 1990s, artificial intelligence generated content (AIGC) evolved from experimental to practical. In 2007, the artificial intelligence system installed at New York University created the first novel completely written by artificial intelligence. Since 2014, AIGC had entered a new era with the development of deep learning algorithms, especially the Generative Adversant Network (GAN). The release of iconic results such as DALL-E and ChatGPM showed that AIGC had made significant breakthroughs in content generation. In terms of technical principles, it uses machine learning technology to generate new data based on large-scale multi-mode data sets, including text, images, videos, sounds, etc., so it has the ability to handle a variety of tasks and scenarios. In terms of applications, it would take the lead in media, e-commerce, film and television, entertainment and other industries with high digital levels and rich content requirements. On December 26, 2023, Generative Artificial Intelligence was selected as one of the top ten scientific and technological terms of 2023. On July 3, 2024, a report showed that China's number of A1 patent applications from 2014 to 2023 was the highest in the world. These all reflected its increasing importance. On May 24, 2024, the Ministry of Human Resources and social protection announced that the application of the generative artificial intelligence system was a new profession, which also reflected the development of the field to promote the emergence and recognition of related professions.
Artificial intelligence related knowledge Q & A
Here are some common questions and answers about artificial intelligence: ** 1. Basic Knowledge ** 1. ** Mathematics Basics ** - In artificial intelligence, linear algebra was used to deal with matrix and matrix operations. For example, the weight matrix operation in neural networks relied on the knowledge of linear algebra. - The theory of probability and statistics is very important for understanding the distribution of data, uncertainty, and model evaluation. For example, in Bayes inference, probability theory was the core foundation. When evaluating the performance of the model, statistics methods such as cross-verification also relied on probability theory and statistics knowledge. - [Derivative: It plays a key role in optimization algorithms. For example, in the back-transmission algorithm of neural networks, it is necessary to use calculative to calculate the loss function's slope with respect to the model parameters, thereby updating the parameters to minimize the loss.] - optimization theory: It helps to find the optimal parameters of the model. For example, in the Support Vector-Machine, the hyperplane that can maximize the classification gap can be found through optimization theory. 2. ** Basic programming ** - Python was one of the most commonly used programming languages in the field of artificial intelligence. It has a rich library, such as NumPy for numerical calculations, Panda for data processing, Scikit-learn for machine learning algorithm implementation, etc., and the grammar is simple and easy to learn and use. - C++/Java: Used when high-performance computing and large-scale system development are required. For example, when developing the underlying code of some deep learning framework that required extremely high running speed, the efficiency advantage of C++ would be reflected. In the development of enterprise artificial intelligence applications, Java was also widely used because of its stability and cross-platform nature. 3. ** Data structure and algorithm ** - Basic data structure: - Array: It is a continuous storage data structure, which is more efficient in storing and accessing data. For example, in the storage of image data, the value of the image is stored in an array. - [Linked-list: It is suitable for dynamic data structures. The insert and delete operations are more flexible than arrays. It is used in scenarios where frequent insert and delete operations are required.] - Tree: For example, decision trees are used in machine learning for classification and regressions. The tree structure can directly represent the hierarchy of data and the decision-making process. - Diagram: It is used in graph algorithms. For example, in Social networks analysis, the relationship between users can be represented by a graph. Graph algorithms such as the shortest path algorithm can be used to analyze the network structure. - Basic algorithm: - Sorting: For example, quick sort, merge sort, etc. Sorting the data in the data pre-processing stage helps to improve the efficiency of search and other algorithms. - Search: Search algorithms such as dichotomous search can quickly locate data, which is very useful when looking for specific data in large-scale data sets. - Graph algorithm: In addition to the shortest path algorithm mentioned above, there is also the minimum spanning tree algorithm, which is widely used in network analysis, path planning, and other fields. ** 2. Machine Learning Basics ** 1. ** Supervising Learning ** - "Liner Regression: It is used to establish a linear relationship between independent variables and dependent variables. For example, it is used to predict the relationship between house prices and factors such as the size of houses and the number of rooms. - [logistic regressions: mainly used for two-class classification problems, such as predicting whether an email is junk mail. The classification result is obtained by combining the input features in a linear manner and then using a logical function (such as sigtoid function).] - [Decision tree: Construct a decision tree structure by dividing the features. Each internal node is a feature test, and the leaf nodes are the category or numerical prediction results. It has the characteristics of strong interpretation.] - The Support Vector-Machine (SSS) was used to find a hyperplane that could maximize the classification interval to classify the data. It performed well in small samples and high-dimensional data. - Random Forest: It is the integration of multiple decision trees. By integrating the results of multiple decision trees (such as voting), the accuracy and stability of the model are improved. - Grade Boosted Tree (GMDT): It improves the performance of a weak learner (usually a decision tree) through successive iterations. Each step constructs a new decision tree based on the residual error of the previous step, which has a strong ability to fit the data. 2. ** Unsupervised Learning ** - K-means Cluster: Divide the data into K clusters, so that the similarity of the data points within the cluster is high and the similarity of the data points between the clusters is low. It is often used for initial exploration and classification of data. - Principal component analysis (Principal component analysis): It is used for data reduction and feature extraction by projecting the data into a low-dimensional space while retaining the data's variation information as much as possible. - Self-Encoders: A neural network structure used for unsupervised feature learning and data coding and decoding. It has applications in data de-noising and generating new data. 3. ** Semi-supervised learning ** - Self-training: Use a small amount of labeled data and a large amount of unlabeled data to learn. First, train an initial model on the labeled data, then use this model to predict the unlabeled data. The prediction results with high confidence will be added to the training set as new labeled data, and the model will be continuously iterated. - Multi-perspective learning: Using multiple perspectives (different feature representation) of the data to learn can improve the performance of the model in some cases. 4. ** Reinforcement Learning ** - Q-learning: It is a reinforcement learning algorithm based on the value function. It chooses the optimal action by estimating the action-value function (Q-function). It is widely used in the markov decision process. - DQN (Deep Q-Network): Combining deep learning and Q-learning to deal with reinforcement learning problems in high-dimensional state spaces, such as playing Atari games. - <<Policygradients>>: Directly optimized the policy function. By adjusting the parameters of the policy function to maximize the cumulative reward, it has a better performance in reinforcement learning tasks in the continuous action space.> ** 3. Deep Learning Basics ** 1. ** neural network ** - Feed forward neural network: It is the most basic neural network structure. Information is transmitted from the input layer through the hidden layer to the output layer in one direction. It is often used to solve various classification and return problems. - Consecutive neural networks (CCN): Specially used to process data with a grid structure, such as image and audio data. It automatically extracted the features of the data through structures such as the convolutive layer and the pool layer, and achieved good results in the fields of image recognition and target detection. - Cyclic neural network (RHN): suitable for processing sequence data, such as time series data and natural language text. There were loop connections between its neurons, which could handle long-term dependence in data, but there was a problem of vanishing or exploding of slopes. - Long Short Time Memory Network (LStellar Memory Network): It is a special type of neural network. By introducing a gating mechanism, it can solve the problem of vanishing of the grads in the neural network and better deal with long sequence data. - Transform: Completely built on the attention mechanism, it has achieved great success in the field of natural language processing, such as machine translation, text generation, and other tasks. 2. ** Deep learning framework ** - TensorFlow: It was developed by Google and is highly flexible and Scalable. It supports a variety of computing devices such as CPU and CPU, and can be used to develop various deep learning models. - PyTorch: It was developed by Facebook. It was popular among researchers and developers for its dynamic computational graphs and simple code style, which made it easy to quickly develop and debuff models. - Keras: A high-level neural network API that is easy to use and fast to build models. It is suitable for beginners, but it is relatively weak in terms of flexibility. - MXNet: A lightweight, distributed deep learning computing platform that supports multi-machine, multi-node, multi-CPU computing with efficient computing performance. ** 4. Data processing and feature engineering ** 1. ** Data Preprocessing ** - Missing value processing: You can use methods such as deleting samples with missing values and filling in (for example, filling in the missing values of numerical characteristics with the mean and the mean, and filling in the missing values of classification characteristics with the mode). - Outlier processing: You can identify and deal with outlier values through a statistical method (such as the 3-sigma-principle), such as replacing outlier values with reasonable values or directly deleting the samples where outlier values are located. - Data Standardisation and Normalisation: Standardisation (such as converting data to a distribution with a mean of 0 and a standard deviation of 1) and normalisation (such as projecting data to the [0,1] interval) can help improve the convergence speed and performance of the model. 2. ** Characteristic Selection and Extraction ** - Character selection method: - The filtering method is to filter the features based on the relationship between the features and the target variables, such as Pearson's coefficient. - Wrapping Method: Using the performance of the model as an evaluation standard, the best feature combination is selected by continuously adjusting the feature set. - Embedding method: automatically perform feature selection during model training. For example, in Lasso regressions, the L1 regularizer is used to shrink the coefficient of some unimportant features to 0, thereby achieving feature selection. - Method of feature extraction: - Principal component analysis (Principal component analysis): As mentioned earlier, it is mainly used for data dimension reduction and feature extraction. - LDA (linear discriminant analysis): It is a supervised feature extraction method. It is used to extract features by magnifying the distance between classes and minimising the distance within classes. It is often used for classification tasks. - [Autonductor: Unsupervised feature extraction. By learning the data's coding and decoding process, a low-dimensional representation of the data is obtained as a feature.] 3. ** Data Enhancement ** - Image data enhancement: Rotation, flipping, trimming, adding noise, and other operations to increase the variety of data and improve the model's generalization ability. - Text data enhancement: For example, text data can be expanded through operations such as synonym replacement, random sentence insert, and delete. ** 5. Development of artificial intelligence ** 1. ** Historical Development ** - The legend of artificial intelligence could be traced back to ancient Egypt, but the modern meaning of artificial intelligence research began in the mid-20th century. In 1950, Turing published a paper on the Turing Test, which laid the theoretical foundation for the development of artificial intelligence. He was also known as the "father of artificial intelligence." - Early artificial intelligence research was inspired by humans 'understanding of their own intelligence, such as trying to build neural networks by imitating the structure of human brain neurons. 2. ** Future Development ** - According to Ray Kurzweil, the Singularity would occur around 2045. At that time, the era of artificial intelligence would arrive, and humans would develop at an exponential rate, possibly surpassing humans and life itself. - With the continuous development of technology, the application of artificial intelligence in medical, financial, education, retail, manufacturing, transportation and other industries will continue to expand, and new application scenarios will continue to emerge. ** 6. Other aspects of AI ** 1. ** Turing Test ** - The Turing Test was a way to determine whether a machine had intelligence. If the computer could answer a series of questions posed by a human tester within five minutes, and more than 30% of the answers were mistaken for human answers, the computer would pass the test. However, the Turing Test also had some controversy. For example, the Chinese Room Experiment questioned the effectiveness of the Turing Test. 2. ** The artistic creativity of artificial intelligence ** - Although computers could already create Bach-style music and paintings, there was still debate about whether artificial intelligence had true artistic creativity. The works created by artificial intelligence might lack the emotions and inspiration of humans. "A Short History of the Future: Legends of the Intelligent Era" was equally exciting. Everyone was welcome to click and read it!