Understanding Generative Artificial Intelligence

Generative Artificial Intelligence is a branch of artificial intelligence that generates content such as text, images, sound, video, and code based on algorithms, models, and rules. This technology can generate content that has a certain degree of logic and coherence, relying on pre-trained multimodal foundational large models, based on user input and relevant information. Unlike traditional artificial intelligence, generative artificial intelligence can not only process input data but also learn and simulate the intrinsic laws of things, autonomously creating new content.

——Excerpted from “Study Guide on the of the 20th Central Committee’s Third Plenary Session”

Improve the comprehensive governance system of the internet. Deepen the reform of the internet management system, integrate the functions of internet content construction and management, and promote the integrated management of news publicity and online public opinion. Improve the development and management mechanism of generative artificial intelligence. Strengthen the rule of law construction in cyberspace, improve the long-term mechanism for ecological governance of the internet, and establish a system for the protection of minors online.

Artificial intelligence can be classified based on foundational models, mainly including decision-based AI and generative AI. Decision-based AI primarily learns the conditional probability distribution in data, that is, the probability that a set of samples belongs to a specific category, allowing machines to judge, analyze, and predict new scenarios. Generative AI mainly learns the joint probability distribution in data, meaning the vector composed of multiple variables presents a certain probability distribution, enabling machines to summarize existing data and create imitative and stitched content using deep learning technologies.

Generative artificial intelligence is also known as AIGC (AI Generated Content), and the forms of content it can generate are quite diverse, including text, images, audio, and video. Its operational principle mainly relies on feeding vast amounts of data to train deep neural networks. In 2014, Ian Goodfellow left his job at OpenAI and returned to Google, where he formed a research group specifically to explore “generative models,” aimed at creating systems that could creatively generate elements such as images, photos, and audio from the real world. This led to the birth of Generative Adversarial Networks (GANs), marking the opening of the innovative path towards artificial intelligence. GANs generate new data by training two neural networks: one is the generator, which creates new data based on input random noise; the other is the discriminator, which determines whether a piece of data is real or comes from the original dataset. Through training these two networks, the generator learns how to produce real data, while the discriminator learns to distinguish between real and generated data. The training characteristic of GANs is adversarial game-playing, and through this continuous training and evaluation, generative artificial intelligence can produce natural forms of output content, such as text, images, and sound.

Data is the fundamental nutrient for generative artificial intelligence, and its intelligent applications primarily rely on generative systems based on large databases and corpora. The value highlight of generative artificial intelligence lies not only in its application itself but also in its ability to analyze vast amounts of data for machine learning; it can extract data from a static database and corpus and establish a lasting interactive relationship with the subject of questioning and conversation. Each interaction with the conversational subject promotes machine learning and growth. Generative artificial intelligence is not about how intelligent its applications are but rather its capacity to absorb enough databases and corpora; it masters not individual information but panoramic information or long continuous texts; it learns not individual numbers and sentences but the overall context of the system.

Algorithms are the logical structure of generative artificial intelligence. Artificial intelligence essentially employs neural network algorithms, breaking problems down into numerous nodes (neurons) that automatically interconnect and form. Based on the neural network architecture of generative pre-trained transformers, machine learning networks are adept at utilizing the generative principles of basic neuron links, learning and imitating any object they converse with, and reflecting this in their expressions through language. The most fundamental principle is a type known as the “Naïve Bayes algorithm,” which can discover subtle correlations within vast continuous corpora and databases. The more interactions it has with generative artificial intelligence, the more complex the requests made, which is beneficial for the generative capabilities of neural network algorithms.

Computing power is the driving force behind generative artificial intelligence. “Computing power determines algorithms,” and the explosive development of generative artificial intelligence, besides being based on neural network machine learning algorithms, is more importantly driven by enhanced computing power. From the current industry development trends, the parameter count of large models is increasing exponentially, the demand for computing power is surging, and the iteration cycles are accelerating, showing distinctive characteristics of “trillions of parameters, thousands of card clusters, and multimodality.” Looking ahead to future industrial development, exploring innovative brain-like computing, quantum computing, and other high-efficiency computing power technologies opens up new paths for enhancing computing power. For instance, brain-like computing chips that mimic the structure of neurons not only build hardware but also employ new technologies such as spiking neural networks to achieve real-time online learning and even integrate multiple different modalities.

Artificial intelligence has evolved from machine automation, primarily as a technological extension of human brain organs, which is what Marx referred to as creating “organs of the human mind.” The technological route of generative artificial intelligence shapes a new type of relationship of “human-machine symbiosis.” Compared to traditional artificial intelligence, generative artificial intelligence is not only a “complex mega-system” that is “self-updating, self-improving, and self-evolving” but also forms a “request-response” relationship of interaction between humans and technology. At this point, machines no longer simply and passively execute human commands but form a “mutually cooperative and enhancing mode” with humans. In this new type of relationship, human-machine interaction shifts “from machine-centered to human-centered,” “breaking the boundaries between humans, machines, and information resources, and reshaping the paradigm of generating and using information resources.” The development of generative artificial intelligence may serve as an opportunity to build a human destiny community based on intelligent technology.

However, in the face of the potential risks of generative artificial intelligence, such as intellectual property disputes, data privacy threats, ethical use challenges, continuation of technical biases, and structural unemployment issues, we must adhere to guiding principles based on laws and regulations, supporting industry organizations, enterprises, education and research institutions, public cultural institutions, and relevant professional institutions to collaborate on technological innovation, data resource construction, transformation application, and risk prevention in generative artificial intelligence. Departments such as cybersecurity, development and reform, education, science and technology, industry and information technology, public security, broadcasting and television, and news publishing should strengthen the management of generative artificial intelligence services according to their respective responsibilities and legal frameworks. National authorities should improve scientific regulatory methods that are compatible with innovative development based on the characteristics of generative artificial intelligence technology and its service applications in relevant industries and fields, formulating corresponding classification and grading regulatory rules or guidelines.