Opinion | "User usage scenarios" are the key to the development of AI platforms

Original: Yang Jihong

**Source:**AI dark horse

Image source: Generated by Unbounded AI

Introduction to AI Dark Horse 👉

"AI is the bow and arrow of the Neolithic Age." The learning ability formed by AI depends on the scale of parameters. GPT-2 has about 1.5 billion parameters, while the largest model of GPT-3 has 175 billion parameters, an increase of two orders of magnitude. According to media speculation but not yet confirmed news, the parameters of GPT-4 may reach 100 trillion.

This article mainly discusses: the capabilities and construction, opportunities and challenges of the interactive AI platform in the new era. Different from some traditional views that the construction of artificial intelligence is mainly about hardware construction, Yang Jihong, deputy director of the Audiovisual New Media Center of China Central Radio and Television Station, started from the perspective of "people" and emphasized that "people", that is, users, are the decisive battle in the battle of artificial intelligence. Strategic resources in the campaign. She believes that "soft technology" is the "golden track" for technological innovation to overtake on curves and change lanes. Firmly grasp the "winner" of "user usage scenarios".

Talking about the Hard Power and Soft Power of the Interactive AI Platform

01 Preface The only constant is change

Innovation never stops, new technologies are constantly emerging, and technological innovation driven by AI is profoundly changing the way we live and work.

If the popularization of computers, networks, and the Internet is a wave of change brought about by computers, then the overwhelming force of AI is stronger and more comprehensive. Traditional industries and occupations are being impacted and subverted by artificial intelligence technology. Many traditional jobs are being replaced, requiring continuous learning and adaptation to new technologies and new models.

When AI is in this article, I try to propose a new perspective to examine with you the capabilities and construction, opportunities and challenges of the interactive AI platform in the new era. Different from some traditional views that artificial intelligence construction is mainly hardware construction, I try to use hard power and soft power to deconstruct the ever-changing artificial intelligence capabilities from a higher dimension, starting from the perspective of "people", emphasizing "people" That is, users are the strategic resources in the decisive battle against artificial intelligence.

02The “hard power” and “soft power” of interactive AI platforms

1. Hard Power in Interactive AI

1.1 Computer hardware represented by CPU/GPU

Computer hardware is the foundation of AI. First, computer hardware directly determines the ability of AI to handle complex computing tasks. Processing large amounts of data, computing and training deep learning algorithms, etc., requires sufficient computing power. The performance of CPU/GPU directly determines the computing speed and efficiency of AI. Secondly, as a hardware device, CPU/GPU needs to support the AI software operating environment. AI involves a large number of data processing and computing tasks, and requires a special software environment to support its operation. In addition, computer hardware also needs to have high scalability and programmability. There are a wide range of AI application scenarios, and different AI solutions need to be provided for different scenarios and applications. The dynamic allocation and expansion of resources can be quickly realized only if the computer hardware has high scalability. Finally, computer hardware also needs to be programmable so that developers can quickly optimize algorithms and solutions. The following figure shows the basic hardware composition and functions of the AI system platform:

2. "Soft power" is the ballast stone of interactive AI

2.1 Knowledge graph shapes AI basic disk

The knowledge graph of AI is a structured knowledge base, which contains a wide range of domain knowledge, and is used to support the learning and reasoning of artificial intelligence systems. A knowledge graph consists of entities (such as people, places, events, etc.) and the relationships between them. It can include various types of knowledge, including definitions, attributes, categories, associations, etc.

The construction and use of knowledge graphs is an important part of AI technology. It can help AI systems acquire, organize and store knowledge and semantic information in various fields, and improve the cognitive capabilities and intelligence of AI systems. Knowledge graphs can be used in many aspects such as natural language processing, recommendation systems, question and answer systems, and information retrieval.

To give some familiar examples, Google Knowledge Graph, Baidu Encyclopedia, and Wikipedia are all knowledge graphs.

2.2 Algorithm Level Driving Capacity Upgrade

2.2.1 Importance of Algorithms

Algorithms are an important means to achieve various tasks of AI. The design and improvement of algorithms are also key to promoting the development and progress of AI. The importance of algorithms is reflected in the following three aspects:

① Affect the accuracy and efficiency of the model: Different algorithms will affect the accuracy and computational efficiency of the model. Choosing the right algorithm can improve the accuracy of the model and reduce the computation time.

② Meet different needs and scenarios: Different scenarios and applications have different needs, and different algorithms need to be selected to meet them.

③ Provide decision support: According to different problems and data, the algorithm can provide decision support, and effectively analyze and predict the data.

2.2.2 Common algorithms

There are many algorithms used in AI to adapt to different application scenarios and needs. There are four common categories: machine learning algorithms, deep learning algorithms, natural language processing algorithms, and recommendation algorithms.

① Machine learning algorithm:

(1) Supervised learning algorithms: such as linear regression, logistic regression, support vector machine, decision tree, random forest, etc.

(2) Unsupervised learning algorithms: such as K-Means clustering, hierarchical clustering, Expectation–Maximization algorithm, etc.

② Deep learning algorithm:

(1) Convolutional Neural Network (CNN): Mainly used in image processing and computer vision tasks.

(2) Recurrent Neural Network (RNN): Mainly used in serialization tasks such as natural language processing and speech processing.

(3) Generative Adversarial Network (GAN): Mainly used for tasks such as generating images and text.

(4) Transformer network (Transformer): Mainly used in tasks such as machine translation and text summarization.

③ Natural language processing algorithm:

(1) Keyword extraction: such as TF-IDF algorithm, TextRank algorithm, etc.

(2) Named entity recognition: such as conditional random field model, etc.

(3) Syntax analysis algorithms: such as rule models, transfer models, etc.

(4) Sentiment analysis algorithms: such as dictionary models, machine learning models, etc.

④ Recommendation algorithm:

(1) Content-based recommendation: such as TF-IDF algorithm, LDA algorithm, etc.

(2) Recommendation based on collaborative filtering: such as UserCF, ItemCF, LFM algorithm, etc.

(3) Deep learning recommendation: such as DeepFM, Wide&Deep, DIN, BERT4Rec, etc.

Algorithms are the cornerstone of artificial intelligence to achieve various tasks. With the continuous development and innovation of technology, new AI algorithms are also emerging. The key to improving AI capabilities is to choose the right algorithm for the right scene. From this point of view, the key to artificial intelligence is still artificial design.

2.3 Training scenarios affect the evolution rate

2.3.1 Importance of training

Training is the process by which an AI model learns and acquires knowledge and skills, and is therefore crucial to artificial intelligence.

① Improve model accuracy: Through reasonable data set selection and sufficient training times, the AI model can gradually learn the patterns and features in the data, thereby improving the accuracy and precision of the model.

② Support the generalization ability of the model: Training enables the model to have generalization ability, that is, it can handle new data samples outside the training set and play a role in other situations.

③ Improve the robustness of the model: training can make the AI model have better processing ability and robustness for some noisy data, forged data and interference data.

④ Update and iterate the AI model: Through training, the AI model can be continuously updated and improved to meet actual application needs.

⑤ Improve model interpretability: The interpretability of AI data models is very important in many fields. Through training, the interpretability and transparency of the model can be improved, making it easier to understand and use.

2.3.2 Training scenarios affect evolution rate

Setting and selecting training scenarios is crucial to the performance and performance of the AI model, which directly determines the quality and usability of the model. During the training process, it is necessary to preprocess data, select appropriate algorithms, optimize algorithm hyperparameters, control under-fitting and over-fitting, etc., so as to make the results of AI model training more accurate and reliable. These most important tasks essentially rely on manual participation, which also reflects the status of "people" as the core strategic resource.

① Supervised learning scenario: By providing the model with a labeled data set, it indicates which category or target value the data belongs to.

② Unsupervised learning scenario: The training data set does not have specific labels or target values. The model needs to discover the patterns and features in the data on its own based on the statistical characteristics of the data set to process and classify the data.

③ Semi-supervised learning scenario: Contains labeled and unlabeled data, and the model needs to learn how to classify these unlabeled data.

④ Reinforcement learning scenario: The model continuously interacts with the environment, observes and interacts with the environment at every moment, and adjusts the model's strategy based on the feedback information.

⑤ Natural language processing scenarios: The training data set is generally a large amount of text data, and the model needs to learn how to understand text data and the relationship between texts.

⑥ Computer vision scenario: The training data set is usually image or video data, and the model needs to learn how to understand and process these image data to achieve tasks such as target detection and image recognition.

03 "User Usage Scenario" is a scarce strategic resource

1. Limited users, unlimited data

Currently, there are certain limitations and limitations in improving the accuracy and intelligence level of artificial intelligence products by expanding computer hardware and expanding databases. On the one hand, expanding computer hardware and expanding databases require a lot of investment in manpower, financial resources, and time, and as the scale of the system continues to expand, the demand for resources will increase. On the other hand, the current technological development in the field of artificial intelligence still faces many uncertain factors, such as the effectiveness of algorithms and the effect of self-learning. These factors will affect the development space of artificial intelligence products.

Compared with unlimited data, the competition for user resources is a zero-sum game. If an artificial intelligence product has a large number of users and high frequency of use, the intelligent upgrade of the artificial intelligence product can be realized through the monitoring and analysis of user usage data, and the product system can be continuously improved to increase the value of use.

More users create a more advanced platform; a more advanced platform creates a better experience; a better experience attracts more users.

The differences in capabilities of AI platforms under different ideologies in the future are fundamentally determined by differences in the number of users and frequency of use. To gain more users and frequent use, it is necessary to continuously improve and optimize the product user interface, functions, service quality, marketing, etc., to continuously meet user needs and improve user experience, and to win the trust and loyalty of users.

2. "Matthew Effect" cannibalize whales

The Matthew effect refers to the phenomenon that under certain conditions, outstanding talents are continuously supported, resulting in a growing gap between talents.

In the development process of artificial intelligence, it is often easier for advanced technologies or advanced companies to obtain the first batch of users and market share, resulting in more investment and better returns. This phenomenon will further strengthen the leading position of the industry.

On the other hand, the development of AI technology also requires a large number of professionals. Leading companies and platforms can continue to invest more funds and resources to obtain talent advantages, so that latecomers will completely lose their ability to catch up.

The most important thing is that leading platforms will provide better user experience. Currently, the Matthew Effect in the AI to C market has become prominent. From the perspective of user psychology, users who are accustomed to using one AI product need to relearn and adapt when switching to another product of the same type, which will consume a lot of time and energy. From the perspective of data scale, the accumulation of a large amount of user data is of great significance to the use effect and analysis of products. Retaining user data is the key to establishing such products, and new products have inherent disadvantages. The "flywheel effect" inherent in AI systems further amplifies the "Matthew Effect" in user experience.

04 Conclusion

An interesting statement: AI is the bow and arrow of the Neolithic Age.

The learning ability of AI formation depends on the scale of parameters. According to academic experience, the learning ability of deep neural networks is positively correlated with the parameter scale of the model, that is, the more model parameters, the stronger the learning ability. GPT-2 has about 1.5 billion parameters, while GPT-3's largest model has 175 billion parameters, an increase of two orders of magnitude. According to media speculation but unconfirmed news, the parameters of GPT-4 may reach a scale of 100 trillion.

Some experts in the industry believe that hard technology provides key components and hardware interfaces for the development of "soft technology" and innovates around the needs of the "soft technology" supply chain. "Soft technology" is the "golden track" for scientific and technological innovation to overtake on curves and change lanes.

CCTV has figured out a set of effective combination punches: first, refine user needs into a series of AI empowerment directions such as medical assistance, agricultural assistance, beauty assistance, and educational assistance, and then lift the main station through fuzzy search semantic matching Massive media resources build high-frequency usage models in various application scenarios, forming self-organized self-learning of "soft power" such as knowledge bases and algorithms, which forces the construction of hard power such as GPU computing power and CDN bandwidth. Firmly grasp the "user usage scenario" as the "winner".

What is exciting is that the technical direction of GPT is now clear, and there are no insurmountable technical obstacles. We can use the "long-termism" spirit that the Chinese are best at to shoot this "Neolithic Age" bow and arrow most accurately and farthest.