Opportunities and challenges

The model capabilities are gradually converging

The gap between large models is gradually converging, and more and more large models are comparable to GPT4 in terms of capabilities and practicality. OpenAI's core capabilities are mainly reflected in three aspects: 1) the ability to collect and process data, including data sources, cleaning methods, data structures, etc.; 2) Model structure, including attention, hidden layer design, etc.; 3) Training methods, including various hyperparameters, learning rate, etc.

The cost of large models has been significantly reduced: the call cost of OpenAIGPT4 has been reduced by more than 90% in 23 years, and the latest GPT4-O multimodal is free and GPTs, image understanding, code interpreters, etc., which were only available to paid users before, are all free. It is expected that with the continuous development of computing power and model architecture, GPT-4 level models will be free of charge just around the corner.

There are more and more refined models for vertical categories to ensure the cost performance of subdivided scenarios: different industries have their own vertical models, and models of different sizes are suitable for different scenarios. Due to the differences in enterprise scenarios and business processes, different domain model functions, datasets, model sizes, and performance are quite different, and customized models or small models are more cost-effective.

AI empowers tech giants

Case: Microsoft itself has strong user traffic entrances such as Office and Windows, and bundles OpenAI to take the lead in using leading large models, combined with its own strong software technology and cloud computing resources, of which 24Q1AI contributes 7% of Azure cloud revenue and drives Microsoft's cloud business to achieve 31% YoY accelerated growth.

Microsoft's rich use cases can deeply integrate GPT to enhance the competitiveness of existing products, and it is one of the companies that have significantly benefited from AI technology in downstream application fields.

AI-powered advertising and hardware

Meta: AI + advertising and social media, significantly improve advertising ROI

(1) MetaAI: The AI assistant can answer questions, obtain real-time information, and generate images, which was launched in September 23 and is already available on Facebook, Ins, WhatsApp, and Messenger.

(2) Recommendation model: The extended recommendation model can better adjust the number, time, place, and object of ad impressions, about 30% of the posts in the Facebook push are delivered by the AI recommendation system, which has increased by 2 times in the past few years, more than 50% of the content on Instagram is recommended by AI, and the average price of ads has accelerated for the fifth consecutive quarter.

(3) METAAdvantage+: Generating personalized ads based on user preferences, optimizing copy and placement, segmenting targeting, and expanding lookalike audiences, its revenue has more than doubled since last year, and the average cost per click has been reduced by 28%.

Overseas hot AI application directions

From the perspective of SaaS products with rapid implementation of AI functions, the core of AI applications in the future will be empowered in three directions: 1) increasing the unit price of products; 2) increase the payment rate; 3) Expand the user base.

(1) Improve product efficiency, add SKUs, and increase ARPU. For example, the Microsoft M365 Copilot feature is priced at $30/month, and the SalesforceSalesGPT feature is priced at $50/month.

(2) Generate new payment points and increase the payment rate. For example, Meitu, a Hong Kong-listed company we track, has added a large number of AI functions (AI avatars, AI filters, etc.) to its original membership benefits, and the membership payment rate will increase from 2.3% to 3.7% in 2023.

(3) Expand the depth of use of the original products and the scope of customers. For example, Palantir said that AI drives the growth of commercial customers and lowers the barrier to product use.

Artificial intelligence scenario requirements

How can you tell if AI is empowering or disrupting existing businesses?

From the core perspective of judging whether AI will be impacted or empowered, we should focus on whether the core pain points solved by the original business model will be replaced by GPT. For example, Chegg's business model, which is also an educational IT company, solves the core problem of helping students find the questions and answers of previous exams, and GPT may replace some of the needs to find answers. The core of Duolingo's business model is to develop apps that are willing to actively learn by users by designing gamification mechanisms.

At the level of independent applications, we are optimistic about two types of AI needs:

1) Rapid iteration and promotion of product functions after productivity improvement. For example, Meitu launched seven AI products and dozens of AI functions in 23, and WeShop, an AI commercial auction tool on Mushroom Street, had 300,000+ users worldwide in March. For example, Remini's recently launched "Clay Effects" filter is quickly available on Jianying, the Meitu App, Midjourney, Lora and other products after becoming popular all over the Internet.

2) The need for AI to accompany the scenario. For example, in the field of education, Duolingo launched a voice dialogue AI assistant to meet the needs of users for oral learning and flexible interaction. For example, the emotional companionship and virtual social products Glow and Character.AI, and the domestic Insop, XHer, Dream Island, Caiyun Xiaomeng, and AuraAI have attracted a large number of OC (original characters), AGC enthusiasts and users who are eager to accompany and communicate.

Case comparison under the wave of artificial intelligence

The challenge posed by artificial intelligence – cybersecurity

The revenue contribution of AI is low, and cost performance has become the core pain point. For example, overseas SaaS products with rapid progress in the commercialization of AI functions are expected to contribute to the low single-digit level this year, and users are still cautious about the ROI of AI functions.

The boundaries of AI products are not clear, the market is not transparent, and the supervision is not clear, so enterprises have to explore. Taking AI software as an example, Xiaofan Table interviewed and reported that the two major ways for AIGC to go overseas to make money are restricted-level online articles and video face swapping, and with the development of GenAI technology, a large number of false, vulgar, and low-quality content floods the Internet.

Investment and financing market

Financing of AI projects in China in the first quarter

According to the data of Qibusiness Card Pro, there were 81 domestic AI application financing projects in the first quarter of 24. At present, domestic AI application financing projects are concentrated in the fields of intelligent robots, industrial intelligence, intelligent driving & transportation, AIGC, medical treatment, and energy solutions.

Financing of AI projects in foreign countries in the first quarter

Foreign AI application financing projects focus on machine intelligence, industrial, medical, commercial, and security solutions.