2023 AI Industry Special Report META's AI Capability Analysis

Core Ideas:

CORE TAKEAWAY: THIS REPORT FOCUSES ON META'S AI CAPABILITY ANALYSIS. The purpose of Meta's AI is simple and clear, that is, "AI can help companies better understand user needs": Meta has a large number of social and metaverse algorithms, established an artificial intelligence research institute, launched the Pytorch framework, released hardware support algorithms such as Zion, etc. This report specifically compares Meta with Microsoft, Google, and Amazon in the field of AI, and we find that Meta is on par with or even slightly better than the other Big Three in some aspects of the AI field. We believe that Meta's artificial intelligence has grown into a non-negligible existence under the leadership of Yann LeCun, and Meta is actually one of the number one AI players in the world to be paid attention to.

The company started with social and is currently focusing on the metaverse on the basis of social. Formerly known as Facebook, Meta was founded in 2004 and is one of the world's largest Internet companies. The company's main business consists of two parts, namely Family of Apps (FoA) and Reality Labs (RL). Family of Apps (FoA) is the vast majority of revenue, with advertising accounting for more than 95%. FoA includes social media messaging software based on Facebook, Messenger, Instagram, WhatsApp, as well as business software Workplace, and metaverse platform Meta Horizon, Reality Labs (RL), mainly engaged in AR/VR devices, AI technology and metaverse development. The company's total operating income in fiscal 2022 was US$116.609 billion and net profit was US$29.146 billion.

AI runs through the company's entire business process, and the company is a major AI player in the world. The purpose of Meta's AI efforts is simple and clear, that is, "AI can help companies better understand user needs". At present, Meta has a rich accumulation of social algorithms and metaverse algorithms, which play a pivotal role in the company's business development process. Meta established the AI Research Institute in 2013, mainly focusing on artificial intelligence (AI) and machine learning (ML), the two most popular and cutting-edge technology fields in science and technology, the institute is led by Yann LeCun, chief scientist of Meta, the three giants of deep learning, and the proposer of convolutional neural networks. In 2017, Meta launched the PyTorch framework, whose dynamic graph programming can help users build models efficiently, and the launch of PyTorch 2.0 in 2023 has improved the speed of model compilation by a notch. In terms of AI hardware, the company launched a number of AI hardware products, including the Zion platform, to train and fine-tune AI models. In addition, the company released Glow and ONNX for the compilation of AI models and the opening of the ecosystem.

We compare Meta with Microsoft, Google, and Amazon in the field of AI, and the conclusions are as follows: 1) From the perspective of AI management strategy, the number of times "AI" related to "AI" appeared in the minutes of Meta's performance exchange meeting was second only to Google, and even surpassed Google in 22Q1-22Q3, which is of course closely related to Meta's vigorous layout of the meta-universe; 2) From the perspective of AI engineering practice, we find that in terms of the number of fans and projects of the Github website, Microsoft is undoubtedly the leader, followed by Google, Meta, Amazon; 3) From the perspective of AI social influence, the number of Meta social media fans is second only to Google; 4) From the perspective of AI academic capabilities, Meta has fewer academic papers than Google and Microsoft;

5) From the perspective of AI applications, Meta AI is more concentrated in the C-end and visual fields, Microsoft AI is also biased towards C-end users, while Google and Amazon are more inclined to B-side; 6) From the perspective of AI power, because of Yann LeCun's AI concept, Meta pays more attention to AI self-supervised learning ability; 7) From the perspective of AI large models, the layout time and number of Meta in AI large models are weaker than those of Google and Microsoft, of which LLaMA is the most typical representative.

First, the Meta basic disk

Starting with social networking, he is currently focusing on the metaverse

Meta, formerly known as Facebook, is one of the world's largest internet companies. Founded in February 2004 by Mark Zuckerberg, the company is a leader in social media platforms and AR/VR. In April 2012, the company was listed on NASDAQ under the symbol FB, and later changed its name to META. Since 2011, the company has continued to deepen the field of social media platforms, forming a social media application matrix consisting of Facebook, Instagram, Messenger and WhatsApp. In 2014, the company acquired Oculus to develop AI, VR/AR and metaverse-related businesses. In 2021, in order to achieve the strategic goal of the metaverse and reshape the brand image, the company announced that it changed its name to "Meta" and established Facebook Reality Labs to focus on the research and development of AR/VR and other virtual reality technologies. The Meta Quest standalone VR headset launched by the lab can be applied to the social media platform Meta Horizon, gaming, office and other fields.

Meta's main business structure

The company's main business consists of two parts, namely Family of Apps (FoA) and Reality Labs (RL). In fiscal 2022, the company's operating income totaled US$116.609 billion; in 22Q4, the company achieved revenue of US$32.165 billion. Family of Apps (FoA) is the vast majority of revenue, with advertising accounting for more than 95%. FoA includes social media messaging software such as Facebook, Messenger, Instagram, WhatsApp, as well as business software Workplace, and metaverse platform Meta Horizon. Reality Labs (RL) mainly focuses on the development of AR/VR devices, AR/VR technology, AI technology and metaverse.

Due to the external environment, operating income declined slightly

Since 2022, Meta's revenue profit has continued to decline due to overseas macro pressure. The company's total operating revenue in fiscal 2022 was US$116.609 billion, down 1.1% year-on-year. Among them, the revenue of FoA business was US$114.45 billion, down 1% year-on-year; RL revenue was $2,159 million, down 5% year-over-year. The company's net profit for fiscal 2022 was US$29.146 billion, down 25% year-on-year. The reasons for the decline in operating income and net profit are: 1) the sales of VR equipment and software are not as expected due to the development of the metaverse is not as expected; 2) weak overseas macroeconomics and weakening demand; 3) With the development of the metaverse strategy, the company increased R&D spending, and the company's net profit declined in fiscal 2022. 4) In 2022, the pace of business expansion accelerated, and capital expenditure increased significantly.

Advertising business is the main source, and the R&D expense ratio has increased significantly

From the perspective of business type, advertising business revenue is the most important source of Meta's operating income. In fiscal 2022, the company's advertising revenue was 113.441 billion US dollars, accounting for 98.15%. From fiscal year 2015 to FY2019, the proportion of advertising revenue to operating revenue increased year by year, from 95.28% in FY2015 to 98.53%. In fiscal 2022, the company's gross margin was 25%, a significant decrease from fiscal 2021, mainly due to the increase in research and development expenses. In fiscal 2019, due to Facebook user data leaks, the company's gross margin fell to 34%. The Company's R&D expense ratio was 30% in FY2022. The company announced its transformation from 2021 to penetrate into the metaverse, resulting in a significant increase in the R&D expense ratio in fiscal 2022 compared to 21% in fiscal 2021.

North America is the main source of revenue

In terms of geographical distribution, North America is the main distribution area of business revenue. In 22Q4, North America generated revenue of US$15.636 billion, accounting for 48% of total revenue. It was followed by Europe and Asia Pacific, with revenues of US$7.05 billion and US$6.05 billion, accounting for 19.1% and 22.1% respectively. Other regions generated operating income of US$3.429 billion, accounting for 10.8%. For many years, the United States and Canada have been the two main countries for Meta's advertising business revenue, and the Asia-Pacific region is growing rapidly. Since 2022, the proportion of revenue in the Asia-Pacific region to total operating revenue has been increasing, from 17% in Q4 200 to 21% in Q3 202, with the rapid growth of daily active users in the Philippines, Indonesia, Vietnam and other countries driving the continuous growth of operating income in Asia.

The number of users showed a slow growth trend, and the per capita revenue declined to a certain extent

Facebook's new user growth has slowed. As of 22Q4, Facebook's monthly active users (MAU) reached 2.96 billion, a year-on-year increase of 0.2%; Daily active users (DAU) reached 2 billion, up 0.8% year-on-year. Among them, the number of daily active people in North America was 199 million (accounting for 9.95%); The number of daily active people in Europe is 300 million (15%; The number of monthly active people in Asia Pacific was 854 million (42.7%); In the rest of the world, there were 643 million monthly active people (32.15%). Meta's per capita revenue in fiscal 2022 was $7.95/person, down 6% year-over-year. In 22Q1, due to macroeconomic headwinds, advertiser bidding demand weakened, and inventory did not grow significantly, resulting in a significant slowdown in revenue and a certain decline in per capita revenue.

Instant messaging platforms: WhatsApp and Messenger

WhatsApp is a software focused on the communication experience, it allows users to send text voice messages, make voice and video calls, share images, documents, user location and other content, and can share messages, photos and videos with up to 256 contacts at once. After linking a bank card, WhatsApp's payment feature allows users to transfer money to friends or split bills. Similar to WhatsApp, Messenger is an instant messaging app and platform for sending messages and exchanging photos, videos, stickers, audio, and files, as well as supporting voice and video calls. In addition, Messenger also supports functions such as friend transfer and business communication. For cross-platform communication, Messenger can use Facebook, Instagram, Portal and Oculus to chat.

Metaverse strategy

Since 2014, Meta has gradually entered the field of virtual reality through the acquisition of virtual reality technology companies such as Oculus VR, and launched a series of VR headsets, such as the Oculus Rift and Oculus Quest 2, to promote the popularization of virtual reality technology to the consumer market. In 2020, the company changed its name to Meta and officially announced its metaverse strategy, planning to shift its focus to the metaverse over the next 10 years, investing billions of dollars in R&D and development, launching products such as Horizon Workrooms and Spark AR Studio to improve remote work, drive user-generated content and social interaction, and strengthen collaboration with cross-industry partners. In terms of metaverse technology, Meta has carried out a lot of technology research and development in its Reality Labs laboratory, including virtual reality, augmented reality, computer graphics, artificial intelligence and other fields. The company invests a lot of resources in technology research and development, and promotes continuous breakthroughs and innovations in virtual reality technology, including graphic rendering, sound, interaction, human tracking, eye tracking, etc.

Second, a longitudinal introduction to AI capabilities

AI Research Institute: In 2013, Meta established the FAIR Research Institute

Meta established the FAIR (Facebook AI Research) Institute in 2013, focusing on artificial intelligence (AI) and machine learning (ML), the two most popular and cutting-edge technology fields in technology and business. Its research objectives include improving the user experience of social media, enhancing information security and privacy protection, promoting the fairness, transparency and explainability of AI, and promoting the application of AI in the social and economic fields. These studies involve cutting-edge AI technologies such as large-scale data processing, deep learning, reinforcement learning, generative adversarial networks (GANs), self-supervised learning, and transfer learning. FAIR Institute is committed to promoting the application of its research results in different industries, such as healthcare, finance, transportation, education, metaverse and other fields, and promoting the landing and application of artificial intelligence and machine learning technology.

The AML team pays more attention to application implementation

AML (Applied Machine Learning) is Meta's machine learning application team, responsible for developing and optimizing machine learning models applied to Meta's products and services, including model development, data mining and feature engineering, model deployment and optimization, real-time monitoring and model iteration, and cross-team work. The AML division works in a variety of areas, including recommendation systems, ad sequencing, speech recognition, image recognition, etc., for Meta's products and services, such as Facebook, Instagram, WhatsApp, etc. By continuously optimizing and iterating on machine learning models, AML plays a key role in Meta, driving the company's technological innovation and business development to provide users with a better product and service experience.

AI Scientist: Yann LeCun is Meta's Chief Scientist

Meta has a strong AI team, including Yann LeCun, founder of deep learning, Andrew Bosworth, chief technology officer, Hassan Sawaf, head of Facebook AI, Mike Schroepfer, senior researcher, and artificial intelligence scientists Jerome Pesenti and He Kaiming. Their extensive experience and outstanding achievements in the field of artificial intelligence and machine learning have driven Meta's leadership in technology and innovation. These core personnel have led and promoted Meta's multiple projects in the field of artificial intelligence and machine learning, including ad sequencing, recommendation systems, etc., which play a vital role in the company's technological development and business innovation. Their expertise and experience enable Meta to continue to push the frontiers of AI technology, meet future challenges and remain at the forefront of the global technology industry.

AI hardware: Meta launched a number of AI hardware products, including the Zion platform

Zion is Meta's next-generation large-memory unified training platform designed to efficiently handle a range of neural networks, including CNNs, LSTMs, and SparseNN. The Zion platform delivers high memory capacity and bandwidth, flexible, high-speed interconnect, and powerful computing power for critical workloads of AI models. The Zion system consists of three modules: Angels Landing, Clear Creek and Emerald Pools, of which Angels Landing is the CPU box; Clear Creek is an interconnect box; Emerald Pools are accelerator boxes. Meta's Kings Canyon is an AI inference acceleration module designed to boost computing performance in artificial intelligence and machine learning. It can be embedded in a variety of devices to provide efficient AI reasoning capabilities. Kings Canyon's Inference Acceleration Module accelerates deep learning inference tasks across threads in the device for faster, smarter computations.

Compared to other AI frameworks, Pytorch is more flexible

PyTorch is an open-source deep learning framework that has one of its biggest advantages, and one of its biggest advantages is the flexible way of programming dynamic graphs. Compared to the static graph framework, PyTorch allows users to build and modify the computational graph in real time during the module construction and debugging process. This means that researchers and developers can debug and optimize models more intuitively, and experiment and iterate more efficiently. In addition, the dynamic graph programming method also makes model construction more flexible, allowing users to introduce complex computational logic such as control flow and dynamic shape into the model, so as to meet the needs of more complex deep learning tasks.

In 2023, Meta released PyTorch 2.0

PyTorch 2.0 was released in March 2023, and it adds a number of powerful features to the new version, including full support for distributed training and quantitative reasoning, making it even better at model compilation, large-scale training, and efficient inference. At the same time, PyTorch 2.0 also provides comprehensive tools and libraries to simplify model deployment and model production environment management, making the implementation of deep learning applications easier and more reliable. PyTorch 2.0 focuses on the developer experience, providing a more concise, flexible and easy-to-use API, making model building, training and debugging more convenient. In addition, PyTorch 2.0 supports multiple hardware acceleration and model optimization techniques, including GPU, TPU, and quantization acceleration, to provide higher performance and efficiency.

AI ecosystem: ONNX, an open ecosystem for machine learning

ONNX (Open Neural Network Exchange) is an open ecosystem released by Meta that enables AI developers to choose the right tools as their projects grow. ONNX provides an open-source platform for AI models, including deep learning and traditional ML. It defines an extensible computational graph model, as well as built-in operators and definitions of standard data types. Currently, the platform focuses on improving reasoning (scoring) capabilities. ONNX enables models to be trained in one framework and then exported and deployed to other frameworks for inference. ONNX models currently run well in frameworks such as PyTorch, Caffe2, Microsoft Cognitive Toolkit, Apache MXNet, and Chainer, and are scalable to platforms such as Core ML, TensorFlow, Qualcomm SNPE, Nvidia's TensorRT, and Intel's nGraph.

Web framework: React

The React Framework is a JavaScript library for building user interfaces. Developed and maintained by the Facebook team, the framework is a flexible, efficient, and powerful front-end framework for building web applications of all sizes and complexity. It has become an important tool in modern front-end development and is widely used in many enterprises and communities. React can be used as the basis for developing single-page, mobile or server rendering applications with frameworks such as Next.js. However, React only focuses on state management and rendering state to the DOM, so creating a React application often requires the use of additional tooling libraries for routing implementations, as well as some client-side functionality. React's design philosophy is unique, revolutionary, with outstanding performance and very simple code logic. Therefore, more and more people are beginning to pay attention to and use it, thinking that it may be the mainstream tool for web development in the future.

Third, horizontal comparison of AI capabilities

AI management strategy layer: The number of "AI" appeared in the minutes of the Meta performance exchange meeting

We believe that the number of times "AI" related to "AI" appears in the minutes of the company's performance exchange meeting represents the importance of management to the relevant AI layout, and the more times AI appears, the more attention the management attaches to AI at the strategic level, so we counted the number of AI occurrences in the performance meeting of several manufacturers. According to the number of AI mentioned in the quarterly performance briefings of Google, Microsoft, Meta, and Amazon, Google increased from 4 times in 2020Q1 to 64 times in 2022Q4, the most obvious increase; Meta increased from 15 times in 2021Q2 to 37 times in 2022Q4, reflecting Meta's relatively early layout for AI; Microsoft decreased from 26 fluctuations in 2020Q1 to 7 in 2022Q4; Amazon barely mentioned anything related to AI in the earnings briefing.

AI social impact: Meta Twitter is second only to Google in terms of followers

In terms of social influence, we believe that the number of Twitter followers can reflect the influence of major manufacturers at the social level to a certain extent, and the greater the number of account followers, the wider the spread and coverage. Therefore, we counted data from Google (Google, Google Research, DeepMind), Microsoft (Microsoft, Microsoft Research, OpenAI), Meta (Meta, Meta AI), Amazon (aws) on Twitter. Specifically, Google's Twitter platform has 29,100,000 followers, ranking first among manufacturers, with a clear lead; Meta and Micosoft have 13,900,000 and 12,800,000 followers respectively, ranking behind Google but also relatively large, reflecting their greater social influence to a certain extent.

AI academic capabilities: Meta has fewer academic papers than Google and Microsoft

In terms of the total number of research projects, Meta has a total of 4478 studies and Amazon has 2679 studies. In terms of specific areas, Meta's research covered a total of 14 areas, mainly distributed in Artificial Intelligence (1200 papers), Computer Vision (949 papers), Computer Vision (508 papers) and Natural Language Processing & Speech (472 papers). Amazon's research involves a total of 13 fields, mainly distributed in Conversational AI / Natural-language processing (1001 articles) and Machine learning (756 articles), accounting for 65.58%, and the distribution is relatively concentrated.

AI application fields: Meta AI is more focused on the C-end and vision fields

Meta AI capabilities are mainly focused on C-end and vision, including image recognition and object detection, image generation, augmented reality (AR) and virtual reality (VR), face recognition and human face related technologies, as well as social network data processing and privacy protection. These technologies can be applied in social media platforms to provide better user experience and social interaction. Technologies such as face recognition login, face feature extraction, and face emotion recognition can enhance the entertainment attributes of social media platforms; Meta AI also provides data processing and privacy protection technology to process and manage users' data on social media platforms, to protect user privacy and data security while improving functions such as content recommendation and user interest modeling.

Fourth, the Meta hotspot model

Meta's main big model: OPT

The OPT (Open Pre-trained Transformer Language Models) model is a series of open-source large-scale causal language models proposed by Meta, which can be used to generate creative text, solve basic math problems, answer reading comprehension and other problems. Meta has open-sourced OPT-175B, this model has 175 billion parameters, trained on multiple databases such as BookCorpus, CC-Stories, The Pile, Pushshift.io Reddit, CCNewsV2, etc., with a total of 180 billion characters of training data and more than 800GB of text data. While the performance of OPT-175B is similar to that of GPT-3, the training process only uses 1/7 of the computing resources of GPT3.

Meta main large model: LLaMa

LLaMa (Large Language Model Meta AI) is Meta's latest open-source large-scale language model with parameters ranging from 7 billion to 65 billion. LLaMA achieves results comparable to very large models with relatively few parameters, and the LLaMA model with 13 billion parameters can outperform GPT-3 with 175 billion parameters on most benchmarks; The LLaMA model with a maximum of 65 billion parameters is comparable to Google's Chinchilla-70B and PaLM-540B. And the model can run on a single GPU, which greatly reduces the difficulty of using NLP developers. LLaMa was trained on a corpus of more than 1,400 billion tokens and more than 4,800 GB with training datasets including CCNet (67%), C4 (15%), GitHub (4.5%), Wikipedia (4.5%), Books (4.5%), ArXiv (2.5%), and Stack Exchange (2%).

Meta's main big model: SAM

On April 5, 2023, Meta released the SAM Model (Segment Anything Model), which is the basic model in the field of CV image segmentation. SAM consists of an image encoder, a prompt encoder and a mask decoder, which has the following characteristics: 1) task learning using prompt engineering, interactive clicking, selection box and even text manipulation for object segmentation; 2) With fuzzy perception function, multiple effective masks can be output when there is ambiguity in the face of segmentation objects; 3) Similar to ChatGPT, SAM cleverly combines manual labeling with big data (data engine) to finally achieve the function of "dividing everything".

Meta's main big model: DINOv2

DINOv2, is the first method released by Meta on April 17, 2023 to train computer vision models using self-supervised learning. Models using the DINOv2 method, which can be trained on any image set, provide powerful performance without fine-tuning the model, and are suitable as the backbone for many different computer vision tasks. Thanks to the addition of self-monitoring feature learning and lightweight task-specific execution modules, DINOv2 performs significantly better than other algorithms in depth estimation. The test results show that DINOv2 has very powerful predictive power for tasks such as classification, segmentation and image retrieval. At the same time, the model performs better than other models in scenarios such as image network classification, video classification, target recognition, and intensive recognition tasks.

(This article is for informational purposes only and does not represent any investment advice from us.) For information, please refer to the original report. ）

Source: Future Think Tank

Related reports

• Artificial Intelligence Industry Special Report: One of the world's AI "number one players", META.pdf
• (Artificial Intelligence) Artificial Intelligence Foundation Exam Syllabus .docx
• AI empowers smart bridge solutions .pptx
• How much economic growth can AI drive? .pdf
• The U.S. Artificial Intelligence Bill of Rights Blueprint .pdf
• Internet industry special report: wave by wave big model, Internet giant's AI ambition .pdf
• U.S. technology stocks review, performance summary and outlook: 23Q1 exceeded expectations, 23Q2 growth stabilized, and actively promoted AI commercial .pdf
• Flush (300033) research report: Finance + AI opens the upward engine, which is expected to drive the continuous growth of performance .pdf
• Special report on the communication industry: Open source race, the "Linux moment" of AI big model has come .pdf
• Media Industry Feature: From repair to growth, AIGC opens a new era of media Internet .pdf