EB SECURITIES: AI Agent is the key to breaking the bottleneck in the development of AI applications.

EB SECURITIES released a research report stating that OpenAI's latest model, o1, performs significantly better than GPT-4 in reasoning-intensive tasks such as programming and science competitions, but is weaker in some natural language tasks. Due to limitations in model performance, AI applications are facing a bottleneck, and the sustainability of 26 years of capital expenditure by North American technology giants and the performance growth of the upstream computing power industry chain are being questioned. The cutting-edge research papers and the underlying technologies such as reinforcement learning reasoning and thinking chains demonstrated by o1 are key to the development of the AI industry and boosting investment sentiment. The new Scaling Law, RL+COT, is crucial for achieving autonomous planning AI agents. North American technology companies are entering a new cycle of AI investment, and the significant increase in capital expenditure may put pressure on companies, leading technology giants to pay more attention to the cost-effectiveness of AI investments. Event: On September 12, 2024, OpenAI released the latest model, o1, which performs significantly better than GPT-4 in reasoning-intensive tasks such as programming and science competitions, but is weaker in some natural language tasks. o1 has global thinking ability, which complements the demand for complex reasoning and expands into vertical scenarios such as academic education. According to the evaluation by EB SECURITIES, the characteristics of o1's thinking chain can be summarized as follows: 1) Prioritize forming global methods: before answering, o1 will first analyze the problem and summarize the underlying rules; 2) Continual questioning and reflection: before outputting the final answer, o1 will continually reflect on the answering process and make improvements, with its complete thinking chain reaching several hundred lines. o1 exhibits autonomous planning ability in programming, and is expected to benefit earliest in the AI + low code/network security field. 1) Low code: o1 has strong autonomy in programming, which can to some extent offset the high cost and high latency of o1. 2) Network security: o1 performs well in network security attack and defense, can decompose complex tasks into multiple sub-tasks, possesses preliminary autonomous planning ability, and also reflects the potential threat of AI-assisted network attacks. AI-driven network security upgrades will become the main trend in the future. AI Agent is the key to breaking the bottleneck in the development of AI applications - can o1 open the path to Agent? Due to limitations in model performance, AI applications are facing a bottleneck, and the sustainability of 26 years of capital expenditure by North American technology giants and the performance growth of the upstream computing power industry chain are being questioned. The cutting-edge research papers and the reinforcement learning reasoning and thinking chains demonstrated by o1 are key to the development of the AI industry and boosting investment sentiment. The new Scaling Law, RL + CoT, is crucial for achieving autonomous planning AI agents. Reinforcement learning allows AI to autonomously explore and make continuous decisions, meeting the autonomous planning ability needed for Agents. Self-play, through autonomous games, generates high-quality data, which is advantageous for breaking the current shortage of external training data. Thinking chains can greatly enhance the model's reasoning ability involving mathematics and symbols, but the improvement in other issues is not significant, and may even harm the model's performance. The separation relationship between reasoning ability and the model's instruction following ability will become a core issue in building AGI. Under the RL paradigm, the demand for reasoning power has greatly increased, but it does not mean that the demand for training power will stop growing. The o1-preview generates about 5.9 times the output tokens of GPT-4o with the same content, with 72% of the tokens generated in the reasoning process, and using the output from o1-preview costs about 36 times that of GPT-4o. The Scaling Law shifts from the training side to the reasoning side, increasing the performance requirements for reasoning chips, and a significant amount of computing power is also needed in the pre-training phase. Reinforcement learning reasoning does not mean that model parameters will stop expanding because improving the main model parameters may lead to better reasoning pathways. North American technology companies are entering a new cycle of AI investment, and the significant increase in capital expenditure may put pressure on companies. It is expected that by 2024, the capital expenditure/operating cash flow of technology giants will exceed 40%. In the current situation where the return on investment in AI is not yet clear, technology giants will pay more attention to the cost-effectiveness of AI investments. Investment recommendations: 1) AI Power: Constellation, NRG Energy (NRG.US) 2) AI computing power industry chain: AI GPU: NVIDIA Corporation (NVDA.US), AMD (AMD.US) ASIC chip design: Marvell Technology (MRVL.US), Broadcom Inc. (AVGO.US) Storage: SK Hynix, Samsung Electronics, Micron Technology, Inc. (MU.US) Servers: LENOVO GROUP, Super Micro Computer, Inc., Dell Technologies, Inc. Class C (DELL.US), Hewlett Packard Enterprise Co. (HPE.US), Foxconn Industrial Internet CoWoS: Taiwan Semiconductor Manufacturing Co., Ltd., , Amkor Technology (AMKR.US) Network: Zhongji Innolight, Eoptolink Technology Inc., Coherent, Afnor, Arista Networks, Inc.ANET.US3. AI Applications: Cloud service providers: Microsoft Corporation (MSFT.US), Alphabet Inc. Class C (GOOGL.US/GOOG.US), Amazon.com, Inc. (AMZN.US), Oracle (ORCL.US) AI + Development/Data Analysis: ServiceNow (NOW.US), Palantir (PLTR.US), Datadog (DDOG.US) AI + Cybersecurity: Microsoft Corporation (MSFT.US), CrowdStrike (CRWD.US), Fortinet (FTNT.US) AI Agent: Microsoft Corporation (MSFT.US), Salesforce (CRM.US), Workday (WDAY.US) AI + Education: Duolingo (DUOL.US), Coursera (COUR.US) Risk analysis: Bottlenecks in AI technology research and product development; Increasing competition in the AI industry poses risks; Commercialization progress falling short of expectations pose risks; Domestic and international policy risks.