Smart Finance and Economics APP learned that Everbright Securities released a research report stating that the latest model o1 released by OpenAI outperforms GPT-4o significantly in programming, science competitions, and other reasoning-intensive tasks, but is weaker in some natural language tasks. Limited by model performance, AI applications are entering a bottleneck, and the sustainability of capital expenditure by North American technology giants for 26 years, as well as the performance growth in the upstream computing power industry chain, are being questioned. Recent cutting-edge papers and the reinforcement learning reasoning 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 Al Agents capable of autonomous planning. North American technology companies are entering a new cycle of AI investment, and a substantial increase in capital expenditure may put companies under cost pressure, prompting tech giants to pay more attention to the cost-effectiveness of AI investments.

Event: On September 12, 2024, OpenAI released the latest model o1, which outperformed GPT-4o significantly in programming, science competitions, and other reasoning-intensive tasks, but was weaker in some natural language tasks.

o1 possesses global thinking ability, complements complex reasoning for long-tail demands, and explores academic education and other niche scenarios.

According to the evaluation by Everbright Securities, the thinking chain feature of o1 can be summarized as: 1) Prioritizing global methods: Before answering, o1 will first analyze the problem and summarize the underlying rules; 2) Continuous questioning and reflection: Before outputting the final answer, o1 will continuously reflect on the answering process and make improvements. Its complete thinking chain can reach hundreds of lines.

o1 demonstrates autonomous planning ability in programming, and the AI+ low code/network security field is expected to benefit earliest.

1) Low code: o1 exhibits strong autonomy in programming, to some extent mitigating the high cost and high latency issues of o1; 2) Network security: o1 performs well in network security offense and defense, can decompose complex tasks into multiple sub-tasks, possesses preliminary autonomous planning capabilities, and also reflects the potential threat of AI-driven network attacks. AI-driven advancements in network security offense and defense will be the main theme in the future.

AI Agent is the key to breaking the bottleneck of AI application development, can o1 open the path to Agent?

Limited by model performance, AI applications are facing bottlenecks. The sustainability of capital expenditure by North American technology giants in 26 years and the performance growth of the upstream computing power industry chain are being questioned. The underlying technologies such as reinforcement learning reasoning and cognitive chains demonstrated in recent cutting-edge papers and o1 are crucial for boosting the development and investment sentiment of the AI industry.

The new Scaling Law, RL+CoT, is crucial for achieving self-planning AI Agents. Reinforcement learning enables AI to autonomously explore and make continuous decisions, aligning with the self-planning capabilities required by Agents. selfplay generates high-quality data through autonomous gameplay, which helps overcome the shortage of external training data.

Cognitive chains can greatly enhance the model's reasoning ability involving mathematics and symbols, but the improvement in other areas is not significant and may even harm the model's performance. The separation between reasoning ability and the model's instruction-following ability presents a challenge in building AGI, as balancing the relationship between the two becomes a core issue.

Under the RL paradigm, the demand for reasoning computing power significantly increases, but it does not mean the training computing power demand will stop growing. The o1-preview generates about 5.9 times as many output tokens as GPT-4o, with 72% of tokens generated during the reasoning process. The cost of using o1-preview output is about 36 times that of GPT-4o. Scaling Law shifts from the training side to the reasoning side, increasing the performance requirements for reasoning chips. Moreover, a significant amount of computing power is also needed in the pre-training phase. Reinforcement learning reasoning does not imply a halt in model parameter expansion, as improving the main model parameters may lead to better reasoning paths.

North American technology companies are entering a new AI investment cycle, with a significant increase in capital expenditure potentially putting pressure on companies' costs. In 2024, the capital expenditure/operating cash flow of technology giants is expected to exceed 40%. In the current situation where the return on AI investment is not yet clear, technology giants will pay more attention to the cost-effectiveness of their AI investments.

Investment advice:

1. AI Electrical Utilities: Constellation, NRG Energy (NRG.US)

2. AI computing power industry chain:

AIGPU: Nvidia (NVDA.US), AMD (AMD.US)

ASIC chip design: Marvell Technology (MRVL.US), Broadcom (AVGO.US)

Storage: SK Hynix (000660.KS), Samsung Electronics (005930.KS), Micron Technology (MU.US)

Server: Lenovo Group (00992), Super Micro Computer (SMCI.US), Dell Technologies (DELL.US), Hewlett Packard Enterprise (HPE.US), Foxconn Industrial Internet (601138.SH)

CoWoS: Taiwan Semiconductor (TSM.US), Silitech Technology, Amkor Technology (AMKR.US)

Network: Zhongji Innolight (300308.SZ), eoptolink technology inc., (300502.SZ), Coherent (COHR.US), Finisar (AVF.US), Arista Networks (ANET.US)

3. AI applications:

Cloud computing service providers: microsoft (MSFT.US), google (GOOGL.US/GOOG.US), amazon (AMZN.US), Oracle (ORCL.US)

AI + Development / Data Analysis: ServiceNow (NOW.US), palantir (PLTR.US), datadog (DDOG.US)

AI + Cybersecurity: microsoft (MSFT.US), crowdstrike (CRWD.US), Fortinet (FTNT.US)

AI Agent: microsoft (MSFT.US), Salesforce (CRM.US), Workday (WDAY.US)

AI + Education: Duolingo (DUOL.US), Coursera (COUR.US)

Risk Analysis: AI technology research and product iteration encounter bottlenecks; intensified competition in the AI industry increases risk; commercial progress falls short of expectations introduces risk; domestic and foreign policy risks.