深度｜Anthropic CEO：AI行业的盈利本质上源于对市场需求的低估，而亏损则因为高估了需求，2030年AI行业营收将跃迁至万亿美元级

Core Insights - The core insight of the article revolves around the nearing end of exponential growth in AI technology and the adherence to the scaling hypothesis, which posits that only seven key factors drive technological advancement in AI [3][5][6]. Group 1: Technological Development - The underlying technology has developed exponentially, aligning with expectations, with significant advancements in code-related fields surpassing initial predictions [3][4]. - The scaling hypothesis, established in 2017, remains unchanged, asserting that the core elements driving technological progress are limited and include raw compute power, data scale, data quality, training duration, and the potential for infinite scaling of objective functions [5][6]. - There is a 90% confidence that within ten years, data centers will produce genius-level AI comparable to a nation, with models expected to achieve end-to-end code development within 1 to 2 years [3][12]. Group 2: AI Industry Profitability - The profitability of the AI industry hinges on accurately predicting compute demand, with profits arising from underestimated demand and losses from overestimated demand, contrasting with traditional industry profit logic [3][12]. - The API business model retains long-term viability, with the emergence of diverse pricing models based on the value delivered in different scenarios [3][12]. Group 3: General AI Development Predictions - There is a strong belief that general artificial intelligence will be achieved within this century, with a high probability of significant breakthroughs occurring within the next decade [12][13]. - The company has observed a remarkable revenue growth trajectory, with projections indicating a rise from $0 to $1 billion in 2023, and further growth to $10 billion in 2024, and $9 to $10 billion in 2025 [19][20]. Group 4: Model Learning and Generalization - The learning process of models differs fundamentally from human learning, with models requiring vast amounts of training data to achieve generalization capabilities, unlike humans who learn from fewer examples [8][9]. - The current model training process involves pre-training and reinforcement learning (RL), with the latter showing similar scaling laws to pre-training, indicating a convergence of learning methodologies [6][8]. Group 5: AI Technology Penetration - The penetration of AI technology into the economy is expected to be rapid but not instantaneous, influenced by various factors such as organizational change management and system integration [20][21]. - The company emphasizes that while the capabilities of models are growing exponentially, the actual implementation and integration into economic systems will take time due to existing limitations [21][22].