知名科技博主犀利指出，各大实验室通过RL（强化学习），耗资数十亿美元让大模型"排练"Excel、网页操作等技能，恰 恰暴露其距真正AGI仍远。若AI真接近类人智能，就该像人类一样从经验中自主学习，而非依赖"可验证奖励训练"。而真 正突破在于"持续学习"能力，这一过程或需5-10年才能完善。 硬·AI 作者 | 龙 玥 编辑 | 硬 AI 在人工智能迈向通用智能（AGI）的竞赛中，当前最受追捧的强化学习（RL）路径，可能正将我们引向一 条歧路——RL越强，距离真正的AGI或许越远。 12月24日，知名科技博主、Dwarkesh Podcast博客主持人Dwarkesh Patel今日发布了一则引发行业深思 的视频，直击当前大模型发展的痛点。在硅谷普遍对Scaling Law（缩放定律）和RL（强化学习）持极度 乐观态度的背景下，Patel提出了一个反直觉的犀利观点： 对RL（强化学习）的过度依赖和投入，可能非 但不是通往AGI的捷径，反而是其远未到来的明确信号。 Patel的核心论点在于，当前顶尖AI实验室正耗费巨资，通过基于可验证结果的强化学习，为大模型"预 制"大量特定技能，例如操作Excel或浏览网页。然 ...

Core Viewpoint - Nvidia's third-quarter earnings report exceeded expectations, with revenue of $57.01 billion and net profit of $31.91 billion, reflecting year-on-year growth of 62% and 65% respectively, which may alleviate concerns about AI industry valuations in the stock market [2] Group 1: Financial Performance - Nvidia's Q3 revenue was $57.01 billion, surpassing market expectations of $54.92 billion, and showing a year-on-year increase of 62% [2] - The net profit for the same period was $31.91 billion, marking a significant year-on-year increase of 65% [2] Group 2: Market Dynamics - The current AI boom in the U.S. is largely driven by supply-side investments from major tech companies like Microsoft, Google, and Meta, which are heavily investing in Nvidia's GPUs to build computing power centers [2] - There are concerns that the capital expenditures for AI infrastructure are exceeding current actual demand, drawing parallels to the internet bubble of 2000 [3] Group 3: Technological Evolution - Historical tech revolutions often experience bubbles as a necessary phase, with capital flowing in before technology matures, which can lead to resource misallocation but also provides funding for technological advancements [3] - The accumulation of computing power globally may be a necessary step towards achieving Artificial General Intelligence (AGI) [3] Group 4: Future Challenges - The tech giants are entering a challenging phase where the expectations for technology commercialization must catch up with rising anticipations [4] - Investors are increasingly demanding tangible revenue and profit margins rather than just optimistic future projections, indicating a shift in focus from merely accumulating computing power to demonstrating real profitability [4] Group 5: Valuation Concerns - A potential resolution to the current valuation debate could involve a "time for space" process, where gradual technology application leads to more reasonable valuations, requiring patience from market investors [5]

Core Insights - OpenAI's marketing strategy for GPT-5 has faced criticism for being overly hyped without substantial product information [3][5][7] - The anticipated release of GPT-5 has been delayed, with various rumors about its launch date circulating since last year [8][23][26] - GPT-5 is expected to feature significant upgrades in multi-modal capabilities, software engineering, and AI agent functionalities [9][10][12][15] Group 1: Marketing and Anticipation - OpenAI has been criticized for its marketing tactics, which include frequent but vague updates about GPT-5, leading to public skepticism [3][5][7] - The excitement around GPT-5 has diminished despite OpenAI's continuous benchmark improvements, indicating a shift in marketing strategy to maintain interest [7][8] - Speculation about the release date has intensified, with predictions pointing towards a launch in early August 2025 [23][26][31] Group 2: Technical Advancements - GPT-5's core upgrades include a unified foundational and reasoning model, enhancing its performance in practical applications [9][12] - The model is expected to achieve "complete multi-modal" capabilities, allowing it to process and generate various types of media more effectively [10][11] - Significant improvements in software engineering capabilities will enable GPT-5 to modify and maintain complex enterprise-level codebases, challenging competitors like Anthropic [12][14] Group 3: AI Agent and Reasoning - GPT-5 is designed to execute complex multi-step tasks with reduced human supervision, marking a step towards autonomous AI agents [15][18] - The introduction of a "universal verifier" technology aims to enhance the model's ability to evaluate responses in subjective domains, improving its performance in creative writing and strategy analysis [16][18] - The model's architecture may include a smart routing system that dynamically selects the most suitable model for user queries based on complexity [22] Group 4: Development Challenges - The development of GPT-5 has faced challenges, including the limitations of traditional pre-training methods and the need for a new approach to scaling [40][41] - Internal projects like "Orion" did not meet expectations, leading to a pivot in strategy towards enhancing reasoning capabilities [40][41] - The successful development of the Q* technology has significantly improved the model's reasoning abilities, allowing it to tackle previously unseen problems [41][42] Group 5: Competitive Landscape - OpenAI's upcoming release of GPT-5 is crucial for regaining its competitive edge in areas like programming, where it has lost ground to competitors [50] - The partnership with Microsoft is aimed at leveraging OpenAI's technology to enhance Microsoft's own products while maintaining flexibility in their collaboration [32][35] - The competitive landscape is intensifying, with other companies like Anthropic and Google DeepMind also preparing to launch their advanced models [48][50]