"谷歌AI落后了"。这是ChatGPT诞生以来，很长一段时间内，科技行业的主流声音。 过去三年，全球科技圈的主流叙事中有且仅有两个主角：英伟达负责卖铲子，提供GPU硬件基础；OpenAI负责挖金 矿，凭借Scaling Law定义前沿模型。 谷歌则长期被认为是在AI竞赛中落后的巨头，模型不如OpenAI，甚至连搜索业务也面临被取代的风险。直到谷歌创始 人谢尔盖·布林（Sergey Brin）回归一线，亲自管理AI业务，并先后Gemini 2.5和Gemini 3系列大模型后，一切才彻底发 生了改变。 上个月，谷歌推出最新的Gemini 3系列模型和第七代TPU Ironwood，彻底改变了AI行业的游戏规则。现在，轮到 OpenAI拉响红色警报了。 12月13日，谢尔盖·布林现身母校斯坦福大学工程学院的百年校庆活动。面对台下数百名年轻的工科生，布林首次坦诚 谷歌在AI浪潮初期的战略误判，并深入剖析了这场绝地反击背后的布局。 当被问及谷歌在AI早期的被动局面时，谢尔盖·布林说，尽管谷歌早在八年前就发布了Transformer论文，但公司内部却 并未给予足够的战略重视。 "我们当时在算力扩展上的投入过于保守，甚至 ...

Core Viewpoint - The article discusses the historical development and significance of the Scaling Law in artificial intelligence, emphasizing its foundational role in understanding model performance in relation to computational resources [2][34][43]. Group 1: Historical Context - The Scaling Law's origins are debated, with claims that it was first proposed by OpenAI in 2020 or discovered by Baidu in 2017 [2]. - Recent discussions attribute the initial exploration of Scaling Law to Bell Labs, dating back to 1993 [3][5]. - The paper from Bell Labs demonstrated the relationship between model size, data set size, and classifier performance, highlighting the long-standing nature of these findings [5][9]. Group 2: Key Findings of the Research - The NeurIPS paper from Bell Labs outlines a method for efficiently predicting classifier suitability, which is crucial for resource allocation in AI model training [12]. - The authors established that as training data increases, the error rate of models follows a predictable logarithmic pattern, reinforcing the Scaling Law's validity [12][16]. - The research indicates that after training on 12,000 patterns, new networks significantly outperform older ones, showcasing the benefits of scaling [16]. Group 3: Contributions of Authors - The paper features five notable authors, including Corinna Cortes and Vladimir Vapnik, both of whom have made significant contributions to machine learning and statistical theory [18][19][27]. - Corinna Cortes has over 100,000 citations and is recognized for her work on support vector machines and the MNIST dataset [21][22]. - Vladimir Vapnik, with over 335,000 citations, is known for his foundational work in statistical learning theory [27]. Group 4: Broader Implications - The article suggests that the Scaling Law is not a sudden insight but rather a cumulative result of interdisciplinary research spanning decades, from psychology to neural networks [34][43]. - The evolution of the Scaling Law reflects a broader scientific journey, with contributions from various fields and researchers, ultimately leading to its current understanding in deep learning [43].