以下是内容的中文翻译： 里奇·萨顿（Rich Sutton），强化学习领域的先驱以及《苦涩的教训》（The Bitter Lesson）一书的作者，近日与他的联合创始人库拉姆·贾维德（Khuram Javed）共同探讨了他长期以来对人工智能的愿景以及他们的新公司——橡树实验室（Oak Lab）。萨顿常被贴上“激进”的标签，但他坚称：“我不是异类。这个领域才是异类”，尤其是在其当前对学习的理解上。他认为“所有的学习都是持续性的”，而“持续学习”作为一个独立的子领域之所以出现，仅仅是因为近期“人工智能领域的一些异常现象”偏离了基本原则。 萨顿职业生涯中对强化学习的执着，即使经历了“人工智能寒冬”和个人健康挑战，也是源于他坚信“学习是心智的核心部分”。他将《苦涩的教训》描述为一个长期存在的观察：人工智能的进步源于“能够随计算规模扩展的学习方法”，而非通过嵌入静态人类知识。尽管先验知识和学习在理想情况下应是盟友，但在实践中，过度依赖静态人类知识往往因忽视真正的学习而阻碍了进展。萨顿极力倡导学习，因为他觉得其他人过分强调静态知识，并指出当前的大型语言模型（LLM）在部署后并不真正“学习”，它们仅仅是调整上下文。 将“苦涩的教训”应用于大型语言模型，萨顿看到了积极和消极两方面。大型语言模型通过利用计算能力“汲取互联网数据”实现了显著的规模扩展，但它们对有限的互联网数据和训练后的静态权重的依赖使其成为一个反面例子。库拉姆·贾维德和萨顿强烈批评合成数据生成是“一个大错误”。他们通过“大世界假说”（Big World Hypothesis）论证，世界“比任何智能体都复杂得多”。合成数据本质上是一个“小世界”，由人类专业知识创建并受其瓶颈限制，缺乏现实世界经验的无限复杂性。他们建议智能体应该从*自身*经验中学习，建立和完善*自己的*模型和模拟器，而不是依赖人类筛选的数据。 萨顿强调，动物和婴儿的学习本质上是经验性的，而非监督式的。他将正规学校教育视为一种非典型且新兴的学习形式，认为智能的本质并非由人类提供目标。人类通过积累的知识和文化取得卓越成就，但真正的范式转变需要智能体“通过自身经验”，而非仅仅继承人类知识，来发现新的抽象概念。他们认为，当前人工智能中一个关键的缺失环节是，缺乏“学习一个模型然后利用该模型进行规划”的能力，并运用自主发现的抽象概念，这种能力在复杂的物理环境中导航的动物身上可见一斑。 橡树实验室旨在解决这些根本性差距，其驱动力是萨顿的“阿尔伯塔人工智能研究计划”（Alberta Plan for AI Research）。该计划的核心是“持续深度学习”（continual deep learning），它旨在解决“灾难性遗忘”（catastrophic forgetting）问题。他们的解决方案涉及算法改进，例如“步长优化”（step size optimization，其中每个权重都具有通过元学习获得的、独立调整的步长）和“持续反向传播”（Continual Backprop），后者在标准反向传播的同时，不断注入新的、随机初始化的单元。这使得算法能够“元学习如何学习”，让系统能够从头开始并持续学习而不破坏先验知识。尽管承认当前存在技术挑战，例如实现一个万亿参数模型在20瓦功率下运行的目标，但他们相信由于算法效率和持续的计算进步，这在5-10年内是可实现的。他们认为，当前大型实验室“陷入局部最优”，只专注于扩展现有范式，这使得它们难以追求真正新颖、更高效的方法。 橡树实验室的最终目标是创造一个“自我维护”和“自我一致”的心智，能够涵盖从低级感觉运动技能到高级抽象概念的完整知识谱系，并统一处理它们。他们设想的是一个单一的*设计*，可以产生众多心智，每个心智都由其经验独特塑造，而非一个单一的全知系统。橡树实验室正在组建一个小型、高度协作的团队，以推动人工智能的这一基础性转变。

Rich Sutton, a pioneer in reinforcement learning and author of "The Bitter Lesson," along with his co-founder Khuram Javed, recently discussed his long-held vision for AI and their new venture, Oak Lab. Sutton, often labeled "radical," asserts, "I'm not weird. The field is weird," particularly in its current understanding of learning. He contends that "all learning is continual," and the concept of "continual learning" as a distinct subfield only arose due to a recent "AI craziness" that deviates from fundamental principles. Sutton's career dedication to reinforcement learning, even through an "AI winter" and a personal health battle, was driven by the conviction that "learning is a central part of the mind." He describes "The Bitter Lesson" as a long-standing observation: AI progress stems from "learning methods that will scale with computation," not by embedding static human knowledge. While prior knowledge and learning should ideally be allies, in practice, over-reliance on static human knowledge has often hindered progress by dismissing genuine learning. Sutton champions learning because he feels others over-emphasize static knowledge, pointing out that current large language models (LLMs) don't *learn* after deployment, they merely adjust context. Applying "The Bitter Lesson" to LLMs, Sutton sees both positive and negative aspects. LLMs achieved significant scaling by leveraging computation to "drink in the internet," but their reliance on finite internet data and static weights after training makes them a negative example. Khuram Javed and Sutton strongly criticize the use of synthetic data generation as a "big mistake." They argue, via the "Big World Hypothesis," that the world is "massively more complex than any agent." Synthetic data is inherently a "small world," created and bottlenecked by human expertise, lacking the infinite complexity of real-world experience. They suggest agents should learn from *their own* experience, making and refining *their own* models and simulators, rather than relying on human-curated data. Sutton emphasizes that animal and baby learning is inherently experiential, not supervised. He dismisses formal schooling as an atypical and recent form of learning, arguing intelligence's essence isn't about human-provided targets. Humans excel by building on accumulated knowledge and culture, but true paradigm shifts require an agent to discover new abstractions "through its own experience," not just inherited human knowledge. A critical missing piece in current AI, they argue, is the ability to "learn a model and then plan with the model" using self-discovered abstractions, a capability seen in animals navigating complex physical environments. Oak Lab aims to address these fundamental gaps, driven by Sutton's "Alberta Plan for AI Research." Key to this plan is "continual deep learning," which tackles "catastrophic forgetting." Their solution involves algorithmic advancements like "step size optimization" (where each weight has a meta-learned, individually tuned step size) and "Continual Backprop," which continually injects new, randomly initialized units alongside standard backpropagation. This allows algorithms to "meta-learn how to learn," enabling systems to start from scratch and learn continually without destroying prior knowledge. While acknowledging the current technological challenges, such as the goal of a trillion-parameter model running on 20 watts, they believe this is achievable within 5-10 years due to algorithmic efficiency and continued computational advancements. They argue current large labs are "stuck in a local minimum" of scaling existing paradigms, making it hard for them to pursue genuinely new, more efficient approaches. The ultimate ambition for Oak Lab is to create a "self-maintaining" and "self-consistent" mind capable of a full spectrum of knowledge, from low-level sensorimotor skills to high-level abstractions, treating them uniformly. They envision a single *design* that can lead to many minds, each uniquely shaped by its experience, rather than a single omniscient system. Oak Lab is building a small, highly aligned team to pursue this foundational shift in AI.