David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience

在这期播客节目中，主持人索尼娅（Sonia）迎来了精英人工智能实验室ineffable intelligence的两位联合创始人戴夫·西尔弗（Dave Silver）和莱斯·埃斯佩霍尔特（Lese Espeholt）。讨论的重点是他们雄心勃勃的使命：通过强化学习构建一个“超级学习者”，并在不依赖人类数据或大语言模型（LLM）产品作为中间步骤的情况下，实现与超级智能的“首次接触”。 戴夫·西尔弗因其在博士期间以及在DeepMind将强化学习应用于围棋游戏的开创性工作而闻名，他回顾了人工智能的发展历程。他指出，尽管他最初对大语言模型在经过广泛的预训练后能在小型数据集上进行微调的速度之快感到惊讶，但他的核心信念依然没有改变：通往人工超级智能最有前景的路径，是通过通过强化学习和经验自行发现事物的系统。戴夫解释说，创办ineffable是因为他想在一个不受严重聚焦于基于大语言模型的产品的组织约束的环境中，全力以赴地投入到这种特定的方法中。莱斯·埃斯佩霍尔特曾在DeepMind和谷歌大脑（Google Brain）花多年时间在GPU上扩展强化学习，他分享了自己在与早期投资者接触并与熟悉的面孔重聚后如何加入这家初创公司的经历。 嘉宾们定义了他们的核心概念。“超级学习者”是一个在深度和广度上都拥有令人难以置信能力的系统——从对一切一无所知开始，就像人类大脑中有一团“灰色浆糊”一样，并在模拟、数字或物理领域中自主发现远远超出人类能力的知识。为了培养这一点，他们正在构建“熔炉”（crucibles），这些是具有巨大涌现复杂度容量的模拟环境（例如复杂的物理引擎或多智能体交互），智能体可以在其中自主发现技术、科学、数学和语言。 对话的一个中心主题是有意消除人类先验。将此与AlphaZero进行比较——AlphaZero丢弃了人类数据，并在数小时内在国际象棋和围棋中击败了人类指导的迭代版本——创始人认为，依赖海量的人类知识数据集会拖累系统并锁定历史偏见。ineffable没有用面包屑式的线索引导智能体走向诸如人类语言等预定目标，而是赋予智能体在复杂环境中的简单、开放式优化目标，从而允许复杂行为和新颖工具自然涌现。 讨论还涵盖了基础设施和安全。莱斯详细介绍了他们与谷歌云平台（Google Cloud Platform）和英伟达（Nvidia）的紧密合作，强调了他们获得即将投入使用的海量GPU集群的权限，并利用了源自他过去工作（如Impala和Seed RL）的规模化强化学习架构。在对齐和安全方面，戴夫强调了在模拟中迅速发展“超级智能科学”的重要性。通过研究超级智能智能体如何在受控的虚拟世界中和平互动，他们的目标是在安全地把关真实世界应用之前，建立基础性的洞察。 最后，戴夫分享了他的超级预测家预测，估计ineffable将在一年内拥有极其强有力的证据，证明他们的方法走在正确的轨道上，并预测由于加速的研究能力，真正的超级智能将在三年内出现。

In this episode of the podcast, host Sonia welcomes Dave Silver and Lese Espeholt, two co-founders of the elite AI lab ineffable intelligence. The discussion focuses on their ambitious mission to build a "super learner" through reinforcement learning and to make "first contact" with super intelligence without relying on human data or large language model (LLM) products as intermediate steps. Dave Silver, known for his pioneering work applying reinforcement learning to the game of Go during his PhD and at DeepMind, reflects on the journey of AI. He notes that while he was initially surprised by how rapidly LLMs fine-tuned on small datasets after extensive pre-training, his core conviction remains unchanged: the most promising path to artificial superintelligence is through systems that discover things for themselves via reinforcement learning and experience. Dave explains that ineffable was started because he wanted to give this specific approach its all, outside the constraints of an organization heavily focused on LLM-based products. Lese Espeholt, who spent years scaling up reinforcement learning on GPUs at DeepMind and Google Brain, shares how he joined the venture after connecting with an early investor and reuniting with familiar faces. The guests define their core concepts. A "super learner" is a system possessing incredible capacity in both depth and breadth—starting from knowing nothing at all, much like humans with a "gray mush" in their heads, and autonomously discovering knowledge that goes far beyond human capabilities across simulated, digital, or physical domains. To foster this, they are building "crucibles," which are simulated environments with massive capacity for emergent complexity (such as complex physics engines or multi-agent interactions) where agents can autonomously discover technology, science, mathematics, and languages. A central theme of the conversation is the deliberate removal of human priors. Comparing this to AlphaZero—which discarded human data and outperformed human-guided iterations in chess and Go within hours—the founders believe that relying on massive datasets of human knowledge weighs systems down and locks in historical biases. Instead of "breadcrumbing" agents toward predetermined objectives like human language, ineffable gives agents simple, open-ended optimization goals within complex environments, allowing complex behaviors and novel tools to emerge organically. The discussion also covers infrastructure and safety. Lese details their strong partnership with Google Cloud Platform and Nvidia, highlighting their access to a massive upcoming GPU cluster and utilizing scaled-up RL architectures derived from his past work (such as Impala and Seed RL). On alignment and safety, Dave emphasizes the importance of developing a "science of super intelligence" rapidly in simulation. By studying how super intelligent agents interact peacefully in controlled virtual worlds, they aim to build foundational insights before safely gating real-world applications. Finally, Dave shares his super forecaster predictions, estimating that ineffable will have extremely strong evidence within a year that their approach is on the right track, and forecasting true super intelligence within three years due to accelerated research capabilities.