Ineffable Intelligence：用经验学习打造超级学习者Ineffable Intelligence: Building a Superlearner Through Experience
The Takeaway：真正的超级智能不会来自堆砌人类数据的LLM，而会来自能从零开始、在复杂环境中自主发现知识并持续学习的超级学习者。
David Silver（AlphaGo核心人物、前DeepMind强化学习负责人）与Lasse Espeholt（强化学习规模化先驱）共同创办Ineffable Intelligence，目标是打造能从任何经验流中无限学习的超级学习者，并与超级智能实现首次接触。Silver坚信从经验中学习才是通往超级智能的正确路径，而非当前主流的LLM预训练+后训练路线。他指出，人类集体从空白大脑发现语言、科学与技术的过程，正是AI尚未破解的关键。Ineffable放弃人类先验数据，转而构建能涌现复杂行为的“坩埚”环境，让智能体在简单目标下自然涌现出技术、数学甚至语言。Silver回忆AlphaZero在放弃人类数据后，仅用数小时就超越了两年人工迭代的AlphaGo，证明“放手人类先验”能解锁更强能力。团队约20人，拥有大规模GPU集群，专注研究导向文化，预计一年内看到强证据，三年内实现真正超级智能。
“Collectively, we started with this kind of gray mush in our head... Everything we learned, we discovered it for ourselves. And that's an incredible thing that no one has cracked yet in AI.”
David Silver（AlphaGo核心人物、前DeepMind强化学习负责人）与Lasse Espeholt（强化学习规模化先驱）共同创办Ineffable Intelligence，目标是打造能从任何经验流中无限学习的超级学习者，并与超级智能实现首次接触。Silver坚信从经验中学习才是通往超级智能的正确路径，而非当前主流的LLM预训练+后训练路线。他指出，人类集体从空白大脑发现语言、科学与技术的过程，正是AI尚未破解的关键。Ineffable放弃人类先验数据，转而构建能涌现复杂行为的“坩埚”环境，让智能体在简单目标下自然涌现出技术、数学甚至语言。Silver回忆AlphaZero在放弃人类数据后，仅用数小时就超越了两年人工迭代的AlphaGo，证明“放手人类先验”能解锁更强能力。团队约20人，拥有大规模GPU集群，专注研究导向文化，预计一年内看到强证据，三年内实现真正超级智能。
“Collectively, we started with this kind of gray mush in our head... Everything we learned, we discovered it for ourselves. And that's an incredible thing that no one has cracked yet in AI.”
The Takeaway: True superintelligence will not come from LLMs stacked with human data, but from a superlearner that can start from nothing, discover knowledge autonomously in complex environments, and keep learning forever.
David Silver (AlphaGo pioneer and former DeepMind RL lead) and Lasse Espeholt (pioneer of large-scale RL) co-founded Ineffable Intelligence to build a superlearner capable of infinite learning from any stream of experience and make first contact with superintelligence. Silver has always believed learning from experience is the path to superintelligence, not the dominant LLM pre-training and post-training approach. He points out that humanity’s collective discovery of language, science and technology from blank minds remains the key unsolved challenge in AI. Ineffable deliberately abandons human prior data and instead builds “crucibles”—environments rich in emergent complexity—where agents pursue simple objectives and spontaneously invent technology, mathematics and even language. Silver recalls that once AlphaZero dropped human data, it surpassed two years of human-in-the-loop AlphaGo work in just hours, proving that letting go of human priors unlocks far greater capability. The ~20-person team, backed by a large GPU cluster, runs a research-led culture and expects strong evidence within a year and true superintelligence within three years.
“Collectively, we started with this kind of gray mush in our head... Everything we learned, we discovered it for ourselves. And that's an incredible thing that no one has cracked yet in AI.”
查看原文 →
David Silver (AlphaGo pioneer and former DeepMind RL lead) and Lasse Espeholt (pioneer of large-scale RL) co-founded Ineffable Intelligence to build a superlearner capable of infinite learning from any stream of experience and make first contact with superintelligence. Silver has always believed learning from experience is the path to superintelligence, not the dominant LLM pre-training and post-training approach. He points out that humanity’s collective discovery of language, science and technology from blank minds remains the key unsolved challenge in AI. Ineffable deliberately abandons human prior data and instead builds “crucibles”—environments rich in emergent complexity—where agents pursue simple objectives and spontaneously invent technology, mathematics and even language. Silver recalls that once AlphaZero dropped human data, it surpassed two years of human-in-the-loop AlphaGo work in just hours, proving that letting go of human priors unlocks far greater capability. The ~20-person team, backed by a large GPU cluster, runs a research-led culture and expects strong evidence within a year and true superintelligence within three years.
“Collectively, we started with this kind of gray mush in our head... Everything we learned, we discovered it for ourselves. And that's an incredible thing that no one has cracked yet in AI.”