Ineffable：从纯经验打造超级学习者，通往超级智能的唯一路径Ineffable: Building a Superlearner from Pure Experience as the Path to Superintelligence
The Takeaway：真正的超级智能不会来自人类数据预训练的LLM，而来自能从零开始、在复杂环境中自主发现知识的超级学习者。
David Silver（AlphaGo与AlphaZero核心人物、前DeepMind）与Lasse Espeholt共同创立Ineffable Intelligence，明确放弃LLM路径。他们坚信：人类从灰色脑浆开始集体发现语言、科学、数学的过程，才是超级智能的本质。超级学习者需要深度（从一无所知到超越人类知识）与广度（可投入任何模拟、数字或物理经验流）。他们正在构建“坩埚”（crucibles）——充满涌现复杂性的环境，让智能体在简单奖励下自然涌现技术、语言与科学，而非被人类先验牵着鼻子走。
Silver回忆AlphaZero：放弃人类棋谱后，仅用几小时就超越了两年人工迭代的AlphaGo。他直言当前LLM公司都在“半自主+人类闭环”里打转，必须主动抛弃万亿级人类数据，才能发现真正能持续学习的算法。对齐方面，他们主张先在模拟中建立超级智能科学，搞清楚多智能体和平互动的条件，再决定是否进入真实世界，并承诺以模拟证据作为门控。
团队仅20人，却拥有GCP上最大规模集群之一，计算力人均极高。Silver预测：一年内出现强证据，三年内实现真正超级智能。
“那就是通往超级智能的本质——一个能永远持续学习与发现知识的系统。”
David Silver（AlphaGo与AlphaZero核心人物、前DeepMind）与Lasse Espeholt共同创立Ineffable Intelligence，明确放弃LLM路径。他们坚信：人类从灰色脑浆开始集体发现语言、科学、数学的过程，才是超级智能的本质。超级学习者需要深度（从一无所知到超越人类知识）与广度（可投入任何模拟、数字或物理经验流）。他们正在构建“坩埚”（crucibles）——充满涌现复杂性的环境，让智能体在简单奖励下自然涌现技术、语言与科学，而非被人类先验牵着鼻子走。
Silver回忆AlphaZero：放弃人类棋谱后，仅用几小时就超越了两年人工迭代的AlphaGo。他直言当前LLM公司都在“半自主+人类闭环”里打转，必须主动抛弃万亿级人类数据，才能发现真正能持续学习的算法。对齐方面，他们主张先在模拟中建立超级智能科学，搞清楚多智能体和平互动的条件，再决定是否进入真实世界，并承诺以模拟证据作为门控。
团队仅20人，却拥有GCP上最大规模集群之一，计算力人均极高。Silver预测：一年内出现强证据，三年内实现真正超级智能。
“那就是通往超级智能的本质——一个能永远持续学习与发现知识的系统。”
The Takeaway: True superintelligence will not come from LLMs pretrained on human data, but from a superlearner that discovers knowledge autonomously from pure experience in complex environments.
David Silver (AlphaGo/AlphaZero pioneer, former DeepMind) and Lasse Espeholt co-founded Ineffable Intelligence to pursue pure reinforcement learning from experience, deliberately rejecting the LLM path. They argue that the human journey—from gray mush to collective discovery of language, science, and mathematics—is the essence of what AI still lacks. A superlearner must deliver both depth (from zero knowledge to beyond human) and breadth (droppable into any stream of simulated, digital, or physical experience). They are building “crucibles”—environments rich in emergent complexity—so agents can invent technology, language, and science under simple rewards rather than being led by human priors.
Silver recalls AlphaZero: after discarding human game data, the system surpassed two years of human-in-the-loop AlphaGo work in mere hours. He sees today’s LLM labs as stuck in semi-autonomous, human-in-the-loop loops; only by actively abandoning trillion-token human datasets can we discover algorithms that keep learning forever. On alignment, they prioritize developing a science of superintelligence inside simulation—understanding the conditions for peaceful multi-agent interaction—before any real-world deployment, with simulation evidence as an explicit gate.
With a deliberately small team of ~20 and one of the largest GCP clusters, compute-per-person is extremely high. Silver’s forecast: strong evidence within a year, true superintelligence within three.
“That is the essence of what it will take to get to superintelligence, like a system that can just keep learning and discovering knowledge forever.”
查看原文 →
David Silver (AlphaGo/AlphaZero pioneer, former DeepMind) and Lasse Espeholt co-founded Ineffable Intelligence to pursue pure reinforcement learning from experience, deliberately rejecting the LLM path. They argue that the human journey—from gray mush to collective discovery of language, science, and mathematics—is the essence of what AI still lacks. A superlearner must deliver both depth (from zero knowledge to beyond human) and breadth (droppable into any stream of simulated, digital, or physical experience). They are building “crucibles”—environments rich in emergent complexity—so agents can invent technology, language, and science under simple rewards rather than being led by human priors.
Silver recalls AlphaZero: after discarding human game data, the system surpassed two years of human-in-the-loop AlphaGo work in mere hours. He sees today’s LLM labs as stuck in semi-autonomous, human-in-the-loop loops; only by actively abandoning trillion-token human datasets can we discover algorithms that keep learning forever. On alignment, they prioritize developing a science of superintelligence inside simulation—understanding the conditions for peaceful multi-agent interaction—before any real-world deployment, with simulation evidence as an explicit gate.
With a deliberately small team of ~20 and one of the largest GCP clusters, compute-per-person is extremely high. Silver’s forecast: strong evidence within a year, true superintelligence within three.
“That is the essence of what it will take to get to superintelligence, like a system that can just keep learning and discovering knowledge forever.”