Industry

Google DeepMind 研究员 Tim Genewein 撰文提出通往超级智能的四条路径

Why we picked it

DeepMind 研究员 Tim Genewein 把通往超级智能的路线拆成四条:堆数据、找新范式、大规模智能体协作、递归自我改进。最有信息量的是算力限制挡不住其他路径这个判断,做监管或投资判断的人都该看看原文。

Google DeepMind 研究员 Tim Genewein 在 DeepMind Institute 新文章中表示 AGI 已"在可及范围内",并提出通往超级智能的四条路径:更大的模型与数据、新范式、数百万协同智能体、递归自我改进。

Body

他估计有效算力每年至少增长 10 倍,十年内可运行的当前智能体副本数量将超过地球人口。文中还给出一个成本推算:如今花费 10 亿美元算力的工作,届时成本约降至 1000 美元。

Genewein 特别指出,仅限制训练算力这类单一输入无法阻止进展,因为其余路径仍在持续推进。这篇文章代表了 DeepMind 内部对 AGI 之后路线的思考,也暗示监管若只针对算力,效果可能有限。

Original · kimmonismus

Holy, feel the acceleration: Google DeepMind researchers are already thinking past AGI:

In a new DeepMind Institute essay, Tim Genewein says AGI is "coming within reach" and lays out four paths to superintelligence: -bigger models and data, -new paradigms, -millions of coordinated agents, and -recursive self-improvement.

- Effective compute grows at least 10x per year. In ten years you could run more copies of today's agents than there are people on Earth.

- Work that costs $1 billion in compute today would cost around $1,000.

And: restricting one input like training compute won't stop progress, because the other paths keep moving.

So yeah, ASI it is. Looks like the path is clear and everything is within reach. Im so excited :)

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and AI regulation, compliance frameworks, and sovereign compute policy. Key entities and technical topics under continuous tracking include Google DeepMind, AGI, agents, reasoning models.

Source attribution: curated from kimmonismus (@kimmonismus).