読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · 面壁智能密度法则重大3 分で読めるDisconfirmAI 編集部

Liu Zhiyuan, Chief Scientist at ModelBest, Proposes "Density Law": Open-Source Large Model Intelligence Density Doubles Approximately Every 100 Days

Liu Zhiyuan argues that the core evolution of large models is not merely an expansion of parameters, but the formation of a general technical route capable of continuously absorbing new types of data; as data and computing power become new constraints, increasing "intelligence density" is more sustainable than simply scaling up. He draws an analogy to Moore's Law, estimating that intelligence density doubles approximately every 100 days, and believes that edge AI will become a crucial carrier for AI's entry into the physical world.

2026-10-31 11:31 GMT+8

Liu Zhiyuan, Co-founder and Chief Scientist of ModelBest (面壁智能), presented a testable hypothesis at the 8th Intelligent Driving Conference held in Suzhou, China: the intelligence density of global open-source large models increases by approximately one doubling every 100 days, meaning that half the parameters and half the computational cost can achieve similar intelligence capabilities as before every 100 days. He termed this observation the "Density Law."

Liu Zhiyuan believes that the core change in large models over the past eight years has not been solely the continuous expansion of parameter scale, but rather the formation of a general technical route capable of continuously absorbing different types of data. However, as models continue to expand, data and computing power are becoming new constraints. On the data side, starting around 2024, large-scale reinforcement learning has driven models from imitation learning toward exploratory learning, enabling models to actively explore and generate new training data; on the computing side, if capability improvements rely solely on expanding parameter scales, computational demands will ultimately become unsustainable. Therefore, large models need to increase "intelligence density," i.e., achieving the same level of intelligence with fewer parameters and less computing power.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

Liu Zhiyuan analogizes the Density Law to Moore's Law in the semiconductor industry: in the future, models will continue to shrink on one hand, while terminal computing power for devices such as smartphones, cars, and robots will continue to improve on the other. These two trends will eventually converge, allowing large models that previously could only run on servers to gradually enter terminals. According to the Density Law, if intelligence density continues to double approximately every 100 days, within less than two years, the number of parameters and computing power required for equivalent capabilities may drop by two orders of magnitude, compressing some capabilities of originally trillion-parameter models into the hundred-billion range and enabling deployment on terminal devices.

He categorizes future AI into three forms: super intelligence, cloud-side intelligence, and edge-side intelligence, among which edge-side intelligence will become an important carrier for physical AI. There are three additional trends for artificial intelligence entering the physical world: high intelligence density, which reduces the parameters and computing power required for model operation; full-modal perception, which understands the real world through vision, sound, and other sensors; and autonomous agents, which continuously make decisions, learn, and act in open environments.

ModelBest has already applied edge-side models to scenarios such as smartphones, automotive cockpits, satellites, and robots, and is collaborating with companies including Huawei and Samsung. Liu Zhiyuan believes that in the future, large models will continue to enter terminals such as mobile phones, PCs, cars, robots, wearable devices, and IoT devices, making edge-side intelligence a critical pathway for AI to enter the physical world.

ソース

  • leiphone.com↗
トピック

面壁智能密度法则

面壁智能与清华大学联合团队提出的“密度法则”(Densing Law)指出,大语言模型的智能密度呈指数级增长,约每3.5个月翻倍。这意味着每隔3.5个月,即可用一半参数量的模型实现当前最优性能。该理论最初于2024年12月在arXiv发布,…

5 ステップ1 件更新日 2026-11-01 20:58

経緯

  1. 2024-12-05面壁智能与清华大学团队在arXiv发布论文《Densing Law of LLMs》,首次提出大模型能力密度随时间指数增长的规律。arxiv.org ↗
  2. 2025-11-01该研究成果正式发表于Nature子刊《自然·机器智能》,确认开源大模型最大能力密度约每3.5个月翻一倍。infoq.cn ↗
  3. 2026-04-03美国研究机构METR更新技术报告,显示AI能力任务时长跨度约每88.6天翻倍,与密度法则预测高度吻合。eu.36kr.com ↗

未解決の問題

  • 密度法则的加速趋势是否会因高质量公开语料枯竭而放缓?
  • 不同架构(如MoE vs Dense)对能力密度倍增周期的具体影响差异有多大?
  • 端侧设备算力增长能否持续跟上模型小型化的速度以实现真正的普惠?
トピック全文を読む →
リサーチを続ける →