Recording: May 20 · Organized: July 16, 2026_Audio duration about 3 hours 44 minutes_
核心贡献 · Key contributions
提出以愿景驱动、基于共识的组织方式，以克制而非利润最大化为核心准则来追求 AGI。 Proposes vision-driven, consensus-based organization guided by restraint rather than profit maximization to pursue AGI.
提出 API 定价规则：十个月内收回硬件成本，让用户用得起，而非追求收入最大化。 Introduces an API pricing rule: recover hardware costs within ten months, ensuring affordable access without maximizing revenue.
阐明 AGI 发展路线：语言模型→思维链→智能体→持续学习→奇点→具身智能。 Articulates an AGI roadmap: language model to chain of thought, agent, continual learning, singularity, embodied intelligence.
主张在有限利润下，开源与商业化不矛盾，并能提升实现 AGI 的概率。 Argues open source is compatible with commercialization under limited profit margins and increases probability of achieving AGI.
将 C 端与 B 端应用视为 AGI 研究的副产品，从而形成‘降维打击’式的竞争优势。 Treats consumer and enterprise applications as byproducts of AGI research, yielding a 'dimensionality reduction' competitive advantage.
报告自研高层语言编译器 TileLang 及国产芯片适配，可能绕开 Nvidia 的 CUDA 生态。 Reports a high-level compiler (TileLang) and domestic chip adaptation that may bypass Nvidia's CUDA ecosystem.
局限 · Limitations
算力差距：约 2 万张 H 等效 GPU，无法训练 800B 参数模型，Scaling（规模扩张）受限。 Compute gap: roughly 20,000 H-equivalent GPUs and inability to train 800B-parameter models cap scaling.
愿景驱动与共识式管理可能脆弱，团队稳定被视为最大内部风险。 Vision-driven consensus management may be fragile, with team stability identified as the biggest internal risk.
开源可持续性依赖薄利润与需求缺乏弹性；继续降价未必带来更多社会价值。 Open source sustainability hinges on a thin profit margin and inelastic demand; lower prices may not increase societal value.
路线判断（如将世界模型和视频生成排除在智能主线外）仍是推测性的。 Roadmap judgments, such as excluding world models and video generation from the intelligence mainline, remain speculative.
芯片依赖仍存在：华为超节点相当于‘四张顶一张 Nvidia GPU，且落后两年’。 Chip dependence persists: Huawei super nodes are 'four cards for one Nvidia GPU and two years behind'.