📄 本篇文章深度整理自播客《人民公园》，深入探讨了在顶级闭源大模型（如 Claude Fable 5、GPT-5.6）日益昂贵、审核严格以及出现“降智/切脑”倾向的当下，普通用户和中小企业所面临的困境。文章指出，通过将国产性价比模型（如智谱 GLM、Kimi）接入 Claude Code、Codex 等优秀开源或现有 CLI 工具链生态，不仅能够实现大幅度的成本削减，还能突破使用门槛，并对 AI 创业的前景、免费与付费模式的博弈进行了多维度的思辨与展望。
📄 Dwarkesh Patel argues that continuous learning is the key bottleneck for AGI, not raw intelligence. Current LLMs fail to learn persistently like humans, hindering their ability to become true long-term employees. While reasoning capabilities are improving, the lack of adaptive learning prevents AI from automating many white-collar jobs. He predicts significant AI capabilities by 2028 and true continuous learning by 2032, emphasizing that solving continuous learning could lead to rapid AGI development.
📄 文章对比了Claude Opus 4.7、Gemini 3.1 Pro和GPT-5.4在通用能力上的最新评测结果，三者并列第一。文章指出，尽管通用能力趋同，模型在具体场景下各有专长：Claude Opus 4.7擅长执行复杂工作任务，Gemini 3.1 Pro在科研和科学推理方面领先，GPT-5.4则在长周期编程和批判性思维方面占优。作者建议用户根据具体需求选择合适的模型，并提到了Claude Opus 4.7的新功能如任务预算。