Engram 联合创始人兼 CEO Dan Biderman 一边烹饪地中海风味牛肉丸，一边分享他从以色列海军特种部队到创立 AI 研究公司的历程。
Dan Biderman, co-founder and CEO of Engram, shares his journey from Israeli naval special operations to founding an AI research company, while cooking Mediterranean beef meatballs.
要点 · TL;DR
效率与智能不可分割，少做多得才能解决更难的长期任务。 Efficiency and intelligence are inseparable; doing more with less enables harder long-horizon tasks.
未来企业数据将达到互联网规模的万亿级 token，需要基于训练的压缩。 Future enterprise data will reach internet-scale trillions of tokens, requiring training-based compression.
个性化持续学习将在设备端运行，用户拥有并培育自己的模型权重。 Personalized continual learning will run on-device, with users owning and nurturing their model weights.
核心观点 · Key points
效率与智能不可分割；用更少做更多才能解决更难的长期任务。 Efficiency and intelligence are inseparable; doing more with less enables solving harder long-horizon tasks.
未来企业数据将达到互联网规模的数万亿 token，需要基于训练的压缩方法。 Future enterprise data will reach internet-scale trillions of tokens, requiring training-based compression.
模型应自主决定哪些知识存入权重、哪些保留文本，用户反馈可学习但不依赖。 Models should learn autonomously what to store in weights vs. text, with user feedback but not dependence.
当前上下文窗口存在“上下文腐烂”——即使达到千万级，更多 token 也会降低准确性。 Current context windows suffer from 'context rot'—more tokens reduce accuracy even at 10M scale.
个性化持续学习将在设备端运行，用户拥有并培育自己的模型权重。 Personalized continual learning will run on-device, with users owning and nurturing their model weights.
反共识 · Contrarian takes
预训练将整个互联网压缩进权重的魔法应应用于企业知识。 Pre-training's magic of compressing the web into weights should be applied to enterprise knowledge.
参数高效微调（如 LoRA）在整体性查询上比 RAG 更能代表知识。 Parameter-efficient fine-tuning (e.g., LoRA) can represent knowledge better than RAG for holistic queries.
压缩是有损且确定性的；权重中的神经记忆痕迹是必要的补充。 Compaction is lossy and deterministic; neural memory traces in weights are a necessary complement.
模型路由尚未解决；最佳方案是将路由与可信的“副手”模型结合。 Model routing is unsolved; the best solution combines routing with a trusted 'sidekick' model.
增加算力的 Scaling 并非唯一路径；下一范式需要以更少做更多。 Scaling with more compute is not the only path; the next paradigm requires doing more with less.
本期章节 · Chapters（共 11）
开场与介绍Opening and introduction
动机与背景Motivation and background
18个月知识规模Knowledge Scale in 18 Months
长周期智能体与上下文限制Long Horizon Agents and Context Limitations