Engram Bakes Knowledge Into AI Model Weights To Cut Token Costs 100x

Engram 将知识融入 AI 模型权重，将 Token 成本降低 100 倍 直接更新模型权重可消除冗余推理计算并实现真正的持续学习，其表现优于 RAG 与上下文窗口扩展。 Dan Biderman 和 Jessy Lin 认为，当今 AI 的真正瓶颈并非底层基础智能，而是记忆与持续学习能力。Engram 对向上下文窗口堆砌长提示或单纯依赖 RAG 的主流范式提出挑战，转而采用持续适配器微调方案，将团队专属知识直接融入模型权重。该方法压缩了 KV 缓存状态，大幅削减了推理成本，并构建出能作为不断演进的认知伙伴的个性化、持续更新模型。Engram 设想的未来并非依赖单一庞大的前沿模型，而是每个人和组织都能运行属于自己的、具备持续学习能力的专用模型生态系统。 Engram 认为当前 AI 的真正瓶颈并非基础推理能力，而是记忆与持续学习，主张将团队专属领域知识直接内化写入模型权重，而非单纯依赖不断堆叠的上下文窗口或 RAG。 (0:00) 通过结合 LoRA、前缀调优、SFT 与强化学习等轻量化适配架构，将企业环境与规范预先固化进权重，可在推理时将系统提示词与检索相关的 Token 消耗降低最高达 100 倍。 (4:00) 针对事实与概念学习深度交织的特性，Engram 引入离线“做梦”（Dreaming）机制，让模型在沙盒环境中进行自主模拟与边界用例测试，将短期上下文沉淀为长效权重。 (8:00) 不同于前沿实验室打造单一庞大通用 AGI 模型的路线，Engram 致力于构建支持数百万个私有化、持续自我更新的个性化模型生态，打造新型企业“神经数据平面”（neural data plane）。 (15:00) 针对当前单篇文章 KV 缓存可膨胀至 80GB HBM 的基础设施瓶颈，Engram 探索离线压缩等技术，目标将缓存占用缩减近 1000 倍，带来类似“一夜之间真正变聪明的 AI 实习生”般的记忆爆发时刻。 (30:00)

Engram co-founders Dan Biderman and Jessy Lin argue that AI's primary bottleneck is not foundational reasoning but continuous adaptation to evolving private contexts, which cannot be solved by static context windows alone. They advocate for continuous training that bakes domain knowledge directly into lightweight model weights alongside external retrieval. Limits of Context Stuffing and KV Caching (0:00) Lin and Biderman observe that externalized retrieval methods like RAG and massive prompt stuffing become computationally unsustainable as human output approaches tens of millions of tokens daily per person. While external search remains useful for transient details, repeatedly forcing generic models to re-read large document repositories creates severe latency and inference cost bottlenecks. Biderman highlights an infrastructure bottleneck in key-value (KV) caching: storing the active KV cache for a single Wikipedia article on a 70-billion-parameter Llama model can consume roughly 80 gigabytes of High Bandwidth Memory, nearly matching the approximately 100 gigabytes needed to host the entire model's weights. To overcome these hardware limits, Engram argues for burning compute offline to compress context representations into weights, targeting a 1,000x reduction in cache load and up to a 100x decrease in inference token consumption. Furthermore, Biderman and Lin assert that weight-internalized memory circumvents the "addressing problem" of traditional RAG, allowing models to form spontaneous, proactive associations without requiring users to craft explicit search queries. Adapter Fine-Tuning and Offline Dreaming Consolidations Rather than relying on test-time file parsing, Engram trains per-team models integrated directly into enterprise workspaces like Notion, Microsoft, and Harvey. The technical architecture combines lightweight adapter fine-tuning—utilizing LoRAs, prefix tuning, and sparse architectures—with post-training pipelines including supervised fine-tuning, reinforcement learning, and on-policy distillation. The founders state that this strategy capitalizes on a recurring 3-to-6-month capability gap during which frontier foundation models lag behind bespoke, company-specific workflows. Addressing the theoretical division between abstract reasoning and factual storage, Lin notes that fact retention and skill execution are interdependent in cognitive architectures, warning that stripping facts impairs core reasoning. To assimilate new information without catastrophic forgetting, Engram incorporates an offline "dreaming" phase modeled on biological sleep. In this mode, models operate in sandboxed environments away from user traffic, running self-directed simulations to consolidate edge cases and domain rules into stable weight updates. Personal Model Fleets Versus Centralized Monoliths (15:00) Engram's operational philosophy contrasts with the dominant frontier lab roadmap centered on scaling monolithic models for general benchmarks. Lin points out that high-value operational knowledge is frequently confidential, subjective, or conflicting across different organizations, making it incompatible with standardized post-training datasets and standard reinforcement learning environments that depend on clear ground-truth rewards. Biderman contends that while major AI labs currently treat memory as an application wrapper around pre-trained models, continual learning requires an independent research discipline that unifies modeling with live product deployment. Instead of relying on a single centralized system to govern all tasks, Engram forecasts an ecosystem populated by millions of continuously updating, personalized models adapted to individual and team-level habits. Skills Portability and the Enterprise Neural Data Plane (34:55) Regarding user digital profiles, Lin remains skeptical of a unified "memory wallet," noting that users require firm context segregation between their personal activities and corporate workspaces. However, Biderman proposes a framework for "skills portability": while enterprise proprietary data must remain quarantined behind corporate firewalls, the heuristics, workflows, and adapter-level problem-solving methods an individual refines could theoretically transfer between jobs, similar to professional human expertise. Over a 5-to-10-year horizon, the founders project that enterprise software will transition from static file storage to an associative "neural data plane." Biderman envisions Engram serving an architectural role analogous to enterprise database providers like Oracle or Databricks, but engineered specifically to maintain dynamic neural network states rather than traditional relational records. While Engram positions continuous adapter training as an alternative to context engineering, unresolved technical challenges remain. The founders acknowledge that alternative state-space architectures force persistent trade-offs between memory footprint and accuracy, and modifying factual knowledge inside neural weights without destabilizing existing representations remains an active area of research.

Engram 联合创始人 Dan Biderman 与 Jessy Lin 认为，当前前沿大模型的核心瓶颈并非基础推理能力，而在于无法持续吸收演进中的特定上下文；他们主张打破依靠外部检索与静态长上下文的传统补丁，通过轻量级微调将动态记忆直接内化至模型权重，构建持续学习的个性化模型生态。 外部检索的物理瓶颈与权重内化收益 (0:00) Dan 和 Jessy 指出，虽然检索增强生成（RAG）与上下文工程仍有其应用场景，但单纯依赖在推理阶段塞入海量上下文正面临严峻的算力和物理瓶颈。随着协作团队每日生成的 Token 量迈向数千万级，在多轮对话中重复检索并重新解析庞大文档会导致推理成本剧增。Dan 以 700 亿参数的 Llama 模型为例说明，仅加载单篇维基百科长文的 KV Cache 就需占用约 80GB 的 GPU 高带宽内存（HBM），而整个模型权重本身仅占约 100GB，这证明梯度下降在信息压缩上的效率远超上下文加载。Engram 提出通过离线计算将这些表示压缩约 1000 倍。 此外，权重内化能规避传统 RAG 固有的“寻址难题”：用户在复杂任务中往往无法预先提炼出精确的搜索关键词，而内化了组织知识的模型无须依赖显式检索或冗长的智能体调用循环，即可自发形成跨领域概念的抽象联想。创始人估计，通过前期投入微调算力将团队上下文直接烧录至权重中，可消除冗长系统提示词与重复检索，从而使推理阶段的 Token 消耗降低多达两个数量级（约 100 倍）。 轻量级适配微调与离线“做梦”消化 为实现工作空间内的持续适应，Engram 目前避开分散的个人移动设备，重点与 Notion、Microsoft 和 Harvey 等高上下文密度的协作平台合作，在原生环境中为特定团队训练专属模型。其技术方案基于 LoRA、前缀调优（Prefix Tuning）与稀疏架构等轻量级适配器，并融合监督微调（SFT）、强化学习（RL）以及同策略蒸馏（On-policy Distillation）。Biderman 认为，由于前沿模型在通用预训练覆盖特定领域长尾场景上存在约 3 至 6 个月的周期性滞后期，这种针对具体工作流与内部专有工具链的微调填补了关键空缺。 在认知机制上，Lin 反对将事实记忆与概念技能人为剥离的做法，强调两者在深度学习中高度交织，去除具体事实往往会导致基础推理能力受损。借鉴人类睡眠的神经整合机制，Engram 在模型架构中引入了脱离即时交互的离线“做梦”阶段。在该阶段，模型在沙盒环境中自主运行模拟实验、探索工具交互逻辑，并将白天的短时上下文经验和长尾边缘案例消化并固化为长期的权重参数，以此缓解灾难性遗忘并提升生产环境下的应对能力。 分布式个人模型对抗中心化巨石架构 (15:00) Engram 的核心哲学与当前主流前沿实验室形成鲜明分歧。Dan 与 Jessy 认为，主流前沿机构将主要目标设定为通用人工智能（AGI），倾注资源于更大参数、海量数据和测试时算力，以攻克国防、密码学或宏观自动化的中心化巨石模型；在这些机构中，记忆与持续学习多被视为应用层功能或封装外壳，而非核心研究瓶颈。Engram 则倡导并行的“第三条路径”：构建数以百万计的、不断自我更新的个性化小模型网络。 Lin 强调，人类社会最核心的知识往往是高度私有、主观甚至相互冲突的，例如企业特有的审批规则或个人的排版偏好。这类数据既不会进入公共预训练集，也因目标模糊而无法提供干净的奖励信号以供标准强化学习进行监督。因此，持续学习要求打破传统机器学习中研发与产品交付之间割裂的流程，建立由用户即时交互直接驱动权重更新的研发生态闭环。 记忆边界隔离、技能迁移与神经数据平面 (34:55) 针对未来模型记忆是否应统一为跨应用“记忆钱包”的构想，Jessy 持保留意见。她认为用户天然需要在工作空间、个人生活及特定工具之间维持严格的上下文隔离，避免跨场景建议带来体验干扰。然而，Dan 提出了“技能可移植性”的设想：企业专有数据和知识产权必须受到防火墙保护并留存在机构内部，但用户在长期人机协作中磨练出的操作直觉、启发式策略与解决问题范式，应当能够通过经过净化的技能适配器被个人随身带走，类似于人类离职时遵循商业保密规范但保留工作技能。 展望未来 5 至 10 年，Dan 结合计算神经科学中海马体与内嗅皮层对空间导航与情境记忆的统一机制，将 Engram 的终极形态定位为企业级“神经数据平面”（Neural Data Plane）。他认为，未来的企业架构不会仅停留在静态文件系统或原始向量库的查询层，而是会由类似 Databricks 或 Oracle 地位的神经基础设施来支撑，通过维系海量个性化、持续学习的模型权重，以高联想度、低冗余度的“大脑状态”动态映射组织的全部知识。 Engram 提出的权重内化与持续微调范式在理论上大幅削减了长上下文的推理与显存开销，但在工程落地中仍面临权衡：如何在没有明确真实奖励信号的主观业务场景中自动化评估微调质量，以及如何在模型快速迭代中精确剥离企业私有数据与个人可迁移技能，仍是待解的实际挑战。

## Background 本期内容由红杉资本（Sequoia Capital）的 Sonya Huang 与 Shaun Maguire 主持，于 2026 年 6 月 24 日在播客及网络平台发布。节目邀请了专注研究“记忆与持续学习”的创新实验室 Engram 联合创始人 Dan Biderman 与 Jessy Lin，探讨如何打破不断堆砌上下文窗口与外挂 RAG 的局限，将企业与团队知识直接内化至模型权重中。 ## TL;DR 对话深入探讨了 Engram 提出的反共识路线：大模型能力的真正瓶颈不在于原始推理智力，而在于记忆与持续学习。Engram 致力于通过参数高效微调（PEFT）和离线“做梦”消化等机制，将团队私有文档与交互反馈深度固化到模型权重中，从而实现推理 Token 消耗最高降低 100 倍。两位创始人结合计算神经科学背景，对比了生物大脑的模糊压缩与当前臃肿的 KV Cache，论证了事实与技能无法割裂的底层机理。他们认为前沿实验室执念于通用 AGI，而未来的真正范式是“千人千模”——每个人和组织都将拥有专属、去中心化且隔夜变聪明的神经接口底座。 ## Atmosphere & style 对谈整体洋溢着深厚科研底蕴与前瞻思辨碰撞的火花，既有高密度的计算神经科学与系统架构剖析，又穿插着极具极客特色的幽默调侃。主持人与嘉宾之间互动张弛有度，从自嘲式探讨“民间假说”到对前沿实验室垄断神话的犀利解构，现场笑声与思想交锋频现。整场对话既像是一场严谨的认知科学沙龙，又透露出开拓者试图颠覆主流 RAG 与长上下文范式的锐气与笃定，听来令人耳目一新且发人深省。 ## Bulletin - `05:32` Engram 主张跳出 RAG 与超长上下文工程的堆叠模式，通过持续学习将企业与团队知识直接内化写入模型权重，推理端 Token 消耗最高可降低达 100 倍。 - `02:58` 团队正与 Notion、微软及 Harvey 等平台合作，基于开源 Transformer 利用 LoRA、前缀调优及稀疏架构等技术，为企业打造如入职数年老员工般的专属定制模型。 - `30:00` 针对 Transformer 的 KV Cache 显存暴涨痛点（如 70B 模型单篇长文占用达 80GB 显存），Engram 提出通过离线计算与梯度精炼实现千倍级状态压缩，从底层改写推理效率。 - `31:37` 记忆真正的“ChatGPT 时刻”并非拼接更长的 Prompt，而是模型拥有类似人类“做梦”的离线消化能力，让专属模型隔夜之间真正变得更加聪明。 ## Key takeaways - `00:00` 当前大模型的真正瓶颈不是通用常识与原始智力，而是理解演进中的全新上下文环境，记忆与持续学习本是同一枚硬币的两面。 - `00:00` 依赖外部检索和上下文堆叠面临严重成本与认知瓶颈，未来个人与企业每天产生数千万甚至数亿 Token，重复检索在成本和混乱度上均不可承受。 - `02:58` Engram 通过 LoRA、前缀调优、稀疏架构微调结合强化学习与同策略蒸馏，将工作空间沉淀的交互与文档深度固化为直觉权重。 - `05:32` 在回答全局组织架构与优先级等隐式问题时，权重内化模型仅需约 100 Token 即可精确作答，相比前沿大模型的 10 万 Token 上下文检索节约达两个数量级。 - `10:00` 记忆具体事实与高阶抽象概念不可割裂；若模型在每一步思考链条都停下来外部检索基础事实，其思维深度就无法走向长程复杂推理。 - `10:00` 传统计算机科学将数据库与算法生硬分离，而深度学习的本质正是将事实与算法交织融合；重构 AI 必须打破外部上下文与通用模型的机械割裂。 - `12:35` 引入类似人类“做梦”的离线反思与同策略合成阶段，让模型在脱离即时交互后将当天的经验沉淀巩固进参数，远胜于在测试时仓促进行 Prompt 推理。 - `15:00` 前沿实验室押注单一全能的 AGI 与通用推理算力，而 Engram 坚信未来是“人人拥有专属模型”，私密领域的数据与因人而异的价值定义无法靠中心化公开后训练解决。 - `21:40` 状态空间模型（SSM）等超长上下文架构本质上都在显存与模型精度之间妥协；贯彻“苦涩的教训”的最佳解法，是将更多算力直接烧在全新上下文的训练上。 - `24:45` 事实记忆与业务技能在底层没有鸿沟；通过轻量微调直接将专属规范刻入权重，跳过冗长 REPL 工具调用与检索，构成“RAG 杀手”式架构。 - `28:30` Prompt Caching 仅能优化精确前缀复用，无法像神经权重那样产生自发的抽象跨域联想，也无法解决传统 RAG 在未知探索中不知从何搜起的“寻址困境”。 - `30:00` KV Cache 暴食显存反衬出梯度下降惊人的压缩比；通过离线计算精炼大脑状态实现 1000 倍体积压缩，是颠覆推理显存瓶颈的核心方向。 - `33:45` 未来可能出现“脱敏后的技能包”形式的记忆钱包，让员工在不侵犯雇主数据资产的前提下，将个人训练沉淀出的 Agent 能力带入职业生涯流转。 - `42:40` 借鉴海马体中空间导航与情境记忆同源的神经机制，Engram 的长期愿景是成为对标 Oracle/Databricks 的“人类数据平面的神经模型接口”。 ## Detailed walkthrough ### `00:00` 始终在训练：超越预训练与上下文工程的瓶颈 红杉资本（Sequoia Capital）的 Sonya Huang 与 Shaun Maguire 共同主持了本期 Training Data 对话，嘉宾为 Engram 的联合创始人 Dan Biderman 与 Jessy Lin。Engram 是一家专注于“记忆（Memory）”与“持续学习（Continual Learning）”的创新实验室（Neolab），而这两大方向正是当下 AI 研究领域最具争议也最核心的焦点。访谈伊始，Sonya 引用了 Engram 官网上的标语——“我们不从预训练或后训练的视角看待世界，我们的模型始终在训练”，并询问这句话背后的真正含义。 Jessy Lin 指出，当今的顶尖大模型在通用常识和推理能力上已经极为聪明，但限制模型发挥更大实际价值的瓶颈并不是“原始智力（Raw Intelligence）”，而是“理解全新且不断演进的上下文环境”。无论是员工接手的一项新任务，还是具体职位的业务语境，行业当前的做法往往是通过提示词或上下文工程（Context Engineering）进行临时拼接。而 Engram 认为，记忆与持续学习本就是“同一枚硬币的两面”：真正的目标应当是如何像预训练和后训练那样，将新知识深入地固化到模型权重之中。 针对“外挂数据库和上下文窗口算不算真正记忆”的质疑，Jessy 与 Dan 明确表示，工具调用和上下文拼接固然有其价值，但这两种模式绝非终极答案。Dan 形象地比喻道：对人类个体而言，随手记便利贴和笔记本不可或缺，但第二天回到工位时，大脑中必然已经沉淀了某种记忆痕迹，形成了新的直觉与方向感。纯靠外部检索面临两大约束：首先，每个人未来每天产生的数据量动辄达到数千万甚至数亿 Token，反复检索、重读如此庞大的上下文在成本上是不可承受的；其次，若缺乏深层认知架构，将海量杂乱文档一股脑塞入上下文只会让模型陷入混乱。 ```mermaid flowchart TD accTitle: 传统上下文外挂与权重内化方案对比 accDescr: 展示了将数据输入到模型上下文与将数据直接训练进权重两种模式的路径差异 subgraph External["传统方案：外部上下文堆叠 (RAG / Context Engineering)"] A["海量业务文档 / 交互历史"] --> B["向量检索 / 外部存储"] B --> C["拼装超大 Prompt (数万至数十万 Token)"] C --> D["单次推理 (重复读取、易产生幻觉与混淆)"] end subgraph Internal["Engram 方案：权重内化持续学习"] E["团队协作 / 反馈流"] --> F["适配器微调 (LoRA / 稀疏架构)"] F --> G["深度内化至模型权重"] G --> H["极简 Prompt 高效推理 (节省高达 100x Token)"] end ``` --- ### `02:58` 团队级权重内化：Engram 的架构与技术路线 Dan 和 Jessy 阐释了 Engram 的产品机制与技术落地方式。Engram 目前正与 Notion、微软（Microsoft）以及 Harvey 等平台深度合作，这些工作空间沉淀了企业跨越数月乃至数年的核心上下文，包括历史文档、协同交互记录以及用户向 Agent 提供的实时反馈。Engram 致力于在这些工作空间内为每个团队训练专属模型，使其不仅仅是在测试时“临时翻看文件”，而是像“在一家公司工作多年的资深老员工”一样，对全公司的战略方向、做事风格、招聘流程及特定业务惯例了然于胸。 在技术实现层面，Engram 并不主张把一切全部塞入权重，而是构建一套自适应的权重更新与能力微调体系： - **参数高效微调（PEFT）**：运用包括 LoRA、前缀调优（Prefixes）以及稀疏架构（Sparse Architectures）在内的多种适配器技术，以轻量化方式持续学习。 - **数据管线与信号转化**：结合监督微调（SFT）、强化学习（RL）和同策略蒸馏（On-policy Distillation），将原始文档与即时交互转化为有效的梯度更新信号。 - **内化与外挂的边界划分**：模型假定某些知识和工具依然保持外部化，关键难题在于精准界定“哪些需要被刻入直觉内化，哪些适合外部调用”。 - **适配专有定制工具**：不同企业往往拥有极其独特的内部工作流，并非都能套用前沿实验室默认优化的标准 Bash CLI 工具，内化训练能让模型真正掌握团队独有的操作逻辑。 - **模型支持机制**：由于需要白盒访问权限（White-box Access）来直接调节权重，Engram 当前主要基于开源 Transformer 架构展开，但同样可与开放模型权重的闭源厂商开展合作。 --- ### `05:32` 推理 Token 削减 100 倍与团队级记忆的落地优势 在讨论前期训练算力与实际收益的权衡时，Dan Biderman 揭示了一个极其直观的数据指标：通过预先投入少许算力将团队知识写入权重，企业在推理端可以获得**高达 100 倍（两个数量级）的 Token 消耗削减**。 Dan 指出，面对诸如“公司团队结构与当前核心优先级是什么”这类问题，往往无法在单一规整的文档中直接检索到，前沿大模型依靠庞大的系统提示词和上下文检索，往往需要消耗 100,000 个 Token 才能梳理清楚；而经过专属适配器训练的模型，由于已经隐式内化了全局协作背景，往往只需花费 100 个 Token 即可直接给出精确答案。此外，前沿大模型的通用能力更新与垂直领域的个性化探索之间始终存在 3 到 6 个月的能力滞后期，自主、轻量的本地权重持续学习正好填补了这一窗口。 当被问及为何首先切入“工作空间（团队级）”而非“个人设备级”时，Dan 解释道： - 团队级协作拥有沉淀多年、整理更严谨且密度极高的知识资产，是当前信息密度最大的富矿。 - 从团队空间起步更容易收敛和跑通数据飞轮。 - Engram 长期来看必将拓展至每个人的电脑与手机，让个人终端上的专属模型具备伴随式学习能力。 --- ### `07:11` 概念抽象与记忆解耦：什么值得被记住？ 针对 Sonya 提出的经典争论——“大模型将‘法国首都是巴黎’这类事实硬背在参数权重中，究竟是一种特性还是一种缺陷？模型是否应该只抽象学习国家与首都的概念框架，而把事实留给外部存储？”——Jessy Lin 给出了深度回应。 Jessy 认为，记忆与高阶抽象之间并非水火不容，因为记忆基础事实是进一步组合出复杂抽象概念的基石。当前体系的真正痛点在于“大模型缺乏判断什么信息值得被记住的能力”。许多学术基准测试（Academic Benchmarks）热衷于考核模型能否背诵极其冷门的偏门事实（例如某非洲国家某座桥梁的具体长度），这完全是对模型参数容量的无谓浪费，人类大脑也不会为这类琐碎事实分配容量。人类记忆之所以是“有损（Lossy）”的，正是因为大脑演化出了一套精妙的筛选机制，仅把最具泛化价值的表征烙印在神经回路中。 ### `10:00` 事实与算法的融合：为什么必须将知识内化到模型权重中 在智能的核心定义中，极其关键的一项能力是“压缩重要信息，并将其与无关噪音剥离”。Jessy Lin 指出，很多人倾向于将“事实学习”（fact learning）与“技能/非事实学习”（skill learning）割裂开来，甚至有人尝试剥离掉模型中的全部具体事实，试图提炼出一个纯粹的推理核心，但这种做法极其违背自然规律——失去了对基础事实的认知，模型根本寸步难行。当被问及“为什么不能把事实留给外部检索、只让模型负责推理”时，Jessy 解释道，如果模型在思考链条的每一步都必须停下来调用外部检索去回忆最底层的常识，它的思维深度就无法走远；唯有将信息真正“内化”（internalize）进参数结构中，模型才能够将基础知识组合为更高阶的抽象概念，进行复杂且深度的思考。 Dan Biderman 从计算机科学的历史沿革进一步剖析了这一范式转移。在传统的计算机体系中，数据库（Database）与算法（Algorithm）分属两套独立的课程与逻辑：数据库负责存储和查询关于世界的客观事实（如各国首都），而算法则负责高效操纵信息并得出答案。深度学习的奇妙与神秘之处恰恰在于它将这两者强行“揉碎融合”在了一起，以至于像 Anthropic 这样的可解释性研究团队不得不耗费巨量心智去尝试拆解其内部机理。然而，当前 AI 在工业界和经济体系中的落地实践却出现了一种倒退——人们再次把它们生硬切开：企业小心翼翼地维护着自己外部的私有上下文（Context），而上面悬挂着一个对其具体业务一无所知的通用通用模型（Generic model）。Dan 强调，尽管将持续变动的动态事实直接烙印在权重中存在挑战，但让事实、故事与细节以某种节奏重新与模型自身融合，才是通向更高智能的必然收敛方向。 ```mermaid flowchart TD accTitle: 知识架构对比：传统分离模式与 Engram 深度内化模式 accDescr: 对比传统外部检索分离架构与参数内化架构的差异 subgraph External["当前主流模式：事实与算法割裂"] DB["企业外部上下文 / 数据库<br/>(RAG、Prompt 工程)"] GenModel["通用基础模型<br/>(算法/推理核心，完全不了解业务)"] DB -.->|"测试时单次注入 (无法深度组合)"| GenModel end subgraph Internal["深度学习与 Engram 模式：知识深度内化"] Weights["私有模型权重矩阵<br/>(事实与算法深度交织)"] Abstract["抽象概念深度组合与多步长程推理"] Weights --> Abstract end ``` ### `12:35` “做梦”隐喻与离线消化：超越测试时计算的训练魔力 为了打破这种“测试时临时抱佛脚”的局限，Dan 和 Jessy 提出了类似人类“做梦”的认知机制。当前的模型交互形态几乎把所有开销与努力都堆砌在测试时（Test time）：模型临时扫过用户扔进来的 Prompt 上下文，仓促进行即兴推理；这种机制不仅上限极低，而且极易在此过程中频频出错。真正的突破在于引入一个“缺席的阶段”——模型在结束即时交互之后，能够退回到安静的离线状态，将白天的上下文与交互痕迹深度消化、反思并吸收回参数当中。Dan 笑称，宣称要“造出像人类做梦一样的 AI”听起来有些离经叛道，但人类梦境的底层逻辑正是如此：在梦中，人类与自我对话，在极度偏向社交和复杂情境的环境中反复推演边界与可能性，并应对那些长尾极端的假设；给模型留出“做梦”的时间，就是让它离线测试自身在特定环境下的行动能力，从而在下一次面对用户时变得更加笃定与成熟。 在被问及支持这一路线的典型案例时，Dan 给出了一个极具启发性的思维实验： * **国际奥数竞赛的假设场景**：假设 OpenAI 被要求在短短一周时间内赢得一场国际数学奥林匹克竞赛（Math Olympiad），他们会作何选择？他们是会召集一大批标注人员去编纂所有数学教材的目录、精细标定检索哪些章节与图表来做 RAG 检索？还是会迅速整合资料、合成高质量训练数据，果断启动一次为期五至六天的训练任务，并在最后关头评估模型的内化能力？ * **训练的核心优势**：任何真正动手训过大模型的研究者都清楚，将跨领域的思想与能力融会贯通的最佳途径，永远是“训练的魔力”（Magic of training）。 * **愿景下沉**：前沿实验室已经在数学、编程和网络安全等顶尖垂直领域验证了这套训练法则，而 Engram 的使命就是把这种深度的训练魔法普及到成千上万的企业与普通人手中。 ### `15:00` 差异化的世界观：去中心化的私有模型与前沿实验室的 AGI 执念 面对“为什么最终产品不会被头部基础模型实验室通吃”的质疑，Jessy 指出 Engram 与前沿实验室在根本世界观上存在分歧。头部实验室追求的是一个通晓万物、规模越来越大、在全领域通吃的超级大模型（AGI），而 Engram 设想的未来图景则是“人人拥有自己的专属模型”。现实世界中大量极具价值的知识本质上是私密的，永远不可能被清洗整合进公开的 Post-training 数据集里；更重要的是，不同组织乃至不同个体之间对于“何为优秀”的定义存在天然冲突——某家公司的业务逻辑与行为偏好，可能与另一家完全相反。几十年来机器学习赖以生存的工具套件建立在“极其干净的监督信号”和“客观的奖励函数”（Ground-truth reward）之上，这在有清晰编译反馈的代码任务中表现尚可，但面对现实世界里充满模糊性、因人而异的需求时便彻底失效。 Dan 进一步指出，前沿实验室的核心优先级（P0）极其单一明确：倾注绝大部分算力、资金与顶级人才，通过“更多预训练、更大参数、更海量数据、更强 RL 以及更高的推理时计算（Inference-time compute）”，强行叩开 AGI 的大门，以期彻底实现宏观经济层面的自动化，或攻克密码学、国防等国家级硬核命题。尽管这些实验室也关心记忆与持续学习，但在他们眼中那更偏向于一种次级的产品功能（Product feature），而非必须孤注一掷的核心科研攻坚： * **顶级共识与赛道空白**：DeepMind 首席执行官 Demis Hassabis 此前在红杉资本的活动中明确直言，AI 若想更上一层楼，必须在“记忆与持续学习”领域迎来全新的范式突破；前沿实验室受困于 AGI 路线无法全神贯注，而这正是 Engram 独立重注的领域。 * **数据激励与归属权**：通过去中心化的方式让模型持续从广泛的人类与组织交互中学习，同时确保数据无需离开所有者手中、不必派驻工程师贴身驻场，将成为释放企业生产力的巨大解题钥匙。 ### `17:15` 持续学习的三大支柱与产品-研究闭环 要把“千人千模、持续进化”的世界变为现实，现有的技术范式必须经历全方位重塑。Jessy 将这一变革拆解为不可或缺的三大支柱： * **核心科研范式的突破**：攻克持续学习与灾难性遗忘的底层理论难题，寻找更优的记忆整合机制。 * **面向微型模型的训练基础设施**：过去的基础设施全是为了“单次超大规模集群运算”（One big run）而设计；新的世界需要一种能够同时为数百万用户持续微调、高效训练个性化小模型的全新分布式底座。 * **打破隔阂的产品-科研闭环**：传统前沿实验室的研发流程极其割裂，研究员在象牙塔中把模型训练好，便将其“扔过围栏”（Throw it over the fence）交给产品团队，后者再通过 Prompt 和上下文工程在其表面糊上一层产品外壳。但在模型时刻都在训练的新范式下，用户的每次真实输入与反馈都直接构成了驱动模型进化的训练信号，产品设计与底层研究必须深度咬合在一起。 在谈及创立 Engram 的原初驱动力时，Jessy 表示，在通用大模型已经极其聪明的今天，AI 领域最激动人心的未解之谜不再是单纯堆砌原始智力，而是“在边界处学习”（Learning at the edges）——即如何让模型精准捕获那些决定其实际生产力价值的剩余知识。当下的预训练大模型之所以能展现出惊人的涌现与泛化能力，很大程度上像是互联网在无意间“赐予人类的一份礼物”，海量异构文本碰巧孕育了这些奇迹，但学术界至今未能完全透彻理解其背后的机理。解开持续学习的谜团，实质上是倒逼科研人员彻底搞懂模型泛化的真正诱因，并主动驾驭这一过程：让模型在面对一家企业的私有数据资产时，能够像当初吞咽整个公网数据一样，举重若轻地将其内化吸收。 ### `20:00` 从计算神经科学到“自然愚蠢”：生物记忆与机器存储的本质鸿沟 Dan Biderman 的学术背景深深植根于计算神经科学。在转向人工智能前，他最初着迷于人类意识和人类生存境况等终极命题。当被问及现有的 AI 模型是否拥有意识时，Dan 坦率而幽默地表示自己并没有什么超出大众认知的洞见，他认为目前的模型并不具备意识，但“有聪明人在思考这个问题非常重要”。他的切入点更偏向于认知机制本身——人类究竟如何思考与感知。他引用了已故以色列认知心理学家阿莫斯·特沃斯基（Amos Tversky）的名言自嘲：“特沃斯基常说，他感兴趣的不是人工智能（Artificial Intelligence），而是‘自然愚蠢’（Natural Stupidity）。” 顺着探究人类与动物如何感知物理世界的轨迹，Dan 逐渐转向统计学与机器学习领域。他敏锐地察觉到，生物体每天都在处理的记忆与持续学习（Continual Learning）难题，正是当前 AI 系统最亟待解决的瓶颈，而现有技术方案与生物学机制相去甚远。他并不盲目崇拜仿生学，也不认为机器必须全盘复刻人脑——毕竟计算机在很多任务上早已超越人类，例如无损存储整个庞大的代码库并随时调用，普通计算机靠传统文件系统就能做到，根本用不着 AI。 但人类大脑之所以独特，在于它是在极其严苛的物理信息容量限制下演化出来的。人脑并不是死记硬背原始数据，而是构建出一种“模糊表征”（Fuzzy Representations）。这种表征能够在更高维度被抽象化、自动建立不同记忆之间的网络连接，并在睡眠或次日巩固沉淀。Dan 指出，现有的主流大模型除了最初那次耗资巨大的通用预训练阶段外，完全缺乏这种动态内化与抽象关联的机制，而这正是 Engram 致力于从计算神经科学中汲取灵感并着手攻克的核心命题。 --- ### `21:40` 算力苦涩教训与架构之争：SSM 的权衡与“烧算力”的新范式 关于计算机架构与大脑结构的对比，Shaun Maguire 抛出了一个深具哲学意味的疑问：现代计算机拥有 CPU、GPU 以及各类协处理器，大脑同样拥有不同的功能功能分区；在理查德·萨顿（Rich Sutton）著名的“苦涩的教训”（The Bitter Lesson）视角下，LLM 最终会收敛为一个大一统的主导计算单元（所有计算都发生在类似 GPU 的大模型核心中），还是说模型内部会自然涌现出专门负责记忆等任务的协处理器？记忆到底是一种单纯随着算力扩展和数据增加而自然出现的“涌现特性”（Emergent Property），还是必须依赖某种独立的外部专用架构？ ```mermaid flowchart TD accTitle: 传统长上下文方案与持续学习新范式的对比 accDescr: 展示长序列注意力机制与状态空间架构在内存和精度间的权衡，对比 Engram 将算力投入新上下文训练的路径。 subgraph Hardware_Tradeoff["上下文处理架构权衡"] A["Transformer 自注意力"] -->|"二次方计算复杂度 O(N²)"| B["显存开销极大 / 精准度高"] C["状态空间架构 (SSM)"] -->|"次二次方或线性复杂度"| D["显存占用低 / 精度妥协不可避免"] end subgraph Engram_Approach["Engram 苦涩教训范式"] E["海量全新业务上下文"] --> F["将算力转化为模型参数权重"] F --> G["免去单次推理重复塞入长提示词"] G --> H["实现本能级联想与高精度"] end ``` Dan 从工程落地与模型底层架构两个层面作出了剖析： - **工程部署现状**：真实的 AI 落地早已超越单纯的 GPU 运算，行业正在大量使用代码沙箱环境，模型在外部计算机上不断交互试错。 - **模型底层权衡**：Dan 曾深入参与过状态空间架构（State Space Models，如 S4、Mamba 等系列）及其他超长上下文架构的研发。这类架构试图打破标准 Transformer 自注意力机制随着序列长度呈二次方（$O(N^2)$）暴涨的计算与显存开销，力求实现次二次方甚至线性复杂度（包括当前部分顶尖开源模型也在混合层中借鉴了此类设计）。 - **天下没有免费的午餐**：Dan 强调，团队在实际测试中发现，这类架构几乎无一例外都是在**计算机显存开销**与**模型精度**之间做妥协，永远伴随着精度损失。 - **彻底信奉“苦涩的教训”（Bitter-Lesson Pilled）**：Dan 明确表示团队绝不怀疑通用人工智能（AGI）的算力扩展大方向。真正的破局思路不应局限于拼命压榨序列长度算法，而在于“如何把更多的算力直接烧在模型从未见过的全新上下文上”。 - 对此 Dan 幽默地对 Shaun 打趣道：“如果真要让模型彻底理解 Shaun、理解你的全部工作和上下文，单靠每次把文件重新塞进上下文窗口读一遍是绝不可能做到的——特别是像你这么特殊的人物，我们必须为你专门训练一个 100 万亿参数的模型！”Shaun 随即大笑附和：“完全赞同（Co-sign）！” --- ### `24:45` 事实与技能本无分别：轻量化微调如何直击“RAG 杀手”核心 当谈及企业客户最渴望让模型吸收内化什么内容时——究竟是组织内的硬性事实、代码规范（如特定的 CI/CD 流水线流程），还是特定任务技能？Jessy Lin 指出，目前整个应用层开发者耗费绝大多数心力做的事情，本质上就是逼着通用大模型去适配具体的私有业务场景。最普遍的需求包括“按照企业特定的品牌视觉风格来设计网站”、“熟练执行公司内部固有的复杂工作流”或者“完全模仿特定用户的语言风格来撰写文案”。 在 Jessy 看来，所谓“事实记忆”（Facts）与“技能掌握”（Skills）在模型底层并无根本界限，Engram 所采用的持续内化技术对二者一视同仁。Dan 进一步直击当前行业痛点：当今市面上绝大多数应用层产品，本质上都只是拿一个通用前沿基座模型，外包了一层搜索工具、向量数据库和 REPL 循环。企业客户真正想要的，是用一种响应更快、运行更廉价且上下文契合度更高的方式来对接企业海量数据，而不是在每次对话时都逼着模型从头去检索一遍内部知识库。 > “我们正在向客户证明，通过相当轻量化的训练，这些知识和技能就能变成模型的‘本能’。模型根本不需要跑那些极其繁琐漫长的 REPL 循环。从某种意义上说，这就是一个‘RAG 杀手’（RAG Killer）式的方案。” Dan 澄清他们并不彻底排斥在某些特定场景下结合检索，但对于需要高频对接海量数据平面、重复处理类似业务逻辑的企业而言，将重复知识直接刻入模型权重，是实现大规模自动化的唯一高效解法。 --- ### `26:40` 酒店房号还是家门密码？生物记忆取舍与无需严苛过滤的学习 听到“RAG 杀手”这一论断，Shaun 忍不住打断并深入追问：“抱歉我可能有点刨根问底，但我还是没有完全吃透。我们是否面临着‘做 RAG’与‘更新模型权重’之间的二选一取舍？到底什么样的信息应该被刻入权重，什么样的信息应该外置给 RAG？” Dan 坦言这不仅在工程界是一个悬而未决的开放难题，也是认知神经科学领域的核心未解之谜——生物记忆系统同样在时刻面临“哪些记忆该内化、哪些该遗忘或借助外力记录”的选择： - **临时信息的处理**：你完全不需要在神经组织中永久内化一年前住过的某家酒店房间号，这类信息记在便签纸或手机里最合适； - **核心长期信息的处理**：而你当前自家的门锁密码，必须在接下来数年内深深印在大脑记忆中，以保证即时提取而不需要每次翻找记录。 ```mermaid flowchart LR accTitle: 记忆内化与外部检索的决策取舍 accDescr: 比较临时性信息通过 RAG 处理与核心高频信息通过权重内化的区别。 subgraph Input_Data["输入信息类型"] T["临时低频信息<br/>(例如：一年前住过的酒店房间号)"] P["高频关键上下文<br/>(例如：当前住址密码、团队专属规范)"] end subgraph Processing_Path["处理与存储路径"] T -->|"外部存储 / 检索"| RAG["RAG / 外部便签<br/>(按需低频查阅，节省神经开销)"] P -->|"轻量化持续微调"| Weights["刻入模型权重 (Weights)<br/>(形成本能反应，支持抽象联想)"] end ``` 在工程落地中如何将这种认知原则转化为产品逻辑，Engram 团队坚持采用一种尽可能摒弃人为硬编码规则（Heuristics）的哲学。Dan 强调，工程上最偷懒的做法是在数据管道上套一层粗暴的过滤规则，主观指定“这部分数据有用要训练，那部分没用直接扔掉”。但人类现实生活并非如此规整：我们每天刷 TikTok，被海量毫无营养的垃圾碎片信息轰炸，然而正常人脑并未因此彻底“跑偏”或崩溃发疯。同理，一个健壮的持续学习系统本身必须具备抗噪鲁棒性，能够在吸收混乱日常数据的同时，自主维持核心知识框架的稳定。 --- ### `28:30` 为什么 Prompt Caching 救不了检索：联想能力、直觉与权重记忆的不可替代性 Jessy 将话题拉回到产业界眼下最真实的阵痛上：伴随复杂 Agent 架构的普及，让多个智能体昼夜不停地在多轮推理循环中运行，带来的推理计算成本（Inference Cost）简直是天文数字。 这段讨论引发了一段主持人间充满张力的风趣调侃： - Shaun 故作惊讶地唱反调：“等等，高昂的推理成本难道不是大好事吗？！大量消耗 Token 是为了什么？Sonya 可是 Fireworks AI 的投资人，我最爱高推理量了！” - Sonya 会心大笑，Dan 也跟着打趣：“我们也爱推理！” - 但 Jessy 随即切中要害：当前巨大的推理浪费极其荒唐——在同一个团队内部，不同的员工、甚至同一个任务的前后步骤中，模型正在不厌其烦地一遍又一遍完整阅读同一批长文档。“现实中如果有个员工被问到‘你昨天在做什么’，他绝不会当场打开搜索框搜一遍自己的工作记录，他是靠脑子直接知道的。” Shaun 随即提出质疑：“但这难道不能通过 Prompt Caching（提示词缓存）来解决吗？” Jessy 与 Dan 明确反驳指出，Prompt Caching 只能缓解精确前缀匹配时的重复计算成本，但根本无法实现知识的“渐进式积累与网状联想”： - **无法产生无意识的抽象联想**：当模型仅仅依赖 RAG 或静态上下文缓存时，它只能被动回应“你让我搜什么我就搜什么”。而在真实的协作中，如果你看到团队成员正在进行某项特定研究，人类专家会凭直觉联想到某个高层次的抽象关联并提醒对方：“我注意到你在做这个，我联想到了另一件相关的历史方案，你或许需要了解一下”——这种跨维度的隐式联想甚至在用户提问前就已经发生，而这**只能在模型深层权重交互中产生**。 - **检索系统致命的“寻址困境”（Addressing Problem）**：在传统 RAG 和 AI 检索体系中，最艰难的瓶颈往往不是把信息存放在何处，而是**如何在提问时精准地寻址并表达查询需求**。当前的大语言模型缺乏某种本能性的“检索直觉”，在很多复杂的未知探索任务中，模型根本不知道自己该搜什么、该从哪里切入。如果 AI 系统始终被禁锢在关键词检索与被动读取文件的范式中，它就永远无法获得像资深人类专家那样洞若观火的业务直觉。 ### `30:00` KV Cache 的暴食怪兽与千倍压缩愿景 Dan Biderman 指出，Engram 的创立初心很大程度上源于对 KV Cache（键值缓存）以及推理阶段显存瓶颈的深刻反思。在现有的 Transformer 架构体系中，KV Cache 已经变成了一个令人瞠目的“怪兽（monstrosity）”。举个极端的例子：对于一个 70B 参数的 LLaMA 模型而言，仅仅为一篇介绍 Taylor Swift 的维基百科文章缓存上下文，就会在 GPU 上吃掉整整 80 GB 的 HBM 显存。相较之下，整个 70B 模型的全部静态权重加在一起也不过大约 100 GB——而这 100 GB 的权重通过某种有损压缩，却记住了几乎整个互联网的知识。 这种鲜明的信息密度反差展现了一个直观的事实：梯度下降拥有令人惊叹的信息压缩能力，能够将浩瀚的数据塞进极少量的数字参数中；反观当下的 KV Cache，仅仅几十 KB 的纯文本在经过多层注意力投影后，就被膨胀成了数十 GB 笨重不堪的“大脑状态（brain state）”。 虽然业界目前通过显存-磁盘交换（disk-to-HBM offloading）等系统工程手段来缓解这一瓶颈，但 Engram 构想了一条完全不同的路径：能否投入一些离线计算资源，对这 80 GB 的状态进行高倍率压缩，使其在载入显存时体积直接缩小 1000 倍？这种千倍级别的状态精炼不仅将彻底重塑推理加载速度和系统吞吐量，也将从底层改变模型对信息的内部表征保真度。 ```mermaid flowchart TD accTitle: KV Cache膨胀与离线压缩对比 accDescr: 对比传统Transformer的KV Cache显存占用暴涨与Engram离线精炼压缩的思路 subgraph Current["当前KV Cache困境"] RawText["数万字节文本<br/>(例如Wikipedia文章)"] --> InefficientAttn["注意力层展开计算"] InefficientAttn --> MassiveCache["膨胀为 ~80 GB 显存占用<br/>(接近70B模型自身100GB的静态权重)"] end subgraph EngramVision["Engram的解法"] MassiveCache -.-> OfflineCompute["离线计算与梯度精炼压缩"] OfflineCompute --> CompactState["千倍体积缩减 (~1000x Smaller)<br/>极速加载与高效显存利用"] end ``` --- ### `31:37` 记忆的“ChatGPT时刻”：从实习生到越用越聪明的专属模型 当主持人询问未来一到两年内是否存在某种属于“记忆”维度的“ChatGPT 时刻”时，Jessy Lin 表达了她的见解。行业内外谈论终身学习（continual learning）已经很久，大家心目中那个最理想的原型其实就像带一个实习生：你日复一日地指导它、教它新事物，而它确实能随时间推移变得越来越聪明。 当前业内无论采用多么精密的上下文工程（context engineering）技巧，都无法真正触碰这一愿景。真正的“ChatGPT 时刻”，必然是某一天用户由衷地发出一声惊叹：“这家伙真的变聪明了，它和昨天完全不一样！” Dan Biderman 补充强调，当年的 ChatGPT 问世前其实并未被广泛准确预见。从个体的角度来看，假设你今天辞职并立志打造一个最懂你个人的专属 AI，哪怕你 24 小时不停歇地为 OpenAI 或 Anthropic 的前沿基座模型编写上下文提示词，你能撬动的能力杠杆依然极其有限，甚至还不如坐等下一代基座版本发布。Engram 试图开拓的未来，是让用户在特定领域投入的每一分交互时间和精力，都能稳定转化为模型针对该领域真实性能的跃升。实现这一目标极为艰难，其核心突破点在于：必须在不破坏原有泛化能力（避免灾难性遗忘）的前提下，实现计算量与私有领域数据训练的规模化扩展。 - **实习生模型隐喻**：终身学习的临界点并非靠更长的 Prompt 堆砌，而是模型权重具备真正的吸纳与内化能力，实现“隔夜变聪明”。 - **单兵上下文工程的局限**：在现有黑盒 API 模式下，用户手工打磨 Context 的天花板极低，边际收益迅速递减。 - **算力与数据的连续演进**：唯有在保持模型基础智力不受损的前提下，将新增数据与算力融入权重，才能让个人或企业的专属投入沉淀为复利资产。 --- ### `33:45` AI 惊艳时刻、个人记忆钱包与技能便携性 进入快问快答环节，被问及近期在 AI 领域最令自己意外的事情时，Dan 风趣地调侃“看那些融资新闻每天都很惊讶”。抛开玩笑，Dan 与 Jessy 坦言代码代理（coding agents）的能力跃升确实带来了震撼，但由于此前就在该方向深耕尝试，这更像是一种预料之中的演进。在 Dan 心中，真正具有划时代意义的里程碑事件唯有两项：一是 GitHub Copilot 的诞生，二是 ChatGPT 的横空出世。而眼下尚未出现下一次同等量级的质变。 随后，主持人提出了一个开放式设想：未来每个人是否会像拥有代币钱包（token wallet）一样，随身携带着一个数字“记忆钱包（memory wallet）”穿梭于各大企业、软件和工作空间之间？ Jessy Lin 指出，目前记忆的产品形态（product form factor）远未定型。从用户体验和控制权的角度来看，过度融合不同维度的记忆往往会引发混乱和不适——例如现在的 ChatGPT 记忆功能经常出现莫名其妙的上下文污染，如果模型因为你上周在 GPU 上跑了模型训练，就在为你推荐个人床单时给出荒谬的关联，这种体验是非常失败的。工作与个人生活之间必须存在清晰的边界隔离。 Dan 则提出了另一重视角：劳动力技能的“数字便携性”。在传统职场中，员工为公司创造价值，所有业务 IP 归雇主所有，但员工带走的是自己积累的技术经验、处事方法和认知模型；而在数字世界中，若能通过技术手段打造一种“脱敏后的技能包”，让员工既不侵犯原雇主知识产权，又能将自己训练、沉淀出的 Agent 能力带入下一段职业生涯，将构成一种极其强大的正向激励机制，驱动每个人在各自的工作和生活中更深入地拥抱与塑造专属 AI。 --- ### `36:19` 视觉与语言的世纪争霸：生物学视角下的“高带宽”异说 主持人 Shaun Maguire 回忆起自己 2007 年在斯坦福大学开启博士生涯时的往事。当时的 AI 远不如今天这般引人注目，本质上就是统计学习（statistical learning），基本划分为计算机视觉（CV）与自然语言处理（NLP）两大阵营。2012 年 AlexNet 横空出世后，计算机视觉整整压制统治了行业六七年。然而近年来，以 Transformer 和大语言模型为代表的“语言路径”却展现出对视觉路径的全面反超与压制。 Jessy 对此感到颇为意外。她一直将语言视为极具复杂抽象能力的沟通媒介，并坚信从长远来看，多模态必然会融合进一个统一的大脑系统，将各种视听感知输入升华为高阶抽象认知。 Dan 则坦诚地回顾了自己的认知逆转： > “在很长一段时间里，我对纯语言其实完全不感兴趣。在整个动物界中，语言交流方式与人类截然不同，我过去始终和很多 AI 学者一样，坚信智能的关键在于‘具身地在物理世界中体验与行动’——视觉和行动才是核心钥匙。” Dan 在见证 ChatGPT 的爆发后，曾前往 MosaicML 等团队探寻底层究竟是如何“制作香肠（how the sausage is made）”的。他发现语言的学习在数学本质上其实极其反直觉且艰涩：每个词最初都只是高维空间里彼此孤立、毫无语义连续性的 One-hot 向量。然而，业界通过比视觉模型大上数个数量级的巨型参数网络，硬生生把语言抽象出了极其强大的逻辑与推理深度。 最后，Shaun 分享了一个他长期思考的“奇思妙想（crackpot theory）”来解释为什么语言在机器领域跑赢了视觉：在人类生物进化中，视觉相较于语言具有不可比拟的物理与生物带宽优势。人类大脑通过视网膜处理光学数据的比特率（bit rate），在电脉冲信号抵达高级大脑皮层之前，就已经完成了难以想象的多层预处理与信息吞吐。生物系统在视觉处理上的算力与数据流天然极其庞大，而人类在利用机器去复刻这套复杂生物带宽时，却面临着远超纯文本符号流的数据密度瓶颈。 ### `40:00` 脑科学、信息论与多模态：为什么语言统治了 LLM Shaun Maguire 在讨论接近尾声时抛出了一个他自嘲为“疯狂且外行的民间假说”（"crazy ass dumb non-technical crackpot theory"）：在现实物理世界中，声波在信息论意义上的比特率远低于光线，人类大脑必须将低频的声学信号“升频”为电信号送入神经中枢，而对视觉信号的处理则像是把密集的外部光子“降频”为神经电信号；然而在现代计算机体系中，一切信息都被均等化为纯粹的电子流动。Shaun 认为，这种计算介质的转变实际上“削弱了视觉、提拔了语言”（"nerfed vision and promoted language"），将两者拉到了完全同质的起跑线上。这也从信息论的角度解释了为何 Transformer 架构在语言建模上表现得如此得心应手，而在视觉领域却面临截然不同的挑战。他随即向具有神经科学背景的 Dan Biderman 和计算认知背景的 Jessy Lin 求证这一想法是否靠谱。 Dan Biderman 从生物演化与脑科学的视角做出了深刻回应。Dan 指出，人类今天坐在电脑屏幕前阅读文档、撰写备忘录等知识工作，在生物进化史上根本不存在进化适应——“我们并没有为了干这些事情而演化，我们的大脑天生不是用来干这个的”。人类大脑在结构上存在极其强烈的视觉偏好，枕叶（Occipital Lobe）等视觉皮层占据了庞大的脑区空间，而语言区和听觉皮层（颞叶）所占的生物物理“地皮”（real estate）相对小得多，这与高度偏向嗅觉的啮齿类动物形成鲜明对比。谈及脑区比例时，Dan 展现了典型的程序员与科学家式幽默，当 Shaun 打趣问他怎么不直接凭记忆报出具体脑区比例时，Dan 笑着坦白自己得“查查 ChatGPT”，并调侃道：“我个人生活里可是个彻底的 RAG 信徒，全靠外部检索记忆……我唯一内化的只有对你们的深厚感情。” 谈及多模态工程实践，Dan 指出，哪怕在训练当下的视觉-语言大模型（VLM）时，文本语义在底层表征上也往往压倒性地统治着视觉内容。某种生物的大脑结构偏向视觉模态，并不意味着硅基计算系统用相同的方式处理信息就会更加高效。他强调，对于脑机接口（BCI）研发而言，充分理解大脑各个感官通道的输入带宽与皮层面积分配至关重要；但就人类的日常知识工作来说，让大模型以文本作为核心媒介进行交互与内化，已经是高度合理且极具能效的形态。 - **感官比特率与计算平准化**：现实中光波与声波存在巨大物理带宽差异，但计算机内部纯电子化的环境消除了这种物理介质的差距，使语言与视觉站在了同等计算基准上。 - **生物演化与现代知识工作的不匹配**：人类大脑进化出了巨大的视觉皮层（枕叶），却没有针对读写文档演化出专用硬件；但在认知计算中，文本反而成了信息密度和推理逻辑的最佳载体。 - **多模态模型中的语言偏向**：在当代 VLM 训练过程中，语言表征依然占据着支配性的核心地位。 --- ### `42:40` 终局愿景：每个人专属的神经网络记忆与数据平面的神经接口 当 Sonya Huang 问及如果 Engram 的一切探索都被验证是正确的，5 到 10 年后的世界会变成什么样、以及 Engram 在其中扮演什么角色时，两位创始人给出了极富远见的描摹。Jessy Lin 勾勒了一个彻底告别“单体全能前沿模型独裁”的世界。在她设想的未来中，每个个体和每个团队都将拥有专属于自己的模型。这些模型不仅与别人手中的模型千差万别，也不同于中心化的前沿通用模型；它不再只是机械地“了解你”，而是能以一种与你的个人思维习惯高度契合的方式为你提供协作支撑。世界上将不再只有一个单一维度的通用智能，而是遍布着形形色色、专精各异的分布式智能体。 Dan Biderman 则将这个愿景与神经科学中的精妙机制关联起来。他解释道，在大脑深处，空间导航系统与情境记忆（Episodic Memory）共用着极其相似的神经回路——负责在物理空间中标记地标的细胞网络，同样承担着编码生活片段与知识记忆的功能。基于这一原理，Dan 将 Engram 的未来角色定义为“面向全人类数据平面的神经大模型接口”（"the actual LLM interface to the data plane for everyone"）。 Dan 类比了企业软件史上的巨头 Oracle 和 Databricks：如果说它们建立了管理结构化与半结构化数据的基石，那么 Engram 的使命就是通过数以亿计的个性化模型，构建出高度联想化、自适应的“神经记忆系统”。这一接口不再将文件系统当作一堆冷冰冰的扁平文件和路径，而是将整个数据平面动态映射并升华成一种神经网络层面的“大脑状态”（brain state）。这不仅极大降低了上下文推理的能耗，更为未来知识工作创造了前所未有的智能效率。两位创始人在一片对未来充满信心的共鸣声中，结束了这场横跨神经科学与前沿 AI 记忆架构的对谈。 ```mermaid flowchart LR accTitle: Engram 数据平面的神经接口架构愿景 accDescr: 展示从底层原始数据平面到 Engram 神经接口，最终支撑起数亿个性化专属模型的过程 subgraph DataPlane["底层数据平面 (Data Plane)"] direction TB D1["企业内部知识库"] D2["个人交互与情境数据"] D3["动态文件系统"] end subgraph EngramCore["Engram 神经接口"] direction TB E1["类空间导航与联想编码"] E2["权重参数化神经记忆"] E3["动态大脑状态映射"] end subgraph DistributedAI["分布式个性化智能"] direction TB M1["个人专属模型 (知晓个人思考偏好)"] M2["团队协作模型 (共享深层组织常识)"] end DataPlane --> EngramCore EngramCore --> DistributedAI ``` - **记忆与导航的同源神经回路**：在认知神经科学中，海马体等结构中表征空间物理地标的神经元回路，在机理上直接参与了情境记忆的存储与检索，Engram 正是借鉴了这种联想表征哲学。 - **重塑数据平面交互**：对标 Oracle 和 Databricks 在传统数据基础设施上的地位，Engram 旨在成为新一代由数亿个个性化模型构成的神经记忆底座。 - **从机械检索到“大脑状态”**：未来的专属模型不再逐字机械比对文件系统，而是将海量数据直接内化为模型的动态神经表征，实现超低 Token 消耗与高度情境化的认知反应。 ## Notable quotes > `00:00` 我们不从预训练或后训练的视角看待世界，我们的模型始终在训练 — Sonya Huang > `12:35` 训练的魔力 — Dan Biderman > `17:15` 在边界处学习 — Jessy Lin > `20:00` 特沃斯基常说，他感兴趣的不是人工智能（Artificial Intelligence），而是‘自然愚蠢’（Natural Stupidity）。 — Dan Biderman > `21:40` 如果真要让模型彻底理解 Shaun、理解你的全部工作和上下文，单靠每次把文件重新塞进上下文窗口读一遍是绝不可能做到的——特别是像你这么特殊的人物，我们必须为你专门训练一个 100 万亿参数的模型！ — Dan Biderman > `24:45` 我们正在向客户证明，通过相当轻量化的训练，这些知识和技能就能变成模型的‘本能’。模型根本不需要跑那些极其繁琐漫长的 REPL 循环。从某种意义上说，这就是一个‘RAG 杀手’（RAG Killer）式的方案。 — Dan Biderman > `28:30` 现实中如果有个员工被问到‘你昨天在做什么’，他绝不会当场打开搜索框搜一遍自己的工作记录，他是靠脑子直接知道的。 — Jessy Lin > `31:37` 这家伙真的变聪明了，它和昨天完全不一样！ — Jessy Lin > `36:19` 在很长一段时间里，我对纯语言其实完全不感兴趣。在整个动物界中，语言交流方式与人类截然不同，我过去始终和很多 AI 学者一样，坚信智能的关键在于‘具身地在物理世界中体验与行动’——视觉和行动才是核心钥匙。 — Dan Biderman > `40:00` 我们并没有为了干这些事情而演化，我们的大脑天生不是用来干这个的 — Dan Biderman > `40:00` 我个人生活里可是个彻底的 RAG 信徒，全靠外部检索记忆……我唯一内化的只有对你们的深厚感情。 — Dan Biderman > `42:40` the actual LLM interface to the data plane for everyone — Dan Biderman ## Mentioned Dan Biderman, Jessy Lin, Engram, Sonya Huang, Shaun Maguire, Sequoia Capital, Notion, Microsoft, Harvey, Anthropic, OpenAI, Demis Hassabis, Google DeepMind, Amos Tversky, Rich Sutton, Fireworks AI, LLaMA, Taylor Swift, GitHub Copilot, ChatGPT, AlexNet, MosaicML, Oracle, Databricks ## Video description Dan Biderman and Jessy Lin, co-founders of Engram, are building a neolab around memory and continual learning, which they call two sides of the same coin. Their contrarian premise: instead of stuffing ever-larger prompts into the context window or bolting on RAG, bake a team's knowledge directly into the model's weights, so it knows your company the way an employee of several years does.  The payoff: matching or beating frontier models while consuming up to 100x fewer tokens. Working with partners like Microsoft, Notion, and Harvey, the team draws on roots in computational neuroscience and state-space architectures to attack what they see as the real bottleneck in AI — not raw intelligence, but memory and continual learning. In contrast to the frontier labs' race toward one ever-bigger model and AGI, Dan and Jessy imagine a world where everyone has their own model — privately trained, always learning, and good at the things you actually care about. The real ChatGPT moment for memory, they argue, is the day your model feels like an intern that genuinely got smarter overnight. Hosted by Sonya Huang and Shaun Maguire, Sequoia Capital [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI1349350448.mp3)

[0:00] Sonya Huang: 我们不从预训练或后训练的视角看待世界，我们的模型始终在训练 [12:35] Dan Biderman: 训练的魔力 [17:15] Jessy Lin: 在边界处学习 [20:00] Dan Biderman: 特沃斯基常说，他感兴趣的不是人工智能（Artificial Intelligence），而是‘自然愚蠢’（Natural Stupidity）。 [21:40] Dan Biderman: 如果真要让模型彻底理解 Shaun、理解你的全部工作和上下文，单靠每次把文件重新塞进上下文窗口读一遍是绝不可能做到的——特别是像你这么特殊的人物，我们必须为你专门训练一个 100 万亿参数的模型！ [24:45] Dan Biderman: 我们正在向客户证明，通过相当轻量化的训练，这些知识和技能就能变成模型的‘本能’。模型根本不需要跑那些极其繁琐漫长的 REPL 循环。从某种意义上说，这就是一个‘RAG 杀手’（RAG Killer）式的方案。 [28:30] Jessy Lin: 现实中如果有个员工被问到‘你昨天在做什么’，他绝不会当场打开搜索框搜一遍自己的工作记录，他是靠脑子直接知道的。 [31:37] Jessy Lin: 这家伙真的变聪明了，它和昨天完全不一样！ [36:19] Dan Biderman: 在很长一段时间里，我对纯语言其实完全不感兴趣。在整个动物界中，语言交流方式与人类截然不同，我过去始终和很多 AI 学者一样，坚信智能的关键在于‘具身地在物理世界中体验与行动’——视觉和行动才是核心钥匙。 [40:00] Dan Biderman: 我们并没有为了干这些事情而演化，我们的大脑天生不是用来干这个的 [40:00] Dan Biderman: 我个人生活里可是个彻底的 RAG 信徒，全靠外部检索记忆……我唯一内化的只有对你们的深厚感情。 [42:40] Dan Biderman: the actual LLM interface to the data plane for everyone

[18:30] Dan Biderman: As Amos Tversky, the Israeli psychologist, used to say... he's not interested in artificial intelligence. He's interested in natural stupidity. [18:30] Dan Biderman: If I truly want to understand Shaun... just re-reading files is not going to make it... We got to train a hundred trillion parameters. [34:55] Dan Biderman: A holy grail is like, you go to work, and you just burn through all these tokens, and you create all this value... but somehow the skills you learned, the things you invented, your ways of doing things, some of them you can take with you as well to your next job... [42:00] Dan Biderman: ...we form these memories that happen to be neural memories with models that happen to be personalized... but they're basically a neural interface to the data plane in a way that's very different from what we know. It's more efficient. It's more associative. It's not representing the file system as it is. It's representing a brain state of that file system.

[0:00] What about pre-training or even post-training makes it possible for the models to generalize in these magical emergent ways and controlling that process so that a company has a set of private data? How do we make the models learn that just as well as the models know the capital of France or how to write Python? So I think it's a really fun problem to think about. Welcome to Training Data. We are delighted to have Don Biederman and Jesse Lin, co-founders of Ngram today. Ngram is a neolab focused on memory and continual learning, and two of the hottest topics in all of AI research today. And Sean and I are delighted to dig in on those topics with you today. Awesome. Happy to be here. Great. So maybe to kick off, the Engram website says, we don't see the world through the lens of pre-training or post-training. Our models are always training. What does that mean? So I think like models today obviously you know a lot of things They incredibly smart But we kind of think the bottleneck for making these models more useful these days is not really raw intelligence but understanding new and evolving contexts. So whether it's a new task that you're doing or a particular context for a job or something like this, how do you bake that into the model weights the same way that pre-training and post-training bakes that into the model weights very deeply. And this is kind of why we think of ourselves as working on these fundamental problems of memory and continual learning, which are really two sides of the same coin. How do you make the models learn new things and bake them deeply into the weights of the model? And is your premise then that memory as a separate database or separate, you know, thing that you've shoved into the context window is not true memory and it's not true continual learning? I think all of these tools will kind of come together. So these days, like the way that people are solving these problems is with context engineering. So you take a huge prompt, maybe you keep talking to the model over many, many turns and hours and reorganize the context to better understand what you're trying to do. And we think these kinds of things, like tool use, context engineering will play a part, but I think an under-leverage tool these days is using the same kind of training pipeline or framework or kind of workflow that the frontier labs are using to make these models really good at frontier math or code but applying that to every kind of domain every kind of context that you have like let say in a company Yeah. And to me, it's like as an individual, taking notes and having sticky notes is a very valuable thing. We should never discard this. But whenever we get back to business the next day, we always have some sort of trace of memory in our brain, some new intuition about how things should be and where should we look. So these two things should come together. And current solutions are more kind of externalized memory. And this has two issues. One is that the amount of tokens we will all collectively individually generate is going to be in the tens of millions of tokens per day soon. So just keeping it and searching through it is going to be and rereading it's going to be pretty expensive, but it's going to also be pretty hard, pretty confusing for the models unless we have major, major breakthroughs. Tens of billions of tokens for Sean. That's good. Depends on the day. Could you maybe tell us a little bit about the Engram architecture, the Engram product and how it works? Yeah I mean at a high level I think what we trying to do is take any context like there all these different workspaces let say So we working with partners like Notion Microsoft and Harvey that have these places where people are doing a lot of work over a long period of time There's all this context, both in terms of documents that you've already written as a team, as well as now people are interacting with these agents more and more in these products or having conversations, giving them feedback and figuring out how to have a model that deeply understands that context. So not just reading the files at test time, but really understanding it the way that an employee that's worked at your company for years has. So you kind of understand at a high level, oh, these are the initiatives across the company. This is the way that we do things. You've studied how to run the hiring pipeline or how to do this kind of thing within the company and can operate just as well as anybody else can in the company. And so what we're doing is training per team models within these workspaces that deeply understand those contexts and can improve with time on the things that people care about. So the way that we do this at like a technical level maybe is training these into weights. So we do a lot of like adapter fine tuning. So adapters of many types, like I think people have looked into this for decades at this point, like whether it's Laura's or prefixes or, you know, [5:00] sparse architectures, I think like all of these tools are at our disposal. And then figuring out what the right data is. So how do you turn any kind of raw, like document or interaction into useful training signal for the model. So again, we have like a variety of tools now, like supervised fine tuning, you know, RL, you know, on policy distillation, like all of these things that, you know, the field has kind of developed and trying to fit these pieces together into a model that learns continuously on the things that people care about. Yeah. And it's not a bet that tools are not there. Our models always work under the assumption that some knowledge is externalized, some tools are always there. But what you need to do is you need to figure out, and that's the hard task, is what needs to be internalized and what can be externalized. And even for stuff that's externalized, many individuals and companies have their own bespoke tools and ways of doing things. Not everyone has the same Bash CLI tools that the Frontier models are training on and how to get the models to better understand your bespoke setup, I think, is its own interesting thing. And so is the premise then that my Notion agent will be a custom agent that is Laura fine or you know some way with an adapter tuned so that it constantly learning on new content that added into my notion workspace is that yeah and they working with many models and they're the early users of all the frontier models and they're probably going to keep doing that does this approach work on the frontier models or are the closed so your models we need we need white you know we need white box access to the weights right so you know So we can partner with companies that have closed source weights and do this with them. But it's easiest for us to do it with open source models. But any model that's a transformer model, we can do our thing to it. And what's the tradeoff then when people are comparing the before and after using you? Is it that they're no longer sending so much context? and so the trade-off is like you burn more compute up front to learn your company's way of doing things into the weights and then you're sending less context to the model on every inference pass is that the rough trade-off? That's one thing the fact that you don't have to research things and the fact and reread things and the fact that you don't have to write like monstrous system prompts they're already that can give you you know two orders of magnitude reduction in token inference consumption It not like you know 50 or it can be 100x fewer tokens because many things especially things that relate to people and teams and organization and priorities These are things that you can't really find in one document unless like you really have it really regimented and document everything. And these kinds of things, the model can kind of implicitly learn by training on some of the data and answer, you know, within a hundred tokens, what, what the best frontier models would consume a hundred thousand tokens doing. So these kinds of examples are interesting. And also the quality. There are tasks that are not supernatural for the current generation of the models. And we kind of think there's going to be consistently this gap of three to six months ahead where there's certain things that are bespoke that people are just exploring. The models are not fully great for them. The models will at some point be great for them. But if you can autonomously learn in a very lightweight way, it will give value in that time in terms of capabilities. Why train on the workspace level versus the individual level, for example? Either is fine for us. It's just easier to start with, you know, teams of people have, you know, are more, you know, disciplined in how they collect context and in the amount of context they have over years. And it's easy for us to start there. But every person computer and every person phone one day is a useful you know target for our technologies And in fact it will be very interesting to go there We just think the big deposits of information are now in teams of people collaborating in knowledge work Is it a feature or a bug that there is so much fact memorization basically built into the large language models? And there's a school of thoughts that the models just rote memorizing the fact that the capital of France is Paris is actually a bad thing. And what we would prefer for the models to do is abstractly learn the concepts of countries and capital cities, but not to memorize all these facts in the weights. And so I'm curious what you think about disentangling memorization versus learning, how it's done in the models today, and then how you're thinking of approaching it. Yeah, I think it's a really interesting question. Like, to some extent, you kind of need to remember stuff in order to like compose them into more complex concepts. I think the thing that's kind of missing is figuring out what's important to remember. And I think even now when you think about like learning new knowledge, if you look at a lot of these academic benchmarks, it's like, how can we learn very specific facts like, you know, the length of a bridge in this like African country. And that's not something that you really want the models to devote capacity for. And it's not something that we devote capacity to. So I think if you look at human memory, I mean, you can say a lot more about this, but like it's lossy because. [10:00] As part of the feature of intelligence is compressing what's important and separating that from what's not important. And so I think like you can't really separate fact learning from like non fact learning or skill learning, as some people would like to think. Like if you take a model and like some people have done this with models where you like strip out, you know, like all the facts and just have it like the pure core or something like this. It's very unnatural as a model. It doesn't know basic things. and you kind of need that. But I think... Why do you need that? Like, why can't you look up facts and then just have... I think if you look at, like, how the models think, if you need to recall basic facts in order to, like, take the next step in your thinking, you can't get very far. Maybe that's, like, a high-level intuition, but it's part of, like, the reason why we think training is really important. In order to, like, think more and more complex and deep thoughts about things, you kind of need to internalize something so that you can compose them into more abstract concepts. Yeah, and there have been efforts before that were hard to scale to try and disentangle the two and pre-train the models in a way that allows it to retrieve and search for things and not internalize them The recipe we know to hill climb on collectively right now is this you know fact pre step And I think the magic of or the mystery of this approach is that, you know, traditionally in CS, we would have, you know, databases as its own curriculum, and we would have algorithms and the databases is like facts about the world and capitals of whatever, store them or query them. There's also algorithms of how do you efficiently manipulate information and get some answers in a sample efficient way. And I think the magic of deep learning is that these two things are now mushed together and we need all these smart people and anthropic interpretability to try and break them apart. And I think a lot of what we're seeing now in the adoption of AI into the economy is that these things are gradually separating again, where companies have their own context and they really handle them with care and engineer them with care. And there's a generic model that's completely a stranger to these contexts. And the model is operating on them. But for us, it's clear that there needs to be a certain convergence, at least with some cadence, where the facts and the stories and the details are getting mixed into the model It has disadvantages as well because if you have to you know capitals of countries are you know they can change but it not very frequent But there many other facts that are changing all the time And just imprinting them into weights is a challenging thing to do I see. So you're saying it's a false dichotomy to try to separate algorithms from databases here. What really matters is like how to distinguish what's important to remember versus what's not important. Exactly. And it's an open question. I guess it's part of how we dream. And are you guys taking any inspiration from that in terms of ranking? Very, very loosely, I think. Just the idea that that's kind of a phase that's missing, maybe, where you take a context and you deeply internalize it. Right now, it's like everything happens at test time. You look at the context that the user gives you and you do some thinking on the fly. But again, you can't get very far or you can get so far, maybe. And you make mistakes along the way. How do you digest that back into the model so that next time you do it, you do it the right way and make even more progress? Yeah. And what are dreams? Dreams are pretty crazy things to say. We want to build an AI that's like our dreams. It sounds a little bit like a nut thing to do. There's not a lot of coherence there. But what's interesting there is like what happens in our dreams? We see things, we talk to ourselves and we experiment with the affordances of what can we do and can we do in the world and social situations And then you know any any it heavily biased towards social stuff right So for us too with things we building is you know we give the models the time to then go back retreat from the actual interaction and experiment with its affordances. What can it do in an environment? What can it, what does it know? How fast can it, you know, handle these kind of tail extreme things that same ones that we dream about at night? You guys come from academic backgrounds. What's a canonical example that motivates this problem or that's a win so far? Yeah, I have one example. Maybe Jesse can give another one. A hypothetical one, for example. Imagine one of the AI labs say OpenAI has to win some math Olympiad in a week time from now. would they construct the catalog of all the math textbooks and really have people annotate which chapters to get and which graphs to see, or will they actually collect this, synthesize some training data, launch a training job, see where it lands in five, six days, start evaluating it and stuff like that. So it's obvious for anyone who's trained models that there's a superior way to integrate across the ideas and capabilities and involves this kind of magic of training. And we are clear that this has to happen in those high [15:00] domains of math and coding and cyber and stuff. We just think much of this magic can actually end up in the hands of many more people in interesting ways. Why isn't it just the foundation model labs that own the end product here? How do you go between giants? Yeah. So I think the worldview that we have is a bit different from the Frontier Lab worldview, where it's like, we want one model that's bigger and bigger, that's more and more intelligent across a variety of domains. Instead, how we see it, like we kind of imagine this world where everybody has their own model. A lot of the things that people want to learn are either private, like things that will never see the light of day in a post-training data set, or even conflicting, like, oh, the way that I want to do the task is different from how another company or another individual wants to. And I think a lot of these things we're already seeing are hard to train into the models with the same tools that we have used for like decades in machine learning, which is like you have really clean supervision, you have like ground truth reward signals and you like create a nice environment and you like train the model to like use the tools to better accomplish this like coding task And instead a lot of the things that actually happen out in the world are very ambiguous or like it hard to say like what makes something good And so I think a lot of these things are very specific to individuals and I think very kind of misaligned or not very aligned with how the frontier labs think about the whole training pipeline and what kind of models will exist in the longer term Yeah. And to add to it, I think, you know, what is the peer zero for the frontier labs? And some of you here are pretty close with them. It's getting to AGI, getting this one generic model that's extremely capable in coding and math and then using it to automate the economy or to solve really hard, you know, long term problems in cryptography and defense or whatever. And it's pretty clear what needs to happen to push this, you know, more pre-training, bigger models, more data, more RL, more inference time compute, that kind of stuff. That's P0. That's where the majority of expenditure and talent goes. And definitely all of them are thinking about memory and all of them are thinking about continual learning. It's just more of a product kind of effort right now. We think it deserves its own attention. We think breakthroughs need to happen there. And Demis and the Sequoia event about a month ago said pretty clearly that we need new breakthroughs around these topics And obviously they thinking about them We just focusing exclusively on this And we think certain things are on incentives of where the data is and who owns the model are pretty interesting. So if you could learn from many humans or organizations at scale without necessarily sending someone to work with them shoulder to shoulder, that would be a pretty big unlock. And maybe another point on that is like, I think a lot of things need to look different in the world. So one is there needs to be new research breakthroughs. Two is new infrastructure for training, like, you know, small models for everybody rather than like one big model, one big run. And then the third, I think, is a different way of kind of combining research and product. So right now, I think like there's like researchers in these frontier labs. They kind of train the model. They throw it over the fence to the product team who then like prompts or contacts engineers like new products, surfaces on top of the core models. But in this world where the models are always training, I think the inputs that users provide are very intricately tied to what the models learn from, like what the training signal is. And so there needs to be a lot more of a kind of integrated loop between research and product. And so while we focused on tackling a lot of the core research challenges and that our background I think we also very focused on how to deploy this as quickly as possible to learn from actual feedback in the real world What motivated you to work on this problem I think it's obviously one of the grand challenges in AI. I think everybody's talking about it these days because the models are so smart. What else is left? I think learning at the edges, learning the remainders of what makes these models useful. It's not just about raw intelligence anymore. It's about like learning new things. And I think it also feels very fundamental because it kind of goes back to really understanding what makes the model so good. So right now the models kind of incidentally know a lot of things from pre-training and we don't really understand why. It's like the internet was just, you know, this gift granted to us where there's like a diverse set of data that contains like all of these different examples of coding and like writing and all these other things, and it just happened that way. And now to figure out how to crack this problem of continual learning, it's about figuring out what about pre-training or even post-training makes it possible for the models to generalize in these magical emergent ways and controlling that process so that a company has a set of private data. How do we make the models learn that just as well as the models know [20:00] like the capital of France or, you know, like how to write Python. So I think it's a really fun problem to think about. And Dan, you came from the neuroscience world, is that right? Yes, yes. So I was initially interested in questions around, you know, consciousness and the human condition and things like that. Are the models conscious? I don't have any advanced thoughts on this more than you would read. I don't think so, but it's important that smart people are thinking about it. I would say like I was interested in how humans think, how humans perceive. And as Amos Tversky, the Israeli psychologist, used to say, like he's not interested in artificial intelligence. He's interested in natural stupidity. So I would say like I started kind of similarly trying to see how people and animals experience the world. Gradually, you know, my inclinations took me to the stats and AI domains. And there I figured that so many of the same problems of memory and continual learning are really, really urgent. and the kind of solutions we have in the current systems are pretty far from what we have in biology. And I'm not one of these people who would say that the machine should be like, you know, like the animal or the human brain. I don't think so. There's many things computers can do better than us. But human memory has these like very different things in it It you know if you want to store a whole code base or you can use a computer you don even need AI on the computer to store everything losslessly and just get it But the human brain evolved to work in these constraints of, you know, information capacity and to have these fuzzy representations that can then, you know, be abstracted and form connections and form the next day. Current systems don't really have that beyond the generic pre-training step. And I was really interested in, you know, what are ways to build that in? What are ways to learn from that? This is more of a philosophical question. You know, you mentioned in the brain, there's a bunch of different real estate, different co-processing units, whatever, modern computer architecture, there's CPUs, GPUs, you know, memory, there's different co-processors. With the bitter lesson, do you think that what's happening is that like LLMs are, converge to say like one coprocessor that's just totally dominant. It's like everything, all compute is gonna happen in the GPU equivalent of like a language model. Or do you think that these models are kind of building a bunch of coprocessors like you know emergently inside the model like like you know and take with memory like do you think that the models themselves will just build you know whatever part of the brain equivalent would be that's good at memory or or do you think there needs to be like another standalone architecture of it yeah like is memory an emergent property exactly versus like And almost everything. Is everything that we need in intelligence will just be emergent with better training data and more scaled compute? Yeah, I would say just on a more superficial perspective on the current deployment of AI, it's way more than just GPUs. And we're seeing all these sandboxes exploding and models operating on other computers trying things. I'm more mean on the model architecture level rather than on the... So other experiments, there have been many previous experiments on different architectures that we contributed to, like the state space family and others, to try and handle very, very long contacts more efficiently. The thing with all these methods, it ends up being a trade-off, usually a trade-off between memory and accuracy. And memory not in the behavioral cognitive sense, memory in the computer sense, right? Instead of having the memory footprint of the transformer attention which is quadratic in the sequence length these models have Some are claiming they have sub Yeah some are claiming and some do have it right And some of the best Chinese models have layers that are you know inspired by those state space architectures and are, you know, not quadratic in cost. Thing is, is that in our hands, we find that, you know, you always compromise accuracy for this memory. There's no free lunch. And what we're saying is like, look, if you're really bitterless and pilled, what do you want to do is you want to think how can i burn more compute and how can i burn it on you know new context that i have not seen before so we're as bitter less and pilled as anyone else and we are not betting that the overall direction of agi is is gonna you know end anywhere soon we just think there's more compute to scale and if i truly want to understand sean and sean's work and sean's context just like rereading files is not going to make it especially for a special person like you We got to train a hundred trillion parameters. Cosign. What are you finding that people care most about their models learning? Like, is it memorizing facts about the organization? Is it remembering like, ah, no, we do CI this way? Is it like, what are people actually hoping to, and then maybe this feeds into how you [25:00] do the ranking of memory slots and all that. Yeah, well, I think if you look at what people are spending their time in the app layer doing these days, it's a lot of just trying to make the model work well for your use case. Like, oh, I want the model to like, you know, let's say like design my website with my brand style. Like that's like a, you know, very common example these days. But there's many kinds of different tasks that people do with agents, like learning how to run of workflow or, you know, kind of your particular way of like writing, let's say. So there's many, many kinds of things. And honestly, like I think when we think about these methods, kind of going back to this distinction between like facts and skills, there really is none. I think the methods are kind of agnostic to that. Yeah, to me, it's like the natural thing. Almost all the app layers are basically, you know, a frontier model wrapped in a loop with search tools and stuff. And what they're all interested in doing with us is finding ways to kind of interface with their data in a way that's, you know, faster, more efficient, and also is more contextual. So almost all of them it like we want to have our you know our firm knowledge you know be encoded in something that more efficient that I don have to research We want to have the model know in a targeted way who the person I should triage a thing to And we're just showing them that with pretty lightweight training, these things can be instinctual to the models. They don't have to have these very involved long REPL loops to solve them. So it's in a sense, it's like, you know, it's a rag killer kind of thing. Again, we can always do RAG and we can always retrieve, but that's the thing that people are interested in, interfacing with very large data planes and automating very repetitive things this way. Yeah. And I want to double click on this RAG killer thing, and I'm sorry to beat a dead horse. I just don't fully grok it yet. Is the premise that there's some tradeoff between doing RAG versus updating your model weights? Is it an idea that you should be doing both? Like what types of things should be done in the weights versus what types of things should be externalized to RAG? I think it's an unsolved problem. I don't think anyone has answered to it. We're all working on it. It's also the fundamental question of biological memory, what should be internalized versus what not. I do think that things that are like, do you need to internalize the room number in a hotel that you were in a year ago Probably no not in your neural tissue Probably that good to write down but do you need to internalize the room number in a hotel that you were in like a year ago Probably no not in your neural tissue Probably that good to write down But do you need to internalize maybe the password to your home right now Probably it useful for the next few years to have that imprinted somewhere. So yeah, how does this translate into like knowledge work and products? This is still something we figure out. And we try to take the approach that we try to use as few heuristics as possible. It's easy to run filters on the data and say like, I'm going to keep this, discard that, train on this, train on that. But as humans, you know, we watch TikTok and we, you know, get exposed to a lot of garbage and still the brain is able to learn and not completely go off the rails. And we think models should be the same as well. Yeah. Maybe concretely in the short term, I think a lot of what people are worried about these days is the huge inference costs of running these agents like for days on end. High inference costs a good thing. I mean, consuming tokens for what? Sonya works with fireworks. I really love Zyatriska. I love Infrance. We love Infrance too. Yeah, so I think it's like, in the short term, I think that's the immediate pain point. Like, why are you reading the same files over and over again, you know, even in the same query but like definitely you know across people in the same company they running the same queries on the same documents over and over again And that should be something the model just knows like in the same way you ask an employee they don type into the search box like what was I working on yesterday They just know. But doesn't caching kind of solve that? I think to some extent, yeah. But I think going back to this question of what should be internalized versus what's something you retrieve at test time, I think, again, a lot of it is about building on your knowledge. So if you are always doing RAG, you can't make associations like, oh, you know, I see somebody, you know, on the team is doing this kind of research. And I kind of like recall at an abstract level, oh, there's this like related thing that you might want to know about. You didn't even ask about it. Right. But I think like these kinds of associations can only happen in weights because they're not really about, you know, you asked me to search for this. I'm going to search for this. And also, I think the main limitation with retrieval systems in general and in AI specifically is like, the problem is not so much what to store and where to put it. The problem is like how to address it, like how to query the thing. Do you know what to look for even? And this involves some sort of intuition that sometimes the models don't have, interestingly enough, they don't know where to look. And especially if you're limited to the current way of doing things, which is keyword search, it is just easier to scale in RL and logistically. [30:00] involved in terms of like infra for embeddings and stuff. So yeah, knowing what to search is something that's intuitive and can happen in the way. And also about caching and inference, like much of this company started with us taking a deep dive into like KV caches and caching. And this is a fascinating thing, right? KV cache is a monstrosity of the current way of doing things that, you know, think about it, a KV cache for a single like Wikipedia article for some, you know, Taylor Swift or something like this, it will be like 80 gigabytes of HBM memory on the GPU and an entire Lama. It's for, say, a 70B Lama model. And the entire weights of the model would be about 100 gigabytes. And, you know, with some distortion, they remember the entire internet. And how come this thing is so, one thing is so bit efficient. And we have this proof of existence that gradient descent can pack a lot of information in very few numbers. Whereas this KV cache thing, you take a few tens of kilobytes of article and it becomes those 80 gigabytes of brain state. So sure you can cache this, you can load this, you have issues with disk to HBM stuff People are working on it it pretty interesting But what if we can take those 80 gigabytes spend some compute offline maybe also in fireworks, but then compress it and make it really, really small so that the thing we load in cache is like a thousand X smaller. That would have tremendous implications for how we load things, how fast we can do things and what the fidelity of the representation is. Super interesting. Yeah. What are some of the things that could happen in the next year or two that would be like the chat GPT moment of memory? Or do you think that that's not how things will play out? It's a good question. I don't know. I think like the first proof of concept of the thing that people keep talking about with continual learning, which is you have an intern that you can teach things over time and it actually gets better. I think everybody's waiting to see that, you know, and no matter how sophisticated the context engineering approaches are these days, they're not getting there. So I think you need all of these tools at your disposal to make that happen. But I think it will be something like that, where it's like, the model's actually getting smarter. Like, whoa, it's different from yesterday. Yeah. And it's important to say that the ChatGPT model was not anticipated. We just read about all the different product directions that certain people had before ChatGPT was um i feel like to me the example is like look if you you know resign from your job today and your sole mission was to make a model that's better for you and you would use open and topic and all these frontier models and you just 24 7 engineer the context right skills your way to move the needle is very limited as an individual you'll just be better off waiting for the next version of the model and and you'll take it from there. And we would like to see a future where actually the more time you spend on the thing actually translates to the quality of performance, at least in the things and domains you care about. And this is pretty hard to achieve. And the only reason we think it could be achieved is if you start scaling, compute, and training on these data without destroying them all, importantly, which is pretty hard. Just a couple, like, this is just for fun, like rapid fire questions going off just memory. When's the last time you reached surprise about something in AI in any area? When reading about fundraising. A lot of surprises every day. I would say all of us felt, you know, a little bit of a change around the capabilities of the coding agents. That's true. But we been you know dabbling with these things and trying to make them work in more effortful ways before So it didn come as a complete surprise But yeah I think to me the main events were GitHub Copilot That for me was just the main event and chat GPT. And then seeing the agentic stuff, we all anticipated, I think, and different people had different expectations on how far it can go and how long the horizon it can go. But I feel, yeah, we're yet to see something fundamentally different. and people are working on completely new ways of doing things now. But yeah, to me, its model is actually changing in a way that's not harmful and learning new things on the fly that are personally and economically viable. That's interesting. Right now, there's this idea of we're each going to have a token wallet that we're going to bring around to companies or to different apps, different workspaces. Do you think that we're going to end up with a memory bank, like a memory wallet that we're gonna move around to across the digital world as we go? I think it's an interesting question. I don't know if we've fully figured out what the right kind of like product form factor is in a sense. In a way, even with like ChatGPT memory, let's say, I kind of don't want it to remember across my like personal and work context. Oh yeah. Like it's like, [35:00] Oh, you know, you might like these sheets because you trained a model on a GPU last week. It's like that's totally irrelevant. And to some extent, it's like because the memory is flawed. But also, I think you do want memory in your, I guess, tools and the products that you use to be separated to have control over that. So I personally think like there needs to be some separation there. But I guess to be determined what that might look like. Yeah. And like, I think a holy grail is like, you go to work, and you just burn through all these tokens, and you create all this value. And somehow, you know, all the IP and stuff stays with the company. But somehow the skills you learned, the things you invented your ways of doing things, some of them you can take with you as well to your next job in a way that's, you know, sanitized and not, you know, harmful to any other company's IP. So I do think like carrying a set of skills will be interesting. We do it in our biology right now and we just, you know, sign NDAs and have like ethical rules around it. But I think doing it in a digital world would be pretty interesting and pretty rewarding because it will force each of us to push the frontier and implement AI more deeply in our companies and our individual life and then be rewarded for it. I started a PhD in the SaaS firm in 2007 at Stanford and AI was boring as hell at the time It was all statistical learning And there basically two areas like computer vision and NLP So like vision and language were kind of the two areas. And I think that's still true. In 2012, Alex, that happened, like vision was dominating for six years or whatever. Are you guys surprised that language seems to be like the language approach seems to be like dominating over vision in progress. Question two, do you think vision has any chance of coming back? How do you think about this? Yeah, I think it is pretty surprising to me. I mean, some people maybe saw it coming, but I think I've always kind of been interested in language as like, I don't know, I guess like a medium for communication and like so many kind of complex abstract things can be done in language. I do think like, you know, I imagine like in the longer term, language and vision will kind of like combine in this more like unified system where, you know, we kind of like take in inputs from all of these different modalities and like understand them in this abstract way. But yeah. Yeah To me like I never been interested in language It seemed to me such a such an advanced capability that you you know is very the entire animal kingdom has very different forms of speech and language than what we, you know, and how we communicate with ourselves in writing. And I was always, as many other leaders in AI had this thought that, you know, the natural thing is you have to experience the world, act in it, in vision, in action, that will be the key. But then I've, you know like anyone else we've seen the the chat gpt moment and went to to do some work at mosaic and stuff like that to learn how the sausage is made on the on the nlp side and the thing that's striking is that like the the language should be pretty hard like each word has this uh one hot embedding vector that's as dissimilar to any other word uh than it is you know uh to you know it's it's a completely high dimensional space and it's really artificial in a sense. And we, we learn it with models that are order of magnitude bigger than the best vision models. And still, you know, things work pretty well. I do think there's a lot of juice to be squeezed in, in an image and video. And I think you guys doing good, good investments in this space, but it's, I think the two would keep being interesting in different ways I mean now I tell you my that was my lead up Now I gonna tell you the crackpot theory Um I like and this this podcast is not for me to pontificate It for you guys, but this is something I've been thinking a lot about and I just, you're the right people to share this with. I, I was pretty shocked that language kind of surpassed vision and I underestimated what was happening with LLMs in like 2018, 2019, 2020, because I just had this bias towards vision. And when I look back on it now, like I think what's basically happening is that in biology, like vision has a massive fundamental advantage over language in biology. And maybe I'm wrong, But basically, like the bit rate that your brain can process optical data through the eye is in this my I'm not a biologist. This is just kind of my dumb assessment. It seems many ors of magnitude greater. And there's a lot of like optical processing that happens like even before you reach, you know, like electrons. And so it's just like the total bit rate that is of training data that's kind of being processed and then making it to your brain seems many awesome [40:00] than the audio data where, you know, it's sound waves where sound waves are fundamentally like much slower bit rate than light. And then there's almost like an upscaling from the acoustics to electronics, which make it into your brain, whereas like there's like a downscaling from photons to electrons with vision. Whereas in computers today, everything is electronic. So it's kind of like you nerfed vision and you like promoted language where the it's like all processing is on the same playing field it's all electronic and i just i think this might this is like my crazy ass dumb non-technical crackpot theory but i think this might be part of why just like from an information theory perspective that like maybe language and vision are on a similar playing field by the time you get to like llms and then llms are we're just a really really smart architecture that better suited for language than for vision um how dumb does this sound especially to you don the neuroscientist just he also has some background in cognitive computational science right So I would say my point here is like look much of what we doing in knowledge work we haven evolved to do right We're sitting on these computers, reading these things, writing these memos, whatever. We are not evolved to do this. It's new to us. Our brains are not wired for this. Still, nevertheless, it's useful to have LLMs to do this for us. And, you know, as humans, we're heavily vision biased. you know other rodents are more olfactory uh biased and i've worked on these things myself before so what's the uh real estate in the brain that's allocated to vision and you know occipital lobes versus like language areas temporal lobe probably more vision i'll have to check with chat gpt but i think that's the situation you don't know from memory no man i'm externalizing i'm a big rag believer in my personal lifestyle but um i think in the limit we're all it's all I internalize just important things like my emotions to you. No, just kidding, sorry. Anyways, yeah, and vision is dominating. When people are training vision language models, they end up the language ends up dominating the vision content there But yeah it hard to say that because a certain brain is more you know biased towards a certain modality doesn mean necessarily that we going to more efficiently do it I do think that efforts on like brain computer interfaces should take this into account. How do you then relay it back to the brain? That's where I think it's really important to think like what real estate do we have there right now? But for knowledge work, it's equally fine if it's text, I think. Last question. If everything was right, what does the world look like in five, ten years? and then what is engrams we're all in it i think i'm imagining like a world where everyone has their own model um that is really different from the other person's model and from the frontier model and all of these kind of serve different purposes and to have a model that really you know i think people often talk about like knowing knowing you um but also like um kind of like helping you in the ways that make sense to you um personally um whether it's like an individual or a team i think there's an element of having different kinds of intelligence everywhere. Yeah. And to me, actually, it's a variant of the story where in neuroscience, we know that memory and navigation are pretty closely related, same circuits in the brain that represent landmarks in space are in charge of some elements of episodic memory and things like this And for me I think the company can be you know the actual LLM interface to the data plane for everyone. So sharing some similarities to great companies like, you know, Databricks and Oracle, where, you know, we form these memories that happen to be neural memories with models that happen to be personalized and happens to be there's hundreds of millions of them, but they're basically a neural interface to the data plane in a way that's very different from what we know. And it's more efficient. It's more associative. It's not representing the file system as it is. It's representing a brain state of that file system. So that's, for me, a vision. Beautiful vision to end on. Thank you guys so much for coming by the share with your building. Awesome. Love it. Thank you guys. Thank you.