我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
再谈DeepSeek-Harness的开源的策略讨论
视频
音频
原始脚本
技术深度思考,解析 Deepseek Harness 开源策略的必然性、取舍与行业局限。 一、时代必然性,大模型已进入。 模型加脚手架双轮驱动时代。 当前大模型的能力上限早已不再单纯由模型权重与训练数据参数量决定。 前端调度范式、工具调用逻辑、上下文裁剪策略、任务拆解流程等脚手架能力,直接决定同一模型的最终表现。 在目前的技术阶段可以明确,相同大模型挂在不同 Agent 的脚手架,最终任务成功率、 Token 利用率、推理稳定性、复杂问题解决能力可以出现数倍差距。 以 Claude Code 为标杆,可以清晰看到顶尖模型的能力释放依靠的是模型与专属调度框架的双向深度绑定联合迭代。 固定的工具范式、统一的交互习惯、稳定的上下文结构,能够产生高度可归因、干净、可用于持续精调的训练数据,形成完美自洽的技术迭代闭环。 基于这一行业趋势,Deepseek 推出自研 Harness 存在极强的战略必然性。 如果长期依赖派、 Hermes 第三方 Cloud Code 套壳等外部框架调用模型,会产生致命问题。 模型需要被动适配五花八门的调度逻辑、工具习惯、任务流程。 不同社区、不同框架、不同使用者的调用范式高度碎片化,导致训练数据混杂,失败案例无法归因,模型迭代方向发散混乱,没有专属官方脚手架。 就没有统一的进化基线。 因此,Deepseek 自研 Harness 建立自身的 Agent 运行范式,是长期发展的必经之路。 二。 架构本质先进的实验型原框架,而非定型生产级框架。 Deepseek Harness 基于 Cordis 架构构建,核心优势集中在事件驱动解耦。 组件可插拔、完整生命周期管理、可逆副作用体系。 这套架构最大的价值不是成熟落地能力,而是极强的实验能力与组合能力。 行业当前的 coding agent 赛道并没有绝对最优的标准范式。 Agent 的工具组合、检索策略、调用顺序、任务拆解逻辑报错回滚机制,不存在唯一标准答案。 这类工程体系带有极强的调制属性,类似配方调试,基础原理固定。 但最优搭配、最优流程、最优组合,需要海量试错、排列组合、反复迭代,才能慢慢收敛。 正因行业无标准答案、无统一最优解。 Deepseek 现阶段没有推出固化封闭、写死的成品框架,而是选择放出一套高度自由、支持无限重组、支持插件自由增减。 支持运行时改造流程的原框架。 三、开源策略的真实底层逻辑。 官方试错为主,社区辅助发散。 现阶段提前开源,Harness 并非完全成熟产品的对外交付。 而是一套公开化的大规模实验方案。 其真实策略可以概括为两点。 一、内部快速迭代。 官方团队以 Corda 原框架为实验底座。 暴力试错,快速排列组合,验证各类 Agent 工作流,主动筛选最优范式、最优工具链、最优调度逻辑。 二。 借助社区扩大试错边界,通过开源释放自由度,让全球开发者提供多样化插件、差异化流程、特殊场景组合,补足官方内部场景覆盖不足的问题。 这是一种典型的双管齐下加速探索的工程策略。 在时间紧迫、赛道竞争激烈、自身尚未定型的背景下,这种做法可以最快速度完成范式探索。 抢占行业话语权,构建开发者生态,制造行业热度。 四、核心问题与本质矛盾,可以借社区发散,不能靠社区收敛。 虽然现阶段开源探索具备合理性,但必须认清一个核心边界。 开源社区天然擅长发散、创新、试错、多元化探索。 但是完全不具备自发收敛、标准化、工程定型的能力。 社区可以贡献插件、贡献新思路、贡献特殊场景方案,但无法替官方完成。 统一交互范式,固化核心工作流,模型与框架双向适配,标准化数据闭环构建,工业级最优解筛选与定型。 这也是当前这套开源策略最大的隐患。 高度自由的原框架开放后,社区会衍生出成千上万种差异化的 Harness 运行形态,导致全网调用范式再次碎片化。 如果长期放任这种发散状态,无法沉淀统一基线,最终会反向伤害模型迭代的数据纯净度与迭代效率。 因此可以得出关键判断。 试错可以交给社区,收敛必须官方自己完成。 社区只能作为探索的辅助增量,绝对无法替代官方的核心研发与标准化工作。 如果官方自身投入不足、迭代迟缓、过度依赖社区自发优化,这套看似先进的架构最终只会停留在无数零散实验案例。 无法沉淀真正可用的行业标准。 五、最终整体判断一,Deepseek 自研 Harness 是绝对正确的战略选择。 补齐模型前端调度闭环。 统一交互范式,构建可归因的训练数据链路,是未来大模型能力突破的核心前提。 二、现阶段开源属于提前布局,实验优先。 热度优先的阶段性选择。 在无最优标准答案、赛道竞争激烈、需要快速占位的背景下,通过高自由度原框架实现内外双试错。 具备明确的工程合理性与商业合理性。 三、当前架构定位是实验平台,而非最终成品。 Cordis 架构的高灵活、高可组合特性适合探索迭代,但不适合长期无限制开放用于生产。 过度自由会破坏范式收敛,制造数据混乱。 四、行业最终赢家一定是自己能发散、自己能收敛的团队。 社区负责探索可能性,官方负责筛选唯一性。 发散可以开源,收敛必须集权。 结语,Deepseek harness 当前的开源行为。 并非草率跟风,也并非完全激进造势,而是在行业未定、节奏紧迫、范式未知的阶段,做出的一次权衡取舍式的阶段性最优解。 未来该项目的真正成败,不取决于社区造出多少花样插件,而取决于官方能否在海量发散探索中快速筛选固化。 收敛出唯一稳定、统一、可闭环迭代的官方标准范式。 可以借天下之力试错,不可靠天下生态定型。 探索可以开放,标准必须自研。
修正脚本
技术深度思考,解析 Deepseek Harness 开源策略的必然性、取舍与行业局限。 一、时代必然性:大模型已进入模型加脚手架双轮驱动时代。 当前大模型的能力上限早已不再单纯由模型权重与训练数据、参数量决定。 前端调度范式、工具调用逻辑、上下文裁剪策略、任务拆解流程等脚手架能力,直接决定同一模型的最终表现。 在目前的技术阶段可以明确,相同大模型挂在不同 Agent 的脚手架,最终任务成功率、 Token 利用率、推理稳定性、复杂问题解决能力可以出现数倍差距。 以 Claude Code 为标杆,可以清晰看到顶尖模型的能力释放依靠的是模型与专属调度框架的双向深度绑定联合迭代。 固定的工具范式、统一的交互习惯、稳定的上下文结构,能够产生高度可归因、干净、可用于持续精调的训练数据,形成完美自洽的技术迭代闭环。 基于这一行业趋势,Deepseek 推出自研 Harness 存在极强的战略必然性。 如果长期依赖OpenAI、 Hermes 第三方 Cloud Code 套壳等外部框架调用模型,会产生致命问题。 模型需要被动适配五花八门的调度逻辑、工具习惯、任务流程。 不同社区、不同框架、不同使用者的调用范式高度碎片化,导致训练数据混杂,失败案例无法归因,模型迭代方向发散混乱,没有专属官方脚手架。 就没有统一的进化基线。 因此,Deepseek 自研 Harness 建立自身的 Agent 运行范式,是长期发展的必经之路。 二、架构本质是先进的实验型元框架,而非定型生产级框架。 Deepseek Harness 基于 Cordis 架构构建,核心优势集中在事件驱动解耦。 组件可插拔、完整生命周期管理、可逆副作用体系。 这套架构最大的价值不是成熟落地能力,而是极强的实验能力与组合能力。 行业当前的 coding agent 赛道并没有绝对最优的标准范式。 Agent 的工具组合、检索策略、调用顺序、任务拆解逻辑、报错回滚机制,不存在唯一标准答案。 这类工程体系带有极强的调制属性,类似配方调试,基础原理固定。 但最优搭配、最优流程、最优组合,需要海量试错、排列组合、反复迭代,才能慢慢收敛。 正因行业无标准答案、无统一最优解。 Deepseek 现阶段没有推出固化封闭、写死的成品框架,而是选择放出一套高度自由、支持无限重组、支持插件自由增减、支持运行时改造流程的元框架。 三、开源策略的真实底层逻辑。 官方试错为主,社区辅助发散。 现阶段提前开源,Harness 并非完全成熟产品的对外交付。 而是一套公开化的大规模实验方案。 其真实策略可以概括为两点。 一、内部快速迭代。 官方团队以 Cordis 元框架为实验底座。 暴力试错,快速排列组合,验证各类 Agent 工作流,主动筛选最优范式、最优工具链、最优调度逻辑。 二、借助社区扩大试错边界,通过开源释放自由度,让全球开发者提供多样化插件、差异化流程、特殊场景组合,补足官方内部场景覆盖不足的问题。 这是一种典型的双管齐下加速探索的工程策略。 在时间紧迫、赛道竞争激烈、自身尚未定型的背景下,这种做法可以最快速度完成范式探索。 抢占行业话语权,构建开发者生态,制造行业热度。 四、核心问题与本质矛盾,可以借社区发散,不能靠社区收敛。 虽然现阶段开源探索具备合理性,但必须认清一个核心边界。 开源社区天然擅长发散、创新、试错、多元化探索。 但是完全不具备自发收敛、标准化、工程定型的能力。 社区可以贡献插件、贡献新思路、贡献特殊场景方案,但无法替官方完成统一交互范式,固化核心工作流,模型与框架双向适配,标准化数据闭环构建,工业级最优解筛选与定型。 这也是当前这套开源策略最大的隐患。 高度自由的元框架开放后,社区会衍生出成千上万种差异化的 Harness 运行形态,导致全网调用范式再次碎片化。 如果长期放任这种发散状态,无法沉淀统一基线,最终会反向伤害模型迭代的数据纯净度与迭代效率。 因此可以得出关键判断。 试错可以交给社区,收敛必须官方自己完成。 社区只能作为探索的辅助增量,绝对无法替代官方的核心研发与标准化工作。 如果官方自身投入不足、迭代迟缓、过度依赖社区自发优化,这套看似先进的架构最终只会停留在无数零散实验案例。 无法沉淀真正可用的行业标准。 五、最终整体判断:一、Deepseek 自研 Harness 是绝对正确的战略选择。 补齐模型前端调度闭环。 统一交互范式,构建可归因的训练数据链路,是未来大模型能力突破的核心前提。 二、现阶段开源属于提前布局,实验优先、热度优先的阶段性选择。 在无最优标准答案、赛道竞争激烈、需要快速占位的背景下,通过高自由度元框架实现内外双试错。 具备明确的工程合理性与商业合理性。 三、当前架构定位是实验平台,而非最终成品。 Cordis 架构的高灵活、高可组合特性适合探索迭代,但不适合长期无限制开放用于生产。 过度自由会破坏范式收敛,制造数据混乱。 四、行业最终赢家一定是自己能发散、自己能收敛的团队。 社区负责探索可能性,官方负责筛选唯一性。 发散可以开源,收敛必须集权。 结语,Deepseek harness 当前的开源行为,并非草率跟风,也并非完全激进造势,而是在行业未定、节奏紧迫、范式未知的阶段,做出的一次权衡取舍式的阶段性最优解。 未来该项目的真正成败,不取决于社区造出多少花样插件,而取决于官方能否在海量发散探索中快速筛选固化。 收敛出唯一稳定、统一、可闭环迭代的官方标准范式。 可以借天下之力试错,不可靠天下生态定型。 探索可以开放,标准必须自研。
英文翻译
In-depth technical thinking, analyzing the inevitability, trade-offs and industry limitations of Deepseek Harness's open source strategy. I. Inevitability of the Times: Large models have entered the dual-wheel drive era of model plus scaffolding. The upper limit of current large model capabilities is no longer simply determined by model weights, training data and parameter volume. Scaffolding capabilities such as front-end scheduling paradigm, tool calling logic, context clipping strategy, and task decomposition process directly determine the final performance of the same model. It is clear at the current technical stage that the same large model mounted on different Agent scaffolding can have a multiple-fold gap in final task success rate, Token utilization, inference stability, and complex problem-solving ability. Taking Claude Code as a benchmark, it can be clearly seen that the capability release of top-level models relies on two-way in-depth binding and joint iteration of the model and the exclusive scheduling framework. Fixed tool paradigm, unified interaction habits, and stable context structure can generate highly attributable, clean training data that can be used for continuous fine-tuning, forming a perfectly self-consistent closed-loop for technical iteration. Based on this industry trend, Deepseek's launch of self-developed Harness has extremely strong strategic inevitability. Long-term reliance on external frameworks such as OpenAI, Hermes third-party Cloud Code wrapping to call models will cause fatal problems. Models need to passively adapt to a wide range of scheduling logics, tool habits and task processes. The highly fragmented calling paradigms of different communities, different frameworks and different users lead to mixed training data, unattributable failure cases, divergent and chaotic model iteration directions. Without an exclusive official scaffolding, there is no unified evolution baseline. Therefore, it is the only way for long-term development for Deepseek to develop Harness independently and establish its own Agent operation paradigm. II. The essence of the architecture is an advanced experimental meta-framework, not a finalized production-grade framework. Deepseek Harness is built based on Cordis architecture, and its core advantages focus on event-driven decoupling: pluggable components, complete lifecycle management, and reversible side effect system. The greatest value of this architecture is not mature implementation capability, but extremely strong experimental capability and combination capability. There is no absolutely optimal standard paradigm in the current coding agent track in the industry. There is no single standard answer for Agent's tool combination, retrieval strategy, calling order, task decomposition logic, and error rollback mechanism. This type of engineering system has a strong modulation attribute, similar to formula debugging: the basic principle is fixed, but the optimal collocation, optimal process and optimal combination require massive trial and error, permutation and combination, and repeated iteration to gradually converge. Precisely because there is no standard answer or unified optimal solution in the industry, Deepseek did not launch a solidified, closed, hard-coded finished framework at this stage, but chose to release a meta-framework with high freedom, support for infinite restructuring, free addition and deletion of plugins, and support for runtime process modification. III. The real underlying logic of the open source strategy: official-led trial and error, with the community assisting in divergent exploration. Open sourcing Harness in advance at the current stage is not the external delivery of a fully mature product, but a public large-scale experimental plan. Its real strategy can be summarized into two points: 1. Rapid internal iteration. The official team takes the Cordis meta-framework as the experimental base, conducts massive trial and error, quickly conducts permutation and combination, verifies various Agent workflows, and actively screens out the optimal paradigm, optimal tool chain and optimal scheduling logic. 2. Expand the boundary of trial and error with the help of the community. By releasing freedom through open source, global developers can provide diversified plugins, differentiated processes and special scene combinations, making up for the problem of insufficient coverage of official internal scenarios. This is a typical two-pronged engineering strategy to accelerate exploration. Against the background of tight time, fierce industry competition, and undetermined positioning of the product itself, this approach can complete paradigm exploration at the fastest speed, seize industry discourse power, build developer ecology, and create industry popularity. IV. Core problems and essential contradictions: divergence can be achieved with the help of the community, but convergence cannot rely on the community. Although open source exploration is reasonable at the current stage, a core boundary must be recognized: the open source community is naturally good at divergence, innovation, trial and error, and diversified exploration, but it completely lacks the ability of spontaneous convergence, standardization and engineering finalization. The community can contribute plugins, new ideas, and special scene solutions, but it cannot replace the official to complete the unification of interaction paradigm, solidification of core workflow, two-way adaptation of model and framework, construction of standardized data closed loop, and screening and finalization of industrial-level optimal solutions. This is also the biggest hidden danger of the current open source strategy. After the highly free meta-framework is opened, the community will derive thousands of differentiated Harness operation forms, leading to the fragmentation of the whole network calling paradigm again. If this divergent state is allowed for a long time, a unified baseline cannot be precipitated, which will eventually reversely damage the data purity and iteration efficiency of model iteration. Therefore, a key judgment can be drawn: trial and error can be handed over to the community, but convergence must be completed by the official itself. The community can only be used as an auxiliary increment for exploration, and can never replace the official's core R&D and standardization work. If the official has insufficient investment, slow iteration, and excessive reliance on spontaneous optimization by the community, this seemingly advanced architecture will eventually only stay in numerous scattered experimental cases, and cannot precipitate a truly usable industry standard. V. Final Overall Judgment: 1. Deepseek's self-developed Harness is an absolutely correct strategic choice. Completing the closed loop of front-end scheduling of the model, unifying the interaction paradigm, and building an attributable training data link are the core prerequisites for the breakthrough of large model capabilities in the future. 2. Open sourcing at the current stage is an advance layout, a phased choice that prioritizes experimentation and popularity. Against the background of no optimal standard answer, fierce industry competition, and the need for rapid market occupation, realizing internal and external dual trial and error through a high-freedom meta-framework has clear engineering rationality and commercial rationality. 3. The current architecture is positioned as an experimental platform, not a final finished product. The high flexibility and high composability of Cordis architecture are suitable for exploration and iteration, but not suitable for long-term unrestricted opening for production. Excessive freedom will damage paradigm convergence and cause data chaos. 4. The ultimate winner of the industry must be a team that can achieve both divergence and convergence on its own. The community is responsible for exploring possibilities, and the official is responsible for screening the unique solution. Divergence can be open source, but convergence must be centralized. Conclusion: Deepseek Harness's current open source behavior is neither hasty following nor completely radical hype, but a trade-off based phased optimal solution made at a stage when the industry is undetermined, the pace is tight, and the paradigm is unknown. The real success or failure of the project in the future does not depend on how many fancy plugins the community creates, but on whether the official can quickly screen and solidify from massive divergent exploration, and converge to a unique, stable, unified, closed-loop iterable official standard paradigm. We can draw on the power of all parties for trial and error, but cannot rely on the public ecology for finalization. Exploration can be open, but standards must be developed independently.
back to top