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

顶尖模型公司上市前在忙什么

视频

音频

原始脚本

顶尖模型公司在上市前忙什么？近期两大顶尖 AI 企业相继释放关键信号。
Anthropic 发布 When AI builds itself。
披露超80%代码由 claude 自主编写，模型已实现递归自我提升，自主开展安全实验，正式落地大模型闭环自迭代范式。
OpenAI 内部则定下明确目标，计划将每周模型实验迭代次数从30次提升至300次，实现10倍迭代提速。
核心依靠全链路自动化与 AI 辅助完成高频试错。
两家企业动作同步绝非偶然，背后既是行业技术路线的集体转向，也暗含商业布局与行业竞争的深层考量。
一、技术路线趋同。
数据驱动成主流，自动化迭代替代传统研发。
无论是 anthropic 的递归自我提升，还是 openai 追求的十倍迭代效率。
二者核心路径高度统一，自动化流水线加模型自产数据加增量微调，底层架构重构、全新算法设计、大规模代码改版，都需要人类深度研发、多轮评审与长期调试。
完全无法支撑每周数百次的高频实验。
唯有以数据为核心的轻量化迭代，才能实现快速试错、快速落地。
这也印证了当下行业共识。
现阶段头部企业的发力重心是深挖现有模型潜力，而非冒险探索全新架构。
与此同时，两家企业对外释放相关进展也带有鲜明的商业诉求。
目前两家公司均推进上市进程，AI自主进化，逐步逼近 AGI 的叙事。
能够充分放大技术成长性，拉高资本市场估值预期。
对外宣传始终把握分寸，只展示落地成果，不宣称实现通用人工智能。
在营造想象空间的同时，规避宣传风险。
对内则是持续打磨工程体系与迭代链路，抢先掌握下一代模型的效率优势。
二、迭代核心逻辑。
以失败样本为抓手，走少而精的数据路线。
这套高速迭代体系能够高效运转，关键在于数据筛选逻辑的升级。
也就是闻过则喜的样本使用思路。
常规正向样本价值已经边际递减，这类内容对应模型原本就能稳定完成的任务。
相当于在拟合曲线的平滑区间重复打点，不仅难以推动能力突破，还容易引发过拟合问题。
真正推动模型精细化升级的，是失败、出错、逻辑断裂。
产生幻觉的负面样本。
从函数拟合视角来看，模型本质是复杂高维拟合函数，真实的语义与逻辑是一条平滑曲线。
而模型出错的场景，正是曲线上的畸变拐点，欠拟合区域与能力边界。
将这些精准的问题样本筛选整理后，用于微调。
相当于在偏差最大的位置高密度补点，既能修正局部误差，抹平毛刺，也能清晰界定模型能力边界，补齐逻辑漏洞。
这种模式遵循少而精原则，无需堆砌海量数据。
少量高价值问题样本就能实现明显效果，恰好适配高频迭代的节奏，让每一轮实验都直击模型短板。
三。
自动化体系落地， AI 辅助编码，搭建全链路高速试错通道， AI 参与编码，全流程自动化，是十倍迭代提速的硬件支撑。
Anthropic 大量代码由 Claude 生成，OpenAI 实现实验频次跨越式增长。
模型参与开发的核心目的并非设计新算法与新架构。
而是搭建服务于数据迭代的自动化流水线。
模型生成的代码主要用于批量执行评测任务，自动抓取失败样本，完成数据清洗与格式化。
一键启动微调实验，自动验收迭代效果。
人工则从重复的数据搬运、流程部署、结果统计中解放出来，仅负责制定规则、把控方向。
处理极端问题。
整条研发链路从人工主导转变为 AI 自主流转，迭代周期被大幅压缩。
而每周数百次的高频实验内容也集中在轻量化调整。
比如优化失败样本的配比、调整微调权重、测试不同学习率等超参。
这类改动风险低，落地快。
可并行开展和架构重构，底层代码改写形成明显区分，是保障高迭代频率的关键。
四、本质约束。
迭代再快也难突破模型固有能力天花板高速自动化迭代。
可以缩短模型优化周期，却无法改变系统的底层约束，依旧逃不开能力上限。
所有用于微调的样本、实验用的测试案例。
全部来自当前模型的输出。
这意味着模型只能修复已知错误，无法触及自身认知盲区之外的全新问题。
按照数学埃普斯隆德尔塔极限原理与芝诺悖论逻辑，持续的同源样本微调，只会让模型拟合曲线不断向当前架构的最优解收敛。
迭代次数哪怕再提升数十倍，也只是加快趋近天花板的速度。
而非突破天花板。
短期来看，迭代效率的差距会直接转化为产品体验与市场竞争力的代差。
但长期而言，当现有架构的潜力被挖掘殆尽。
单纯依靠自动化与数据迭代的增长终将停滞。
届时行业想要继续突破，依然要回归人类主导的架构革新、底层算法创新。
五、行业博弈，稳健挖掘存量潜力。
理性看待 AGI 蓄势当前，头部企业集体压住数据驱动的自我迭代，是非常务实的行业选择。
在全新底层技术范式尚未出现前，最大化盘活现有大模型架构的潜力，远比投入不确定性极高的前沿探索更稳妥，也更符合资本市场对稳健增长的偏好。
至于逼近 AGI 的相关说法，更多是技术进展与商业宣传结合后的叙事。
当下的递归自我迭代、高频自动化试错，属于模型的收敛式优化，仅能实现现有能力的打磨精进，并不具备自主重构逻辑、创造全新范式的能力。
距离真正的通用人工智能仍有很远距离。
总结综合两家头部企业的动向不难看出，当下顶尖大模型公司的核心工作就是搭建 AI 自主运转的高速数据迭代闭环。
十倍迭代提速，模型自主编写代码，本质都是服务于用优质失败样本微调模型这一核心目标。
这套模式能在短期内快速打磨模型，拉开行业差距。
也是上市阶段塑造价值的重要筹码。
但从底层原理来看，它始终受限于模型固有架构，迭代收益会逐步衰减，无法实现颠覆性突破。
对于整个行业而言，自动化自迭代是现阶段的最优工程方案，却并非通往 agi 的终极答案。

修正脚本

顶尖模型公司在上市前忙什么？近期两大顶尖 AI 企业相继释放关键信号。
Anthropic 发布 When AI builds itself。
披露超80%代码由 claude 自主编写，模型已实现递归自我提升，自主开展安全实验，正式落地大模型闭环自迭代范式。
OpenAI 内部则定下明确目标，计划将每周模型实验迭代次数从30次提升至300次，实现10倍迭代提速。
核心依靠全链路自动化与 AI 辅助完成高频试错。
两家企业动作同步绝非偶然，背后既是行业技术路线的集体转向，也暗含商业布局与行业竞争的深层考量。
一、技术路线趋同。
数据驱动成主流，自动化迭代替代传统研发。
无论是 anthropic 的递归自我提升，还是 openai 追求的十倍迭代效率。
二者核心路径高度统一，自动化流水线加模型自产数据加增量微调，底层架构重构、全新算法设计、大规模代码改版，都需要人类深度研发、多轮评审与长期调试。
完全无法支撑每周数百次的高频实验。
唯有以数据为核心的轻量化迭代，才能实现快速试错、快速落地。
这也印证了当下行业共识。
现阶段头部企业的发力重心是深挖现有模型潜力，而非冒险探索全新架构。
与此同时，两家企业对外释放相关进展也带有鲜明的商业诉求。
目前两家公司均推进上市进程，AI自主进化，逐步逼近 AGI 的叙事。
能够充分放大技术成长性，拉高资本市场估值预期。
对外宣传始终把握分寸，只展示落地成果，不宣称实现通用人工智能。
在营造想象空间的同时，规避宣传风险。
对内则是持续打磨工程体系与迭代链路，抢先掌握下一代模型的效率优势。
二、迭代核心逻辑。
以失败样本为抓手，走少而精的数据路线。
这套高速迭代体系能够高效运转，关键在于数据筛选逻辑的升级。
也就是闻过则喜的样本使用思路。
常规正向样本价值已经边际递减，这类内容对应模型原本就能稳定完成的任务。
相当于在拟合曲线的平滑区间重复打点，不仅难以推动能力突破，还容易引发过拟合问题。
真正推动模型精细化升级的，是失败、出错、逻辑断裂、产生幻觉的负面样本。
从函数拟合视角来看，模型本质是复杂高维拟合函数，真实的语义与逻辑是一条平滑曲线。
而模型出错的场景，正是曲线上的畸变拐点，欠拟合区域与能力边界。
将这些精准的问题样本筛选整理后，用于微调。
相当于在偏差最大的位置高密度补点，既能修正局部误差，抹平毛刺，也能清晰界定模型能力边界，补齐逻辑漏洞。
这种模式遵循少而精原则，无需堆砌海量数据。
少量高价值问题样本就能实现明显效果，恰好适配高频迭代的节奏，让每一轮实验都直击模型短板。
三、
自动化体系落地， AI 辅助编码，搭建全链路高速试错通道， AI 参与编码，全流程自动化，是十倍迭代提速的硬件支撑。
Anthropic 大量代码由 Claude 生成，OpenAI 实现实验频次跨越式增长。
模型参与开发的核心目的并非设计新算法与新架构。
而是搭建服务于数据迭代的自动化流水线。
模型生成的代码主要用于批量执行评测任务，自动抓取失败样本，完成数据清洗与格式化。
一键启动微调实验，自动验收迭代效果。
人工则从重复的数据搬运、流程部署、结果统计中解放出来，仅负责制定规则、把控方向、处理极端问题。
整条研发链路从人工主导转变为 AI 自主流转，迭代周期被大幅压缩。
而每周数百次的高频实验内容也集中在轻量化调整。
比如优化失败样本的配比、调整微调权重、测试不同学习率等超参。
这类改动风险低，落地快，可与架构重构、底层代码改写形成明显区分，是保障高迭代频率的关键。
四、本质约束。
迭代再快也难突破模型固有能力天花板。高速自动化迭代
可以缩短模型优化周期，却无法改变系统的底层约束，依旧逃不开能力上限。
所有用于微调的样本、实验用的测试案例。
全部来自当前模型的输出。
这意味着模型只能修复已知错误，无法触及自身认知盲区之外的全新问题。
按照数学埃普斯隆德尔塔极限原理与芝诺悖论逻辑，持续的同源样本微调，只会让模型拟合曲线不断向当前架构的最优解收敛。
迭代次数哪怕再提升数十倍，也只是加快趋近天花板的速度。
而非突破天花板。
短期来看，迭代效率的差距会直接转化为产品体验与市场竞争力的代差。
但长期而言，当现有架构的潜力被挖掘殆尽。
单纯依靠自动化与数据迭代的增长终将停滞。
届时行业想要继续突破，依然要回归人类主导的架构革新、底层算法创新。
五、行业博弈，稳健挖掘存量潜力。
理性看待 AGI 蓄势当前，头部企业集体转向数据驱动的自我迭代，是非常务实的行业选择。
在全新底层技术范式尚未出现前，最大化盘活现有大模型架构的潜力，远比投入不确定性极高的前沿探索更稳妥，也更符合资本市场对稳健增长的偏好。
至于逼近 AGI 的相关说法，更多是技术进展与商业宣传结合后的叙事。
当下的递归自我迭代、高频自动化试错，属于模型的收敛式优化，仅能实现现有能力的打磨精进，并不具备自主重构逻辑、创造全新范式的能力。
距离真正的通用人工智能仍有很远距离。
总结：综合两家头部企业的动向不难看出，当下顶尖大模型公司的核心工作就是搭建 AI 自主运转的高速数据迭代闭环。
十倍迭代提速，模型自主编写代码，本质都是服务于用优质失败样本微调模型这一核心目标。
这套模式能在短期内快速打磨模型，拉开行业差距。
也是上市阶段塑造价值的重要筹码。
但从底层原理来看，它始终受限于模型固有架构，迭代收益会逐步衰减，无法实现颠覆性突破。
对于整个行业而言，自动化自迭代是现阶段的最优工程方案，却并非通往 agi 的终极答案。

英文翻译

Here is the English translation of the provided Chinese text, paragraph by paragraph:

What are top-tier model companies busy with before going public? Recently, two leading AI enterprises have successively released key signals.

Anthropic released "When AI builds itself."
It disclosed that over 80% of the code was autonomously written by Claude, the model has achieved recursive self-improvement, autonomously conducts safety experiments, and has formally implemented a closed-loop self-iteration paradigm for large models.

Internally, OpenAI has set a clear goal: to increase the number of weekly model experiment iterations from 30 to 300, achieving a tenfold improvement in iteration speed.
The core relies on full-link automation and AI assistance to complete high-frequency trial and error.

The synchronized actions of the two companies are no accident. Behind this lies not only a collective shift in the industry's technical roadmap but also deep considerations of commercial layout and industry competition.

1.  **Convergence of Technical Roadmaps.**
    Data-driven approaches become mainstream, with automated iteration replacing traditional R&D.
    Whether it is Anthropic's recursive self-improvement or OpenAI's pursuit of tenfold iteration efficiency,
    the core paths of both are highly unified: automated pipelines + model self-generated data + incremental fine-tuning. Underlying architecture reconstruction, entirely new algorithm design, and large-scale code overhauls all require deep human R&D, multiple rounds of review, and long-term debugging.
    They are completely unable to support hundreds of high-frequency experiments per week.
    Only data-centric, lightweight iterations can achieve rapid trial and error and rapid deployment.
    This also confirms the current industry consensus:
    At this stage, the focus of top companies is on deeply unearthing the potential of existing models, rather than冒险探索全新架构.
    At the same time, the release of related progress by the two companies also carries distinct commercial demands.
    Currently, both companies are advancing their listing processes, using the narrative of "AI self-evolution, gradually approaching AGI"
    to fully amplify technological growth potential and raise valuation expectations in the capital market.
    They always maintain discretion in external communications, only showcasing implemented results without claiming to have achieved general artificial intelligence.
    While creating room for imagination, they avoid propaganda risks.
    Internally, they continuously refine engineering systems and iteration chains, seizing the efficiency advantage for next-generation models first.

2.  **Core Logic of Iteration.**
    Use failure samples as a lever, pursuing a "less but better" data strategy.
    The key to the efficient operation of this high-speed iteration system lies in the upgrade of data screening logic,
    namely, the "welcome mistakes" sample usage mindset.
    The value of conventional positive samples has already diminished marginally. Such content corresponds to tasks the model can stably complete originally.
    This is equivalent to repeatedly plotting points in the smooth interval of the fitting curve, which not only fails to drive capability breakthroughs but also easily leads to overfitting.
    What truly drives the refined upgrade of models are negative samples: failures, errors, logical breaks, and hallucinations.
    From the perspective of function fitting, the model is essentially a complex high-dimensional fitting function, where real semantics and logic form a smooth curve.
    The scenarios where the model makes mistakes are precisely the distortion inflection points on the curve: underfitting regions and capability boundaries.
    After screening and organizing these precise problem samples, they are used for fine-tuning.
    This is equivalent to densely supplementing points at the positions with the largest deviations, which can both correct local errors, smooth out burrs, and clearly define the model's capability boundaries, filling logical gaps.
    This model follows the "less but better" principle, without needing to pile up massive amounts of data.
    A small number of high-value problem samples can achieve significant results, perfectly adapting to the rhythm of high-frequency iteration, allowing each round of experiments to directly target the model's shortcomings.

3.  **Implementation of the Automation System.**
    AI-assisted coding builds a full-link high-speed trial-and-error channel. AI participating in coding and full-process automation form the hardware support for a tenfold iteration speed increase.
    A large amount of Anthropic's code is generated by Claude; OpenAI has achieved a leapfrog increase in experiment frequency.
    The core purpose of the model's participation in development is not to design new algorithms or new architectures,
    but to build an automated pipeline serving data iteration.
    The code generated by the model is mainly used to execute evaluation tasks in batches, automatically capture failure samples, and complete data cleaning and formatting.
    One-click launch of fine-tuning experiments, automatic acceptance of iteration results.
    Humans are freed from repetitive data handling, process deployment, and result statistics, only responsible for setting rules, controlling direction, and handling extreme issues.
    The entire R&D chain has shifted from human-led to AI-autonomous flow, significantly compressing the iteration cycle.
    The content of the hundreds of high-frequency experiments per week is also concentrated on lightweight adjustments,
    such as optimizing the ratio of failure samples, adjusting fine-tuning weights, and testing different hyperparameters like learning rates.
    Such changes have low risk and fast implementation, clearly distinguishing them from architecture reconstruction or underlying code rewrites, and are key to maintaining high iteration frequency.

4.  **Inherent Constraints.**
    No matter how fast the iteration, it is difficult to break through the model's inherent capability ceiling. High-speed automated iteration
    can shorten the model optimization cycle, but cannot change the underlying constraints of the system, and it remains unable to escape the capability upper limit.
    All samples used for fine-tuning and the test cases used for experiments
    come entirely from the output of the current model.
    This means the model can only fix known errors and cannot touch entirely new problems outside its own cognitive blind spots.
    Following the mathematical Epsilon-Delta limit principle and Zeno's paradox logic, continuous fine-tuning with homogeneous samples will only cause the model's fitting curve to converge towards the optimal solution of the current architecture.
    Even if the number of iterations is increased by dozens of times, it only accelerates the speed of approaching the ceiling,
    not breaking through it.
    In the short term, the gap in iteration efficiency will directly translate into a generational gap in product experience and market competitiveness.
    But in the long term, when the potential of the existing architecture is exhausted,
    growth relying solely on automation and data iteration will eventually stagnate.
    At that point, for the industry to continue breaking through, it must return to human-led architectural innovation and underlying algorithmic breakthroughs.

5.  **Industry Competition: Steadily Mining Existing Potential.**
    Rationally viewing the current buildup towards AGI, the collective shift of top companies towards data-driven self-iteration is a very pragmatic industry choice.
    Before a new underlying technological paradigm emerges, maximizing the potential of existing large model architectures is far safer than investing in highly uncertain cutting-edge exploration, and aligns better with the capital market's preference for stable growth.
    As for the rhetoric about approaching AGI, it is more of a narrative combining technical progress with commercial promotion.
    The current recursive self-iteration and high-frequency automated trial and error are convergence-type optimizations of the model, which can only refine and polish existing capabilities, and do not possess the ability to autonomously restructure logic or create new paradigms.
    There is still a long way to go before achieving true general artificial intelligence.

**Summary:** Combining the actions of the two leading companies, it is not difficult to see that the core task of top-tier large model companies at present is to build a high-speed data iteration closed loop that operates autonomously by AI.
The tenfold iteration speed increase and the model autonomously writing code are essentially all in service of the core goal of fine-tuning the model using high-quality failure samples.
This model can quickly polish models in the short term, widening the gap in the industry,
and is also an important bargaining chip for shaping value during the listing phase.
However, from a fundamental principle perspective, it is always constrained by the model's inherent architecture, and iteration gains will gradually diminish, making disruptive breakthroughs impossible.
For the entire industry, automated self-iteration is the optimal engineering solution at this stage, but it is not the ultimate answer leading to AGI.