← 返回 Avalaches

作者探讨了递回自我改进(recursive self-improvement)在自动化电子报繁琐工作中的实用性,挑战了先进 AI 训练仅限于中心化实验室的观念。使用 Andrej Karpathy 的 AutoResearch 专案,作者利用 Claude 从头开始训练一个小型语言模型。在 Nvidia DGX 桌上型超大型电脑的支援下,该过程允许 Claude 自主调整参数和训练方案,最终将该模型从生成无意义的重复内容精炼为产生连贯的输出。

为了实现更具体的实用性,作者使用最近获得 1500 万美元(15 million USD)资金的 Prime Intellect,来训练一个名为 Frontier_Paper_Curator 的专用模型。该模型基于大约 100 篇历史电子报内容进行训练,利用了合成数据和强化学习,并由另一个 AI 评估其输出。Prime Intellect 的执行长 Vincent Weisser 指出,他们的目标是将训练基础设施民主化,培育一个由十亿个专业智慧组成的去中心化生态系统,而不是单一的中心化权威。

其他公司,例如由执行长 Sara Hooker 领导的 Adaption,也将 AI 模型训练自动化,以帮助企业减少 Token 使用成本。这种民主化降低了依赖单一前沿模型的风险,当 Anthropic 封锁对 Fable 5 的某些请求时,这一脆弱性显露无遗,这也呼应了 Palantir 执行长 Alex Karp 对于数据所有权流失的警告。作者最终得到的模型在摘要复杂的研究论文方面展示了令人期待的能力,指明了自动化繁琐工作流程的可行路径。

The author explores the utility of recursive self-improvement for automating newsletter busywork, challenging the notion that advanced AI training remains exclusive to centralized labs. Using Andrej Karpathy's AutoResearch project, the author used Claude to train a small language model from scratch. Supported by an Nvidia DGX desktop supercomputer, the process allowed Claude to autonomously adjust parameters and training regimes, ultimately refining the model from generating nonsensical repetitions into producing coherent outputs.

To achieve more practical utility, the author used Prime Intellect, which recently secured 15 million USD in funding, to train a specialized model named Frontier_Paper_Curator. Trained on approximately 100 historical newsletter entries, the model utilized synthetic data and reinforcement learning, with its output evaluated by a separate AI. Prime Intellect CEO Vincent Weisser notes that their goal is to democratize training infrastructure, fostering a decentralized ecosystem of a billion specialized intelligences rather than a single, centralized authority.

Other companies, such as Adaption led by CEO Sara Hooker, also automate AI model training to help enterprises reduce token usage. This democratization reduces the risk of relying on a single frontier model, a vulnerability highlighted when Anthropic blocked certain requests to Fable 5, and echoed by Palantir CEO Alex Karp's warnings regarding data loss. The author’s resulting model showed promising capabilities in summarizing complex research papers, indicating a viable path toward automating tedious workflow tasks.

2026-07-09 (Thursday) · 552521a3a5d9f71f039f620fed062ec6249d2fc0