AI 新闻现实检验:Yegge 关停 Gas Town,Databricks 用 Astra 成本涨 60%
Latent Space 的 AINews 汇总了 9 月 15 日至 16 日的 AI 动态,其中两个案例构成对编码智能体热潮的现实检验。Steve Yegge 关停了 Gas Town,并承认尽管每月在编码智能体订阅上花费数千美元,他只用它建成了 Gas Town 这一个项目。Databricks 的 Patrick Wendell 表示已向约 3500 名工程师推广 GPT-6 Astra,Astra 在复杂系统设计和长周期任务上明显优于 Opus 5 与 Sol 5.6,但整体编码支出增加约 60%,公司为此设立了专门的 Astra 子预算。本期还提到 OpenAI 发布模型失准事件披露框架及六份案例报告、DeepMind Institute 成立、小米 MiMo-V2.6 公开 RL 训练看板,以及 Anthropic 将 Claude Cowork 与聊天合并为统一入口。
y already exhibit “weak” or partial RSI when AI assists researchers or iteratively improves prompts/tools, but “real” RSI would require a complete closed loop. The linked paper is arXiv:2609.14858v1 , which commenters interpret as relevant to AI-discovery automation but not yet model-level self-improvement. There is interest in whether the same technique could transfer from prompt/policy/harness optimization to model development itself, especially in open-source agentic frameworks. The implied t
匹配字段正文

