Thinking Machines 实验室发布 AutoInfra：面向并行实验的训练基础设施
新基础设施把训练实验全流程交给智能体跑，LoRA 训练成本只要 Tinker 标价的不到一成，做实验的同学可以看看。
AutoInfra 是一套面向快速并行实验的训练基础设施，由 TrainForge 智能体决定实验内容，并通过与 Tinker 兼容的训练 API 在共享 GPU 资源上执行。以 GPT-OSS-120B 的 LoRA 训练为例，每百万 tokens 成本为 $0.199，对比 Tinker 标价 $2.33 下降 91.5%，按 1T 训练 tokens 估算可节省约 $2.13M。该系统同时训练驱动 TrainForge 与 AutoR&D-Engineer 的模型，使智能体的实验发现回流到下一版模型。算法、训练和沙箱环境相互隔离，用于防范自我改进循环中的 reward hacking。
AutoInfra: Infrastructure for rapid, parallel experimentation
https://t.co/DxT05IqobY
TrainForge decides what to try. AutoInfra carries each experiment from request to result across shared GPU capacity, through a single Tinker-compatible training API.
It also trains the models that power TrainForge and AutoR&D-Engineer, so what those agents discover feeds directly into the next model version.
→ GPT-OSS-120B LoRA training at $0.199 per million tokens on AWS, compared with $2.33 at Tinker's listed price → 91.5% lower cost per token, an estimated $2.13M saved per 1T training tokens → Isolated algorithm, training, and sandbox environments guard against reward hacking across self-improvement cycles
#FutureOfAI #EnterpriseAI
