Vibe Training

Vibe training is [[LiuZiming|Liu Ziming]]’s analogy to Vibe Coding in 149. 亲历中美 New Labs 资本狂潮，和清华刘子鸣聊：AI for AI、机制可解释性和 Max Tegmark. The imagined user gives a narrative, requirement, or budget-level intent, and a Training Autopilot handles model design, training, deployment, and delivery.

The source treats the idea as aspirational rather than complete. Vibe training would democratize training capability if it worked, but Liu also says large-model training remains expensive and that the eventual user needs are not yet fully clear.

Key Claims

• Vibe training moves model creation from expert-operated training recipes toward intent-driven model delivery.
• It depends on AI For AI because the system must choose experiments and model designs, not only execute a fixed pipeline.
• It depends on Meta-Model Training Curve Prediction when architecture ideas need cheap ranking before expensive runs.
• It could lower the threshold for custom models but cannot remove compute cost by interface design alone.
• The source leaves open whether broad user demand will appear once training becomes easier.

Connections

• Training Autopilot — product system that would implement vibe training.
• Vibe Coding — analogy from programming to model training.
• AI For AI, OPHIS Research Workflow, and Physics Of AI — research automation required underneath.
• Training Compute Allocation and AI Startup Unit Economics — cost and economics constraints.