Training Autopilot

Training autopilot is the long-term product form [[LiuZiming|Liu Ziming]] describes 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. In the source, the short-term product is a training copilot, while the eventual autopilot would let a user state a need and budget, then have the system design, train, deploy, and deliver a model.

This is not just a friendlier MLOps interface. Liu links it to AI For AI, OPHIS Research Workflow, and Meta-Model Training Curve Prediction because the system needs to decide what architecture, data, optimizer, experiment path, and delivery format fit the user’s goal and constraints.

Key Claims

• Training copilot is the near-term assistive product; training autopilot is the end-to-end automation target.
• The product needs research judgment, experiment planning, compute budgeting, training execution, and deployment handoff.
• Its economic promise is model-training democratization, but the source acknowledges that serious training remains cost-constrained.
• Better Physics Of AI could lower training costs if current recipes are meaningfully suboptimal.
• The source leaves user demand partly unresolved: it is not yet clear what everyone would train if training became easier.

Connections

• [[LiuZiming|Liu Ziming]] and [[YuanhuanIntelligence|Yuanhuan Intelligence]] — source proposer and company context.
• AI For AI, Auto Research, and Physics Of AI — underlying research automation.
• Meta-Model Training Curve Prediction, OPHIS Research Workflow, and Training Compute Allocation — technical requirements.
• Vibe Training and Vibe Coding — user-facing analogy.