Training Compute Allocation

Training compute allocation is the problem of deciding how scarce cards and infrastructure should be divided among research exploration, pretraining, post-training, evaluation, and agent rollouts. In 138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权, Luo Fuli / 罗福莉 says agent workflows make card count a sharper bottleneck because agents speed up idea generation, code implementation, and experiment setup.

The source’s concrete heuristic is that a frontier-style team aiming at high agent capability may need research, pretraining, and post-training resources closer to a 3:1:1 ratio, and that top teams may already be moving pretraining and post-training investment toward near parity. The claim is not just about more compute; it is about more parallelism once ML Coding and Agent Post-Training lower the human time needed to produce new experiments.

Key Claims

• Faster idea-to-code loops move the bottleneck from researcher time toward cards, evaluation, and experiment scheduling.
• Agent-era compute demand includes post-training and rollout, not only base-model pretraining.
• Teams need enough research compute to test many ideas in parallel before committing to a large training direction.
• The value of compute depends on Research Taste, because cheap experiments can still waste scarce infrastructure if the question is weak.
• Compute allocation is an AI Organization Design issue as well as an infrastructure issue: rigid group boundaries can prevent resources and people from moving to the current bottleneck.

Connections

• Luo Fuli / 罗福莉, Xiaomi, and Memo VR — source team and model context.
• Frontier Model Scaling — broader scaling pressure around parameters, data, and training efficiency.
• Agent Post-Training and Agent RL — post-training and rollout demand that changes allocation.
• ML Coding, Research Taste, and Problem Definition In Research — faster experiment loops and idea quality.
• AI Inference Cost Structure and Model Routing Cost Control — serving-side counterpart where model cost must be routed by task value.