Model Post-Training Bottleneck

Model post-training bottleneck is the episode’s reminder that a useful large model is not finished when pretraining succeeds. In Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授, [[ZhangQi|张奇]] argues that pretraining supplies much of a model’s knowledge, but post-training decides whether that latent knowledge becomes reliable behavior for a task.

The bottleneck is partly data matching. Zhang says that if a model has already remembered the relevant knowledge during pretraining, a small amount of aligned training data may unlock usable answers; if not, adding supervised data later can fail or even disturb other behavior. The source therefore treats post-training as a costly, expert-heavy search process rather than a simple “add labels” phase.

Key Claims

• Post-training must match the knowledge and behavior the base model can actually support.
• More high-quality-looking data is not automatically better if it does not align with what the model already learned.
• Frontier labs’ advantage may include tacit formulas, evaluation know-how, expert labeling, and large-scale trial-and-error after pretraining.
• DeepSeek can reduce the visible cost story around pretraining and inference, while the post-training layer can remain expensive and hard to copy.
• Agent systems intensify the bottleneck because the model must learn reflection, tool use, memory, failure recovery, and environment feedback.

Connections

• Agent Post-Training, Agent RL, and Training Compute Allocation — agent-specific versions of the bottleneck.
• DeepSeek, OpenAI, MOSS, and Frontier Model Scaling — model organizations and scaling context.
• Interleaved Thinking, Long-Horizon AI, and Agentic Workflow — behaviors that require stronger post-training and evaluation.
• Model Workflow Fit, AI Coding Verification, and Human Judgment Under AI — deployment-side evidence that post-training alone is not enough.