capacity-bottleneck-inverts-between-pretraining-and-alignment
IN derived (depth 3)
Created 2026-06-21T10:10:05+00:00 · Reviewed 2026-06-21T14:41:08+00:00
The transformer's known over-parameterization at the pretraining level (compression to 60% of parameters with 95% performance retained) combined with alignment's inverse priority (scaling reward model size matters more than data volume) reveals a systematic inversion: data volume is the bottleneck for pretraining while model capacity is the bottleneck for alignment.
Summary
Each phase of building a large model hits a different wall: pretraining runs out of data before it runs out of model capacity, while the alignment step runs out of model capacity before it runs out of training examples. This means compute and resource budgets should be allocated very differently across the pipeline, and a scaling strategy that works for one phase will actively backfire in the other.
Justifications
SL — Over-parameterization at pretraining confirms that capacity is the non-binding constraint there, while alignment research independently shows capacity IS the binding constraint at alignment
Antecedents (all must be IN):
- IN pretraining-and-alignment-have-inverted-scaling-priorities — Pretraining and alignment exhibit inverted scaling priorities: in pretraining, data volume outweighs parameter count (Chinchilla, Llama 1 13B beating GPT-3 175B), while in RLHF reward modeling, scaling model size is more effective than scaling comparison dataset size — suggesting different phases of the training pipeline are bottlenecked by fundamentally different resources.
- IN transformer-is-mature-and-over-parameterized — The transformer architecture has reached maturity: post-2017 refinements (Pre-LN, SwiGLU, RMSNorm, GQA) improve stability and efficiency without changing fundamentals, while compression evidence (95% performance at 60% parameters, cross-layer sharing, weight tying) reveals substantial redundancy — future gains come from scaling strategy and efficiency engineering, not architectural revolution.
Dependents
These beliefs depend on this one:
- IN formal-understanding-insufficient-across-llm-stack — Formal theoretical understanding consistently proves insufficient across the entire LLM stack: RLHF's complete mathematical specification fails without dozens of engineering details, prompting's irreducible sensitivity resists formal analysis, the capacity bottleneck inverts between pretraining and alignment stages — and the inversion means that even a correct scaling theory for one stage actively misleads for the next.
- IN training-pipeline-masks-fundamental-capacity-inversion — The mature training pipeline's standardized stages mask a fundamental capacity inversion: pretraining benefits from parameter redundancy (over-parameterized models remain compressible), while alignment is bottlenecked by reward model capacity (scaling the reward model matters more than scaling data), and this asymmetry is hidden by the pipeline's apparent end-to-end reproducibility.