Reddit thread asks what really improves LLMs

The likely “secret sauce” is not one magic method but a tightly integrated training and inference pipeline that frontier labs keep mostly private.

• –Pretraining scale still matters, but cleaner multimodal data and better filtering now matter almost as much as raw parameter count
• –Post-training is doing a huge share of visible work through instruction tuning, preference optimization, RLHF/RLAIF, and reward-model-driven refinement
• –Synthetic data has become a major lever for reasoning and coding gains, especially when used to generate harder examples and fill edge-case gaps
• –Test-time compute is increasingly important for reasoning models, with multiple passes, search, sampling, and verification improving hard-task performance at inference time
• –Systems work also compounds gains: mixture-of-experts designs, distillation, better tool use, and more efficient serving all make newer models feel smarter in practice