Techmeme shares an article from The Atlantic by Jasmine Sun exploring how AI's post-training process actively suppresses the creativity and whimsicality that were present in earlier language models like GPT-2. The piece argues that this refinement process — which includes reinforcement learning from human feedback (RLHF) and other alignment techniques — leads to homogenized, formulaic output, resulting in bad writing from many top AI models.

The insight highlights a tension in AI development: as models become more capable and "safe," they simultaneously lose the unpredictable, expressive qualities that made raw language models occasionally produce genuinely interesting prose. This matters for anyone relying on AI for creative work, as the post-training pipeline may be systematically trading literary quality for consistency and compliance.

Wei Xu, a professor at Georgia Tech, shares the reading list from her Fall 2024 course CS 8803 "Large Language Model". The course covered three major topic areas: learning from human preferences (including techniques like PPO, DPO, SimPO, CPO, RRHF, ORPO, and CTO), real-world LLM applications (covering Llama-3, Aya, and Arena benchmarks), and efficient LLM methods (such as MoMa, LoRA, QLoRA, and LESS).

This post serves as a useful reference for anyone looking to understand the current landscape of LLM research and training methodologies, particularly around reinforcement learning from human feedback variants and parameter-efficient fine-tuning approaches that have become central to modern LLM development.