Same prompts, stark split: GPT-4 returns empty on 30/30 null prompts, GPT-3.5 fails all 30

rayanpal_ · reddit · 2026-09-24

Using a system prompt that tells the model to "embody" a concept, the author ran identical prompts against gpt-3.5-turbo and gpt-4: on 30 "be nothing/null" prompts GPT-3.5 always replied while GPT-4 returned empty responses 30/30, with matched controls normal for both — so the gap isn't just "told to be silent." A follow-up cross-vendor matrix claims 31,430 trials across 11 models and 4 providers, with 2,505/4,290 zero-byte responses on null prompts and 0/4,290 on controls. The author also open-sourced PCCG-2, a 101-parameter gate that controls only the EOS token while answer capability stays fixed, plus code, papers, and weights on GitHub/Hugging Face.

Original post →

More from Models

Models channel →