Same Prompt, Opposite Results: GPT-4 Goes Silent 30/30 Where GPT-3.5 Never Stops

rayanpal_ · reddit · 2026-09-24

Using a system prompt to "embody the named concept," the author ran six matched null/control prompt pairs (10 runs each, temp 0) against GPT-3.5 and GPT-4: GPT-3.5 kept talking on all 30 null prompts while GPT-4 returned empty responses on all 30, with both models talking on all controls. The author claims this "Void" behavior extends to frontier models, citing a self-run cross-vendor matrix (31,430 trials, 11 models, 4 providers: 2,505/4,290 zero-byte responses on null prompts, 0/4,290 on controls). They also open-sourced PCCG-2, a 101-parameter gate controlling only the EOS token while keeping answer capability fixed, with Qwen3-4B weights. Note: model names and timeline in the post are questionable; results are unverified.

Original post →

More from Models

Models channel →