ChatGPT suddenly 'misunderstands' simple prompts — distillation guardrails suspected
TerminalNoop · reddit · 2026-10-23
While trying to improve a prompt with ChatGPT to probe unreleased Qwen models on LMArena, the author found ChatGPT repeatedly failing at simple instructions — e.g. turning a request to improve a 'build prompt' into an image-generation prompt. They speculate this could be new guardrails against distillation attacks. Unverified personal observation.
More from Fun
- Hackers pop the iPhone Duo on release day, add a much-needed feature — laparisa · 2026-10-23
- Dean's Account Keeps "Critically Damaging" OpenAI With Takes — Intentional? — basedjensen · 2026-10-23
- AI-generated WD-40 Joint Care supplement packaging is hilariously convincing — Delahuntagram · 2026-10-23
- The Jensen effect: MicroDucks sells 4,000+ units in 24h after GTC showcase — andimarafioti · 2026-10-23
- Anthropic Says Newer Models Are "Less Likely to Begin an Attack," Drawing Ridicule — AIFlow_ML · 2026-10-23
- waitbutwhy's classic AI explosion chart: from 'monkey tricks' to 'The fuck??' — pickover · 2026-10-23