GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2
10 hours ago
- #Computational Efficiency
- #AI Scaling
- #Model Hallucination
- Major AI labs are growing skeptical of scaling model size and training data endlessly.
- Despite larger models achieving high benchmark scores, intelligence appears to plateau, as seen with open-weight models nearing proprietary ones.
- Larger models, like DeepSeek V4 Pro and GPT-5.5, show high hallucination rates, failing to admit uncertainty or recognize logical flaws.
- Increased model size can lead to worse real-world accuracy and truthfulness, despite superior benchmark performance.
- The AI industry faces a trilemma: balancing raw capability, uncertainty calibration (hallucination rate), and computational efficiency.