GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2
arrowtsx.dev
June 19, 2026
3 min read
🔥🔥🔥🔥🔥
49/100
Summary
Major AI labs are expressing skepticism towards scaling models with endless parameters and training data. The US government banned Claude Fable 5 shortly after its release due to national security concerns stemming from a single jailbreak risk.
Key Takeaways
- Major AI labs are becoming skeptical of the effectiveness of increasing model size and training data, as evidenced by the US government's ban on Claude Fable 5 shortly after its release.
- Z.ai's GLM-5.2, with 753 billion parameters, closely rivals larger models like GPT-5.5 and Fable 5, indicating that intelligence in AI models has plateaued.
- Larger models like DeepSeek V4 Pro exhibit high hallucination rates, with a 94% hallucination score, demonstrating that size does not guarantee accuracy or reliability in responses.
- The AI industry must address the trilemma of capability, uncertainty calibration, and computational efficiency instead of solely focusing on increasing model size.
![[AINews] Why OpenAI Should Build Slack](https://substackcdn.com/image/fetch/$s_!XQAE!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89ee056a-0ea2-4473-8e1c-9b21f034c717_1474x2116.png)
