AI HOT Daily · May 26, 2026

Apple 1.2T Google model to reshape Siri, GPT-5.6 leaks 1.5M token context, Huawei Ascend 1.58-bit edge model, Grok Build Beta, Anthropic co-founder at Pope encyclical.

1. 1. Huawei Ascend 1.58-bit Edge Model BitCPM-CANN

   BitCPM-CANN, China first 1.58-bit edge model trained on Huawei Ascend chips, open-sourced by consortium including Tsinghua.

2. 2. Grok Build Beta Opens to SuperGrok

   xAI launched Grok Build Beta with Plan Mode, image-video generation via Imagine, and CLI automation.

3. 3. Qwen3.7-Max Implicit Cache Launched

   Alibaba Qwen3.7-Max implicit caching launched, auto-enabled for faster and cheaper inference.

4. 4. Apple Customizes 1.2T Google Model for Siri

   Apple using customized 1.2T-parameter Google model for next-gen Siri, larger than Gemini 3.5 Flash 300B.

5. 5. Pope Warns of AI Risks in Encyclical

   Pope Leo XIV warned of AI-driven warfare and labor impacts, calling for legal and ethical frameworks.

6. 6. GPT-5.6 Leaked: 1.5M Token Context

   Unannounced GPT-5.6 (codename iris-alpha) supports 1.5M token context, responds smoothly at 900K tokens, expected June release.

7. 7. Huawei “Tao Law” Logic Folding Chip Tech

   He Tingbo proposed LogicFolding at ISCAS 2026, boosting transistor density by 53% and energy efficiency by 41%.

8. 8. Chris Olah Speaks at Pope Encyclical

   Anthropic co-founder Olah noted frontier AI labs face commercial, research, and geopolitical pressures requiring external oversight.

9. 9. AI Agent Terms: Harness vs Scaffold

   Hugging Face clarifies AI agent terms: models have no memory-loops, Scaffolding defines behavior, Harness controls execution loops.