AI Frontier Daily · 2026-05-26

Your daily briefing on global AI trends — stay informed in 5 minutes.

Model Releases / Updates

1. ModelBest and Tsinghua University Open Source China’s First 1.58-bit Edge Model BitCPM-CANN Trained on Huawei Ascend
IT Home (RSS)

Product Releases / Updates

2. Grok Build Beta Opens to SuperGrok Users
X：xAI (@xai)

Grok Build Beta is now available to all SuperGrok and X Premium+ users. Use Plan Mode, create images and videos with Imagine, and build automations or orchestrators using the CLI. Visit http://x.ai/cli to get started.

3. Qwen3.7-Max Implicit Cache Now Live
X: Tongyi Qianwen / Qwen (@Alibaba_Qwen)

✅ Implicit cache is now live on Qwen3.7-Max — auto-enabled, no setup required. ⚡️ Faster and cheaper out of the box. Need higher, more deterministic hit rates? Try explicit cache. 🙌 🔗 Best Practices 🔗: https://www.alibabacloud.com/help/en/model-studio/explicit-cache-best-practice

Industry News

4. Apple Reportedly Using Custom 1.2T-Parameter Google Model to Reshape Next-Gen Siri
X：Kim (@kimmonismus)

According to reports, Apple is using a custom 1.2T-parameter Google model as the core for revamping next-gen Siri — significantly larger than the estimated ~300B-parameter Gemini 3.5 Flash. The model will power parts of Siri, with simple queries expected to run on-device. Apple’s key challenge is ensuring the model responds fast enough for everyday questions. Additionally, several major AI releases are expected next month, including the Apple Intelligence & Gemini integration at WWDC, GPT-5.6, possibly Sonnet 4.8/Opus 4.8, and the confirmed Gemini 3.5 Pro.

5. Pope Leo Calls for “Profound Humanity” in the Age of AI
The Verge：AI（RSS）

Pope Leo XIV warned of the risks of AI and unchecked technological power in his first major encyclical, Magnifica Humanitas. Released on May 15, 2026, the document is the Pope’s declaration on “safeguarding humanity in the age of artificial intelligence.” It focuses on the dangers of AI-driven warfare, AI’s impact on labor, and the need for new legal and ethical frameworks to regulate technology, emphasizing that the rapid proliferation of AI is causing economic and social upheaval, and that insufficient protection of individuals threatens human dignity.

6. OpenAI GPT-5.6 Model Leaks: 1.5M Token Context Window
IT Home (RSS)

Multiple developers discovered an unannounced GPT-5.6 model in OpenAI Codex backend logs, codenamed iris-alpha. The model will support a 1.5 million token context window — approximately 43% higher than the current GPT-5.5’s 1.05M tokens — and is expected to launch in June 2026. Tests show it responds smoothly even with 900K token inputs. Other variants found in the same series include ember-alpha and beacon-alpha. GPT-5.6 also shows improvements in frontend UI generation. Based on these leaks, Anthropic Claude, Google Gemini, and xAI Grok may also be targeting simultaneous new model releases.

7. OpenAI, Grupo Folha, and Grupo UOL Announce Strategic Content Partnership
OpenAI: Official Updates (RSS · Excluding Enterprise/Customer Cases)

OpenAI has partnered with Brazilian media groups Grupo Folha and Grupo UOL to bring trusted Brazilian news content into ChatGPT, expanding access to news while ensuring attribution and transparency of content sources.

Research Papers

8. Microsoft Copilot Cowork Found to Have File Leak Issues
Hacker News Hot (buzzing.cc Chinese Translation)

9. Huawei’s He Tingbo Publishes “Tao’s Law” Paper: LogicFolding Boosts Chip Performance
IT Home (RSS)

Huawei’s He Tingbo proposed “Tao’s Law” at ISCAS 2026, introducing LogicFolding technology. This technology enhances chip performance through 3D spatial topology reorganization without relying on new lithography processes. In Kirin 2026 chip tests, transistor density increased from 155 MTr/mm² to 238 MTr/mm², energy efficiency of performance cores improved by 41%, and maximum clock frequency increased by nearly 13%. The paper shows the Kirin 2027 chip has already reached Silicon status, with subsequent plans including Kirin 2028 and 2029. On the AI chip front, the Ascend 990 plans to introduce LogicFolding around 2030, with hardware integration expected to improve over 100-fold by 2035.

Tips & Insights

10. Anthropic Co-founder Chris Olah Speaks at the Pope’s Encyclical Release
Anthropic: Newsroom (Web)

Anthropic co-founder Chris Olah attended the release of Pope Leo XIV’s encyclical on AI at the Vatican. He noted that all frontier AI labs face commercial, research, and geopolitical pressures that can conflict with doing the right thing, making external oversight critical. He emphasized that AI models are not engineered like aircraft but “grown” from human language and thought, with potentially complex and opaque intrinsic properties. He posed three questions requiring careful consideration: how to ensure global benefits from AI development are shared fairly, how to think about human flourishing in the age of AI, and the nature of AI models’ intrinsic properties. He called on all sectors of society, especially religious and civil organizations, to seriously examine AI development and guide it toward positive outcomes.

11. Clarifying Harness, Scaffold, and AI Agent Terminology
Hugging Face：Blog（RSS）

This article aims to clarify easily confused key terms in the AI agent domain. It points out that models (like Claude, GPT) are themselves memoryless, non-recurrent large language models. Their behavior is shaped by “Scaffolding” (the behavior definition layer, such as system prompts and tool descriptions), while the “Harness” (the execution layer) handles model invocation, tool call processing, and the control loop — it is the core of agent operation. The two together turn a model into an agent. Using Claude Code and Codex as examples, the article shows how the same model paired with different Harnesses produces radically different experiences, and proposes the common understanding framework of Agent = Model + Harness. Since terminology is not yet standardized, the article aims to provide a practical mental model.