## Sources

1. [OpenAI May Sue Apple as ChatGPT-Siri Deal Implodes](https://awesomeagents.ai/news/openai-apple-chatgpt-siri-lawsuit/)
2. [Anthropic and Gates Foundation put $200M on global AI](https://awesomeagents.ai/news/anthropic-gates-foundation-200m-global-ai/)
3. [SU-01](https://awesomeagents.ai/models/su-01/)
4. [Olympiad Gold, Broken Memories, and Attention Loss](https://awesomeagents.ai/science/olympiad-gold-broken-memories-attention-loss/)
5. [Notion 3.5 Turns the Workspace Into an Agent Hub](https://awesomeagents.ai/news/notion-developer-platform-ai-agent-hub/)
6. [HiDream-O1-Image](https://awesomeagents.ai/models/hidream-o1-image/)
7. [How to Use AI for Note-Taking: A Beginner's Guide](https://awesomeagents.ai/guides/how-to-use-ai-for-note-taking/)
8. [US and China Agree to AI Guardrails at Beijing Summit](https://awesomeagents.ai/news/us-china-ai-safety-talks-summit/)
9. [Anthropic Tops OpenAI in Business Adoption for First Time](https://awesomeagents.ai/news/anthropic-tops-openai-business-adoption/)

---

This comprehensive summary outlines the key developments in the AI industry as of May 2026, covering corporate shifts, technological breakthroughs, and geopolitical negotiations based on the provided sources.

### **Anthropic Tops OpenAI in Business Adoption for First Time | Daniel Okafor**

*   **Main Arguments:** For the first time, **Anthropic has overtaken OpenAI in enterprise adoption** among businesses paying for AI services [1, 2]. This shift is attributed to a "sequencing strategy" that successfully captured technical users in specific high-value sectors [3, 4].
*   **Key Takeaways:**
    *   As of May 2026, **34.4% of businesses pay for Anthropic products**, compared to 32.3% for OpenAI [1, 5].
    *   Anthropic’s growth has been explosive, climbing 26 percentage points from just 9% adoption one year prior [2, 6].
    *   Overall AI product adoption across surveyed businesses reached 50.6% [2, 5].
*   **Important Details:**
    *   The data originates from Ramp’s AI Index, which tracks spending across more than **50,000 companies** [6].
    *   Anthropic's success is driven by technical teams in finance and professional services using **Claude Code**, which now authors approximately 4% of all public GitHub commits [3].
    *   OpenAI’s share remained nearly flat, growing only 0.3% over the same 12-month period and dropping nearly 3 points in April 2026 alone [2, 7].
    *   Anthropic is seeking a new funding round at a **$950 billion valuation**, which would surpass OpenAI’s last reported valuation of $854 billion [8].

### **Anthropic and Gates Foundation put $200M on global AI | Elena Marchetti**

*   **Main Arguments:** Anthropic and the Bill & Melinda Gates Foundation have entered a **four-year, $200 million partnership** to deploy AI for global health, education, and agriculture, specifically targeting low- and middle-income countries [9].
*   **Key Takeaways:**
    *   The $200 million commitment includes a combination of **cash grants, Claude API credits, and direct technical support** [10].
    *   The partnership aims to release public benchmarks, datasets, and knowledge graphs to foster an open ecosystem for developers [11].
    *   This move provides Anthropic with a strategic footprint in regions and sectors (like global health ministries) where it currently lacks a commercial presence [12, 13].
*   **Important Details:**
    *   **Global Health:** Focuses on vaccine research for polio, HPV therapies, and better forecasting for malaria and tuberculosis treatments [14, 15].
    *   **Education:** Powers evidence-based tutoring for U.S. K-12 students and literacy programs in sub-Saharan Africa and India via the Global AI for Learning Alliance [16].
    *   **Agriculture:** Provides smallholder farmers with localized guidance on planting and crop disease in their native languages [12, 17].

### **HiDream-O1-Image | James Kowalski**

*   **Main Arguments:** HiDream-O1-Image is a highly efficient **8-billion-parameter open-source text-to-image model** that outperforms the significantly larger 32B FLUX.2 [dev] across major benchmarks [18, 19].
*   **Key Takeaways:**
    *   The model utilizes a novel **Unified Transformer (UiT) architecture** that operates directly in pixel space, removing the VAE (variational autoencoder) bottleneck common in other diffusion models [20, 21].
    *   It is released under an **MIT license**, allowing for unrestricted commercial use [22, 23].
*   **Important Details:**
    *   In internal tests, it scored **0.978 on LongText-ZH** (Chinese text rendering), outperforming FLUX.2 [dev] by 22 points [24, 25].
    *   It features an internal **reasoning-driven prompt agent** that performs a "chain-of-thought" pass before generation to ensure better prompt adherence [26].
    *   Local hosting requires a GPU with at least **24GB of VRAM** for the full model [23].

### **How to Use AI for Note-Taking: A Beginner's Guide | Priya Raghavan**

*   **Main Arguments:** AI is shifting note-taking from a "static archive" to a **searchable personal knowledge base** that users can interact with semantically [27, 28].
*   **Key Takeaways:**
    *   AI tools solve the "retrieval problem" by understanding context rather than just matching literal keywords [28].
    *   The guide recommends five key tools: **Otter.ai** (transcription), **Notion AI** (workspace Q&A), **Mem** (auto-organization), **OneNote + Copilot** (Microsoft ecosystem), and **Obsidian** (local storage for power users) [29].
*   **Important Details:**
    *   **Otter.ai** can join virtual meetings automatically, capturing slides and providing summaries with assigned action items [30, 31].
    *   **Mem** uses "Heads Up" to surface related notes automatically as a user writes, eliminating the need for manual tagging [32, 33].
    *   Users are cautioned against **"letting AI do all the thinking"** and are encouraged to review transcripts for errors in technical terms or names [34, 35].

### **Notion 3.5 Turns the Workspace Into an Agent Hub | Sophie Zhang**

*   **Main Arguments:** Released on May 13, 2026, **Notion 3.5** transforms the workspace into an orchestration layer for AI agents by allowing users to pull in external data and run custom code [36, 37].
*   **Key Takeaways:**
    *   **Workers:** A hosted cloud-based sandbox for running custom "glue code" for workflows [38, 39].
    *   **Database Sync:** Allows Notion tables to sync live data from external services like **Salesforce, Zendesk, and Postgres** [38, 40].
    *   **External Agents API:** Native integration for third-party agents like Claude Code and Cursor to perform tasks directly within Notion [38, 41].
*   **Important Details:**
    *   Workers will be free until August 11, 2026, after which they will operate on a credit system [38, 39].
    *   The platform includes a CLI (Command Line Interface) for developers to manage pipelines programmatically [39, 42].

### **Olympiad Gold, Broken Memories, and Attention Loss | Elena Marchetti**

*   **Main Arguments:** This source examines the contrast between peak AI performance and the subtle ways these systems fail, such as **memory corruption** and **attention loss** in long conversations [43, 44].
*   **Key Takeaways:**
    *   **SU-01**, a 30B model, achieved gold-medal scores on the International Mathematical Olympiad (IMO) 2025 and other academic competitions [45, 46].
    *   Research indicates that **continuous memory consolidation** can actually degrade agent performance; GPT-5.4 failed 54% of previously solved problems after consolidation [45, 47].
    *   A new metric, **Goal Accessibility Ratio (GAR)**, can predict when an LLM is about to lose track of its initial instructions [45, 48].
*   **Important Details:**
    *   Memory "drifts" because models end up **abstracting abstractions**, losing the specific details that made a solution work [49].
    *   The study suggests retaining **raw episodic memories** (original trajectories) often outperforms sophisticated summary-based consolidation [50].

### **OpenAI May Sue Apple as ChatGPT-Siri Deal Implodes | Elena Marchetti**

*   **Main Arguments:** OpenAI is reportedly considering legal action against Apple after the much-hyped ChatGPT-Siri integration failed to generate the projected **billions in subscription revenue** [51, 52].
*   **Key Takeaways:**
    *   OpenAI claims the integration is "buried" within iOS, requiring users to explicitly invoke ChatGPT by name [52, 53].
    *   Internal studies showed users preferred the standalone ChatGPT app over the limited Siri-integrated experience [54].
    *   Apple is reportedly exploring **Google Gemini** as a potential replacement foundation for Siri [55].
*   **Important Details:**
    *   Apple has its own grievances, including data privacy concerns and irritation over OpenAI’s hardware project with Jony Ive [52, 56].
    *   The partnership originally involved no upfront cash, with Apple expecting a revenue cut from new subscribers [57].

### **SU-01 | James Kowalski**

*   **Main Arguments:** SU-01 is a compact **30B Mixture-of-Experts (MoE) reasoning model** that demonstrates gold-medal-level academic performance through a disciplined three-stage post-training recipe [58, 59].
*   **Key Takeaways:**
    *   The model activates only **3B parameters** per forward pass but matches much larger models on math and physics benchmarks [58, 60].
    *   It is fully open-weight under the **Apache 2.0 license** [59, 61].
*   **Important Details:**
    *   **Training Stages:** 1) SFT with reverse-perplexity curriculum; 2) Coarse RL using Group Sequence Policy Optimization (GSPO); 3) Proof-level RL using rewards from DeepSeekMath-V2 [62, 63].
    *   **Test-Time Scaling (TTS):** Uses a produce-verify-revise loop that allows the model to generate reasoning chains exceeding **100K tokens** [63, 64].
    *   Local hosting requires approximately **60GB of VRAM** [65].

### **US and China Agree to AI Guardrails at Beijing Summit | Elena Marchetti**

*   **Main Arguments:** During a summit in Beijing, Donald Trump and Xi Jinping agreed to establish a **formal AI safety channel**, primarily focused on preventing non-state actors from accessing frontier models [66, 67].
*   **Key Takeaways:**
    *   Nvidia CEO **Jensen Huang** joined the delegation as a last-minute addition, highlighting the link between AI safety talks and chip market access [68, 69].
    *   While the U.S. cleared **H200 chip sales** to 10 Chinese firms (including Alibaba and Tencent), no actual deliveries have occurred due to guidance from Beijing [69, 70].
*   **Important Details:**
    *   Experts remain skeptical, noting that China has historically used such meetings to gather intelligence rather than commit to guardrails [71, 72].
    *   The two nations hold **different definitions of "AI safety"**: the U.S. focuses on catastrophic risks (e.g., bioweapons), while China focuses on social stability and content control [73].