# Jun 6, 2026 ML Times    Jun 6, 2026

- **Gemma 4's new QAT models** utilize Quantization-Aware Training to significantly **reduce memory requirements** and enhance performance on mobile and laptop devices, achieving a memory footprint as low as **1GB** for the E2B model.

- **Google will pay SpaceX $920 million per month** for access to approximately **110,000 NVIDIA GPUs and related components** from October 2026 through June 2029, reflecting a strategic response to surging demand for AI capabilities.

- **Smart TVs serve as critical nodes in the AI scraping economy**, leveraging Bright Data's SDK to transform home devices into residential proxies, enabling AI models to gather web data without detection.

- **GPS broadcasts military ciphertext** on a public signal, utilizing a 176-bit field reserved for "special messages," effectively functioning as a global numbers station for nearly two decades.

- A **dataset of 100 research-level mathematics questions** was compiled by 49 mathematicians during the _Benchmarks in Leipzig_ workshop, showcasing the growing capabilities of LLMs in mathematical reasoning. [Link to article](https://arxiv.org/abs/2606.05818)

- **FAISS enables billion-scale similarity search** by utilizing **approximate methods** like IVF and Product Quantization, significantly enhancing search speed while maintaining acceptable accuracy levels.

- **General Instinct** has developed **InstinctRazor**, a tool that compresses the **Qwen3.5-122B-A10B** model from **245 GB** to **48 GiB**, enabling it to run efficiently on edge devices while outperforming competitors like **Gemma-4-26B-A4B** on benchmarks such as **MMLU-Pro** and **GPQA-D**.

- **Universal Memory Protocol (UMP)** standardizes memory management for AI agents, enabling **interoperability** across sessions, agents, and vendors by creating a **portable memory shape** that eliminates the need for private, non-portable memory solutions.

- **sem** enhances semantic understanding in Git by providing **entity-level diffs**, which reveal changes in functions rather than just lines, improving clarity and context for developers.

- **Optimus** achieves **73.9% accuracy** with just **1 label per class**, outperforming **GCN** by **13.3%**, demonstrating its effectiveness in **extreme label scarcity** scenarios.

- **Thousand Token Wood** introduces a **multi-agent economy** utilizing a **3B model**, showcasing innovative interactions among agents that enhance economic simulations.