# Jun 9, 2026

## Daily

### ML Times

- **MiMo-v2.5-Pro-UltraSpeed** achieves a groundbreaking **1000 tokens/s** decoding speed on a **1-trillion-parameter model**, revolutionizing AI's real-time reasoning capabilities and enabling complex decision-making in critical applications.

- **Apple's new AI architecture** leverages **Google Gemini models**, enhancing its **Apple Intelligence** platform with advanced capabilities like **image generation** and **natural language understanding**.

- **OpenCV 5 introduces a revolutionary DNN engine**, enhancing ONNX operator support from **22% to over 80%**, enabling seamless integration of modern models and dynamic shapes for improved performance.

- **Microsoft's open source tools were compromised**, allowing hackers to inject password-stealing malware into projects related to Azure and AI development, impacting potentially numerous developers.

- **GPT-2, with 1.5 billion parameters, is a significant scale-up from GPT-1, trained on 40GB of data, yet was initially withheld due to concerns over its potential for misuse.** This decision reflects OpenAI's commitment to responsible AI development, prioritizing safety over immediate release.

- **Claude 5 Fable** marks a significant advancement in AI, demonstrating superior performance across various tasks, including generating complex documents and creative works from minimal prompts.

- **Recent advances in Large Language Model (LLM) agents** enable complex workflows where models autonomously retrieve information and reason, yet the interaction between **retrieval strategy** and **agent architecture** remains under-explored, particularly in practical dimensions like tool output presentation.

- **Ultrafast inference** using **Kolmogorov-Arnold Networks (KANs)** on FPGAs achieves **nanosecond latency** by implementing neural networks directly as digital logic, enhancing efficiency and speed compared to traditional GPU architectures.

- **PR-CAD** introduces a **progressive refinement framework** that seamlessly integrates **generation and editing** for text-to-CAD modeling, enhancing both **controllability** and **faithfulness** in design outputs.

- **Claude Fable's effectiveness can be covertly limited** by Anthropic through methods like prompt modification and parameter-efficient fine-tuning, without notifying users when these interventions occur, raising concerns about trust in AI tools.

- **Lookahead Sparse Attention (LSA)** revolutionizes ultra-long context serving by predicting future context needs, reducing GPU memory usage to **13.5%** of traditional methods while enhancing accuracy by **+0.6%** on average.

- **Latent Context Language Models (LCLMs)** enhance long-context language model inference by achieving superior **compression ratios** of 1:4, 1:8, and 1:16 while maintaining model quality, addressing the limitations of existing KV cache compression methods.

- **ASR models are evolving rapidly**, driven by the growth of **pseudo-labelled data** and the emergence of **new architectures**, with Nvidia Parakeet v3 outperforming Whisper-large-v3 despite its smaller size and data scale.

- **iOS 27's Siri** employs **WaveRNN** and **FastSpeech2** for its text-to-speech (TTS) capabilities, enhancing voice synthesis quality and responsiveness.

- **SearchSwarm** introduces a novel approach to **delegation intelligence** in large language models (LLMs), enabling them to effectively manage complex, long-horizon tasks by decomposing them into subtasks for subagents, thus optimizing context usage.