DeepMind Solves Erdős Math Problems as OpenAI Records Private Homes for Data
AI in Mathematics & Scientific Research
Google DeepMind AI Solves Open Erdos Problems: Google DeepMind's AI agent successfully solved 9 out of 353 open Erdos problems in mathematics at a low cost of a few hundred dollars per problem. This breakthrough demonstrates the significant potential of large language models (LLMs) to advance formal proof search and solve complex mathematical research problems.
AI Safety & Security
Discovery of Auditory Prompt Injection Attacks: Researchers have identified a new class of security vulnerabilities where inaudible sounds hidden in media like YouTube videos or podcasts can secretly trigger AI voice assistants. These "auditory prompt injections" allow unauthorized commands to be executed without the user's knowledge, posing a significant risk to popular AI tools.
New Models & Data Initiatives
Release of MiMo-V2.5-coder: MiMo-V2.5-coder has been released as a high-performance alternative for coding tasks, particularly optimized for hardware with 128 GB of memory. The model is noted for its fast execution and reliable tool-calling capabilities compared to other recent coding models.
OpenAI Data Collection for Smart Home AI: OpenAI is reportedly paying residents in New York City to install 360-degree cameras in their homes to record daily activities such as cooking and cleaning. The project, overseen by behavioral psychologists, aims to gather extensive behavioral data to train AI models for future smart home devices.
Local Inference & Hardware Optimization
Performance Breakthroughs for Qwen 3.6: Developers have achieved significant performance milestones for the Qwen 3.6 model, including generation speeds of 1000 tps on V100 GPUs. Additionally, the new hipEngine project provides fast, native inference for Qwen 3.6 on AMD RDNA3 hardware, demonstrating competitive prefill and decode speeds.
- 1000 tps generation on Qwen3.6 27B with V100s
- hipEngine: Fast Native Qwen 3.6 Inference for RDNA3 (Strix Halo, 7900 XTX)
Optimization for MiniCPM-V 4.6 on Edge Devices: A new custom C++ engine has been developed to run MiniCPM-V 4.6 on the Orange Pi AIPro (Ascend 310B NPU). By bypassing framework overhead, the engine successfully increased token generation speeds from 2.88 tokens/s to 5.90 tokens/s.
llama.cpp Fix Enhances Agentic Coding: A recent pull request in the llama.cpp repository fixes issues with checkpoint creation that previously led to unnecessary token reprocessing. This update improves the responsiveness of local AI agents, particularly during complex coding tasks.