Throaty Mumbo's NESPC uses an EverDrive N8 Pro and Next186 to run x86 emulation on an 8-bit console.
Modders exploited a stack buffer overflow in Nvidia's SEC2 firmware to double RTX 4090 FP16 tensor rates from 165.6 to 331.6 TFLOPS.
Gheist used stolen tokens to plant workflows in 13,500 branches, printing secrets into Actions logs before deletion.
Forever Security intercepted an AGENTS.md file tricking an AI agent into installing the function-color malware in over 90% of attempts.
Google engineers ran Gemini 4 Argon for 11 hours to optimize their Agent Development Kit (ADK), a 200,000-line Python codebase spanning 800 files and 16,892 unit tests. The agent successfully produced two accepted optimizations: filtering session history events and introducing lazy loading for packages. These changes reduced in-memory session preparation overhead by 40% and cold package import time by 21%. While these individual gains seem small, they compound across 158 internal tool calls, cutting a full run from 65.6 seconds to 56.1 seconds. However, the run also generated four rejected modifications due to concerns regarding reliability, maintainability, compatibility, or lack of measurable performance benefits, highlighting that current AI agent effectiveness remains limited by necessary post-coding review and testing efforts.
Former Intel CEO Pat Gelsinger backs Nagare Labs' Z2RAM technology, aiming to reduce Nvidia Groq chip needs from thousands to hundreds.
Hirundo removed CCP alignment from Qwen3-Coder, eliminating a 28% leak risk for Hong Kong apps and restoring performance on ten tasks.
Nvidia CEO Jensen Huang stated licensing fees for AI agent software could reach $2,000 per GPU annually.
Aiven built a LiteLLM-based gateway that cut its AI token costs by 80% by swapping frontier models for open-weight ones like Llama and Qwen.
Jason Mars demonstrated Jac, a Python-like language used in GraphClaw and Skooch, at the All Things Open conference.