Elon Musk unveiled Terafab, a new facility aiming to produce 50x the world's current AI compute per year. The facility will integrate logic, memory, packaging, and testing for chips used in Tesla vehicles, Optimus robots, and space-grade AI satellites, scaling AI chip production significantly.
Sources: 2 Covered by The Neuron, The Rundown AI
The NousResearch Hermes Agent autonomously authored, typeset, edited, and published a 79,000-word novel titled 'The Second Son of the House of Bells,' utilizing an autoresearch-style loop for fiction generation, world-building, and adversarial editing. This demonstrates a significant leap in AI's creative and autonomous content generation capabilities, particularly in long-form writing.
Sources: 2 Covered by The Neuron, The Rundown AI
OpenAI has announced its goal to develop a fully automated, agent-based AI researcher capable of solving complex problems independently. They plan to release an autonomous AI intern in September and a multi-agent system by 2028, making this initiative their primary research focus.
Sources: via TLDR AI
The UK government's AI Security Institute found that successive generations of frontier AI systems significantly outperform predecessors in multi-step cyberattacks. The research indicates AI systems are nearing autonomous, end-to-end cyberattack capabilities, which could lower the cost and increase the scale of attacks.
Sources: via Import AI
Researchers from Chinese institutions have developed MERLIN, a multimodal AI model, along with the EM-100K dataset and EM-Bench benchmark, specifically designed for electronic warfare tasks. MERLIN significantly outperforms frontier LLMs like GPT-5 and Gemini-2.5-Pro in perceiving and reasoning about electromagnetic signals.
Sources: via Import AI
Garry Tan, CEO of Y Combinator, has released gstack, an open-source tool that leverages AI agents (like Claude Code) to emulate a full engineering team, enabling individual developers to ship massive amounts of code. This initiative highlights a shift towards highly automated software development using AI.
Sources: via Keep Bookmarks
Researchers at Supermemory have developed an experimental agentic flow, ASMR (Agentic Search and Memory Retrieval), which achieved approximately 99% accuracy on the LongMemEval-s benchmark. This breakthrough, which replaces traditional vector search with parallel agent orchestration, suggests a significant advancement in solving long-term AI memory challenges.
Sources: via Keep Bookmarks
A new arXiv paper demonstrates that the key-value (KV) cache in transformer inference is entirely redundant, as keys and values can be recomputed from the residual stream with zero error. This leads to KV-Direct, a bounded-memory inference scheme that significantly reduces memory usage while maintaining 100% token match compared to existing eviction baselines.
Sources: via arXiv AI
Researchers have developed a new framework leveraging subgoal decomposition for online planning and a reinforcement learning method called MiRA to enhance LLM agents' ability to handle complex, long-horizon tasks. This approach boosts proprietary models like Gemini by 10% in success rate and increases Gemma3-12B's success rate from 6.4% to 43.0%, surpassing GPT-4.
Sources: via arXiv AI
New Zealand's Halter, which uses AI-powered 'Cowgorithm' collars for remote cattle management, is reportedly close to a new funding round led by Peter Thiel's Founder's Fund, valuing the startup at $2 billion. The technology modernizes agriculture by replacing physical fences and manual checks with real-time data and remote control.
Sources: via The Rundown AI