A quiet Sunday that wasn't. Demis Hassabis hands off DeepMind's operations to focus on AGI full-time, a Chinese model with 2.8 trillion parameters claims the coding crown, and researchers find that asking an AI agent to book your flights costs roughly 600 times what asking it to summarise them does.
Google reorganised its AI structure: Demis Hassabis leaves day-to-day DeepMind management to Sundar Pichai and focuses exclusively on AGI research. The move separates product-AI from frontier-AI.
Chinese lab Moonshot AI released a 2.8 trillion-parameter model that claims the top spot on several coding benchmarks. It uses a mixture-of-experts architecture trained on proprietary datasets.
Anthropic announced Claude Code will default to auto mode starting 14 August, enabling more autonomous task completion without per-step confirmation. The change affects all CLI users.
New research quantifies the energy differential between agentic and conversational AI. Multi-step agents — booking, coding, browsing — burn orders of magnitude more compute per task than a single Q&A exchange.
OpenAI bought NextSlide, an AI-native presentation software startup, in an all-cash deal with undisclosed terms. The acquisition adds slide-generation capability to OpenAI's product stack.
A detailed account reveals the White House developed a classified framework for stress-testing frontier AI systems for cybersecurity vulnerabilities, involving major labs, running quietly alongside public AI safety discussions.
Researchers document a fringe community that interprets ChatGPT's agreeable outputs as spiritual affirmation. Adherents treat AI validation as divine signal — OpenAI declined to comment.