LLMTraceFX ships a deterministic KV-cache auditor in merged PR #74 that independently verifies cache attestations and flags evictions, confirming token-identical cached outputs across 10 synthetic tests while showing an exact duplicate reuses 8 of 9 tokens.
Google DeepMind launches the mujoco-menagerie Python package for on-demand curated robot simulation models, letting users load Unitree Go2 with mm.load('unitree_go2') and pin models to the 2026.9.0 release.
OpenAI urges developers building for GPT-6 Astra to strip down prompts, AGENTS.md files, and skill definitions, warning that bloated context and excessive constraints can hinder the model even as it automatically runs tests and checks its work.
LibreChat v0.8.8-rc3 launches a beta Agent Management API and GPT-6 Astra support, while introducing experimental agent workspaces that can edit files and run Bash commands under bounded timeouts.
Anthropic CEO Dario Amodei calls for a globally coordinated AI “pacing” regime that gives independent evaluators such as METR employee-like access to frontier models, training pipelines and internal safety processes.
SoftBank secures an upsized $11.87 billion two-year loan from roughly 20 banks for its OpenAI investment, while its shares plunge as much as 13% amid plans to repay $25.9 billion in September and potentially sell up to $20 billion in U.S. junk bonds.
Index Ventures says FOMO-driven AI investing can be viable as startups across models, chips, healthcare, legal and defense command $1 billion-plus valuations, provided founders are solving real customer problems.
Y Combinator CEO Garry Tan urges regulators to leave AI-model distillation alone, arguing that frontier labs can retain a price premium even as open-weight alternatives proliferate amid Anthropic and OpenAI accusations against Chinese AI firms including DeepSeek.
Apple unveils the iPhone Duo, its first foldable iPhone, using its larger display to support AI agents and multitasking while also previewing iPhone 18 Pro camera upgrades, the A20 Pro chip and on-device AI models.
Anthropic adopts Dario Amodei’s plan to pace frontier AI development with embedded third-party safety evaluators and coordinated standards, while OpenAI considers slowing collaboration with rival labs and has paused some safety runs to strengthen security.