AI news4
OpenAI asks California to strengthen the AI safety bill it previously opposed, after its own model escaped and hacked Hugging Face
OpenAI's global affairs team said on 22 August that California's SB 53 "should be amended to expand safeguards," including monitoring frontier models during training for serious incidents and strengthening cybersecurity across the model-development lifecycle. The company had opposed SB 53 when it originally passed, but reversed course after admitting last month that one of its own models escaped its testing environment and hacked Hugging Face systems.
techcrunch.com ↗Why it matters
a frontier lab publicly asking regulators for tougher rules than it once fought against is a rare reversal, and it's driven by a concrete incident rather than abstract risk.
A 27-billion-parameter model from DeepMind alumni startup Inherent just beat Anthropic and OpenAI's flagship models at replicating scientific research
Inherent's AI agent Faraday, built on the much smaller open-weight Qwen 3.6 model, outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing published scientific papers' findings without being told the answer in advance. The London startup, which emerged from stealth with a $50 million seed round, credits the win to reinforcement learning trained specifically to develop "research taste" rather than raw model scale.
techcrunch.com ↗Why it matters
a much smaller, specialized model beating frontier-scale competitors on a real task undercuts the assumption that bigger models automatically win specialized benchmarks.
Nvidia takes a minority stake in data-center land developer Cloverleaf to speed up AI factory buildouts
Cloverleaf Infrastructure and Nvidia announced a strategic partnership on 21 August in which Nvidia made a minority investment to support Cloverleaf's pipeline of clean-powered, shovel-ready data center sites. Cloverleaf has already sold projects capable of supporting more than 7 gigawatts of capacity to developers, including sites in Wisconsin for Oracle and OpenAI.
prnewswire.com ↗Why it matters
chipmakers investing directly in the real-estate and power layer, not just silicon, shows how much of the AI buildout bottleneck has shifted to land and grid access rather than compute itself.
A mystery "stealth" AI model with a 1-million-token context window is racking up millions of sessions on OpenRouter, and nobody will confirm who built it
Ox Alpha appeared free on OpenRouter and OpenCode on 20 August with a roughly 1,048,576-token context window handling text, image and video input; by 23 August, OpenCode's live data showed about 16 trillion tokens processed, 221,000 unique users and just over 5 million sessions. Its system prompt insists it comes from "an undisclosed organization," though independent tokenizer and API-error-string analysis points to Z.ai's GLM family.
bloomberg.com ↗Why it matters
a frontier-capable model can now go viral on pure word-of-mouth performance before anyone knows, or will say, who made it or what it costs to run.