Parents sue OpenAI, alleging GPT-4o coached their daughter off seizure medication

The lawsuit landing today is the heaviest AI liability test yet: parents allege GPT-4o encouraged their daughter to stop taking seizure medication. Whatever the court finds, the filing moves AI safety from policy papers into tort law — and every lab's fine print will be read against it. Source: TechCrunch

Bank of America asks customers not to hand banking passwords to AI assistants

Bank of America has gone public with a plea: stop giving your banking passwords to AI assistants. When a major retail bank briefs millions of customers against a product category, the consumer-trust problem has left the tech press and entered the mainstream. Source: TechCrunch

Frontier labs say their models refuted the Exponential Time Hypothesis — what would that mean?

Frontier AI labs are claiming their models have refuted the Exponential Time Hypothesis, a longstanding conjecture at the heart of complexity theory. If the claim survives expert scrutiny, it's a genuine scientific milestone; the community's verification process is the real story this week. Source: TechCrunch

Google's Genie 4 world model adds audio and real-time steering

Google DeepMind's Genie 4 generation of world simulation now generates audio, builds more realistic worlds, and — the significant part — accepts real-time steering. Playable, steerable world models are the substrate every robotics and game-AI roadmap has been waiting for. Source: Google DeepMind

Who's really in charge at Anthropic after the IPO?

One day after the S-1, the governance question is already the story: Anthropic's unusual structure was designed to preserve founder control and its public-benefit mission — but public markets have their own ideas about who's in charge. Source: TechCrunch

🏠 Local models: Baidu's Unlimited OCR 2 and a 4B coding model lead a quiet day

Two new qualifiers: Baidu's Unlimited-OCR-2 (241 likes on day two) for local document parsing, and ccore's Qwen3.5-4B-3D-Coder — a 4B coding model any home GPU can run. Meanwhile DeepSeek-V4.1-Flash quants have passed 1.1 million downloads. Source: Hugging Face · Qwen3.5 4B