An expert witness in a $61 million lawsuit over an explosion that killed three people used ChatGPT — with prompts literally instructing it to show a company was "0% at fault." This is the AI-in-legal-proceedings scandal that was always coming, and it's worse than most expected.
After three years of promises, the Mojo programming language hit 1.0 and dropped its source code publicly. Built for AI systems programming, it's one of the most anticipated open-source releases in the ML infrastructure space.
Alibaba's 27B parameter model scores the same as OpenAI's GPT-5.6 Luna on the Artificial Analysis Intelligence Index — while being small enough to run on a laptop. The gap between frontier and local models keeps narrowing fast.
404 Media followed physical book shipments from dealers receiving suspiciously large, price-insensitive orders — and traced them to an Amazon AI training operation. The hunger for high-quality human text is now stripping physical shelves.
Anthropic's text watermarking approach treats words as interchangeable tokens — exposing that AI labs have no coherent framework for evaluating writing quality. It's a philosophical failure dressed up as a technical feature.
Robot maker Unitree's public debut blew past crypto prediction markets that were already pricing in a strong open. Physical AI hardware is commanding multiples that make crypto speculation look conservative.
Ethan Mollick flagged that OpenAI is committing 20% of research inference compute to chain-of-thought monitoring for alignment — and his read is blunt: if the problem warrants that scale of resource, alignment issues are more serious than the company is publicly admitting. When a lab allocates that much compute to watching what its model is thinking, the subtext writes itself.
Three researchers independently cracked a 40-year-old math problem in the same week, all with heavy ChatGPT assistance. This isn't a coincidence story — it's a signal that AI is collapsing the time between "hard open problem" and "solved," and the simultaneous convergence suggests the tool is guiding people toward the same solution paths.
Five weeks after the OpenAI/Hugging Face breach, a16z surfaced the core uncomfortable truth: defenders now have to ask AI models the same questions attackers do, just to understand what's possible. Traditional security assumed human adversaries. That assumption is gone, and most enterprise security stacks haven't caught up.
Today's stories form a single uncomfortable arc: AI is being dropped into high-stakes situations — courtrooms, security infrastructure, mathematical research — faster than any guardrails are being designed for it. The expert witness using ChatGPT to argue a company bears zero fault in deaths isn't an outlier. It's a preview of standard practice if nothing structurally changes.