OpenAI’s AGI number came from a harness, not the model
THE SO WHAT
Benchmark harness choices can swing an “AGI-level” score from 62.7% to 99.9% — that turns evals into a product surface, not an objective truth. If you’re using third-party benchmarks to make spend or safety decisions, demand raw runs, harness details, and independent replication before you move budgets.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIFlock's AI search tool for police officers has been reverse engineered — here's what it shows
Once an AI policing tool is reverse engineered, its decision logic becomes a governance and liability artifact, not a black box. If you’re selling or buying AI into law enforcement, assume internal scoring and ranking rules will eventually be scrutinized in court and the press.
Applied AII’m a father of three who studies the impact of artificial intelligence: this is what parents need to know about AI
Households are becoming de facto AI deployment environments, with parents forced into the role of safety officer and curriculum designer. If you build consumer AI, assume your real buyer is a time-poor parent managing tradeoffs between convenience, learning, and dependency — design controls and defaults accordingly.
Applied AIMy Brief Summer Fling With Siri AI
The novelty curve on assistants is steep — if daily workflows don’t rewire in the first weeks, the product becomes invisible again. For anyone shipping an AI assistant, the bar is not “wow” at launch but sustained, repeated use on 2–3 core jobs that users actually remember to invoke.
Applied AIHow AI child sexual abuse material is colliding with free speech protections
AI-generated CSAM forcing First Amendment tests is a legal red line—once courts start carving out new categories of unprotected speech, generative tools move into a far more regulated zone. Any team shipping open-ended generation should assume stricter liability and provenance expectations are coming, not just better filters.