The brief · 17 June 2026

The AI brief, 2026-06-17

Today, simulated medicine finds its wizard hiding behind the curtain, and ChatGPT learns to remember you without being asked. Between an open model swallowing a million tokens and Ireland picking its own watchdog, here are five stories showing where AI is really being put to the test.

2 min read J / K to navigate

MIRA plays doctor in a simulated patient chart

This AI doesn't just answer questions, it runs an entire consultation solo: interviews the patient, orders tests, reaches a diagnosis, proposes treatment. On simulated cases, it matched or beat human doctors. The open question is how it holds up against a real patient who won't stick to the script.

practitioners › MIRA operates inside a sandboxed electronic health record environment, benchmarked across multiple conditions against human clinicians.

étude MIRA →

JAMA likens medical AI evaluation to the Wizard of Oz

An AI tool can hit 95% accuracy on a benchmark and change nothing about a patient's outcome. The article names the real problem: we measure an algorithm's correctness, almost never its actual effect on a clinical decision. If you use ChatGPT to make sense of a medical report, this is a reminder that a good score doesn't guarantee good use.

practitioners › The article suggests borrowing the Wizard of Oz method from product design: test an AI's perceived impact before it even works properly, judging usage rather than the metric.

JAMA →

ChatGPT can now follow up with you automatically

Set up a task once and ChatGPT remembers to run it for you. Works for one-off reminders or recurring weekly check-ins, with proper task management and clear notifications finally in place.

practitioners › OpenAI rolled out the feature on June 17, 2026, alongside a redesigned scheduled tasks interface and improved notifications.

OpenAI →

Z.ai releases GLM-5.2, an open model that reads a million tokens

A million tokens of context means you can paste an entire codebase or hours of logs into a single conversation without the model losing track. This is built for agents that run long on a task, not for casual chat. And since it's open, you can self-host it if you're working with sensitive code.

practitioners › GLM-5.2, open weights on Hugging Face since June 17, 2026, 1M-token context window, built for agentic workflows and long-context coding.

Z.ai →

Ireland Finally Gets Its AI Watchdog

The Irish government has signed off on legislation transposing the EU AI Act into national law. That means real enforcement teeth for companies deploying AI without following the rules, not just vague principles on paper. If you work with an Ireland-based company, its compliance obligations just became very concrete.

practitioners › The bill establishes the AI Office of Ireland, an independent statutory body responsible for enforcement, coordination, and penalties.

Gouvernement irlandais →