A fully open model now walks through multi-step problems out loud. The capability that was supposed to be the closed labs' last fortress just opened.
Open weights just learned to reason, and the last fortress the closed labs were counting on — the ability to think through multi-step problems out loud — just opened its gate. A fully open model now sketches its own working, checks it, and corrects course without a single proprietary API in the loop.
The moat everyone said would hold longest was never the reasoning itself. It was the compute to train it in secret. Once the recipe is on GitHub, the fortress was always going to be a tent.
A graphic design award shortlisted a campaign that was entirely AI-generated. The judges didn't know. The studio didn't disclose it. The debate over whether that matters is the most honest conversation the industry has had in years.
A London hospital deployed a chest X-ray model that caught 23% more abnormalities than its senior radiologists. Then a patient sued both the machine and the doctor who overruled it.
Enterprise rollouts of coding and operations agents tripled this year. The awkward secret: no manager has figured out how to performance-review a language model.