UBC, Oxford and DeepMind catalog 26 cases of unpredicted AI behavior

GPT-6 Astra Ultrafast is available now in the OpenAI API and to eligible ChatGPT Work and Codex users, offering up to 8x faster token generation than Astra Standard mode. OpenAI's Philippe Tillet credits NVIDIA tooling for making the models good at programming Blackwell and Rubin GPUs.

GPT-6 Astra Ultrafast is available now in the OpenAI API and to eligible ChatGPT Work and Codex users, offering up to 8x faster token generation than Astra Standard mode. OpenAI used its own models to optimize inference on NVIDIA GPUs.

A ComfyUI user ran the numbers on hours spent generating image-to-video and short videos after switching from rented GPUs to a local RTX 5090. They report their usage habits changed once the hardware was owned, making them more willing to run experimental generations.

SoftBank Group equity investors are looking past higher borrowing costs and credit-risk concerns to focus on potential returns from the company's AI bets, per Bloomberg. The report frames the tension between SoftBank's debt load and its AI upside.

A Reddit user ran family photographs taken 90–110 years ago through ChatGPT to imagine modern hair, makeup, clothing, and body size, including one case with modern military uniforms and equipment.

The social club's grant applications now take 2 hours instead of 3 days, and liquor-license materials 3 hours instead of 4 days. It reports freeing 10-15 hours a week as it opens a new location.

AWS post weighs which parts of a large migration program belong to a managed service versus custom automation, using one enterprise program as the example. The post was reviewed and updated in October 2026.