Seventy: Wednesday, 21 October 2026

Wednesday, 21 October 202610 stories made this issue.

What happened. An independent Schlotzsky's and Cinnabon franchisee says ChatGPT Work shortened menu repricing by 95% and cash reconciliation by 94%. The franchisee also set up a process that could release $100K in cash.

What happened. A European Commission study under the Digital Services Act catalogues common online money scams, the platform features fraudsters exploit, the harm to users and society, and steps platforms can take to block them. It notes these scams appear on all kinds of online services, including the largest platforms and search engines.

What happened. The Barcelona Supercomputing Center's AI Factory Accelerator Programme is now fully running after a pilot, giving startups sovereign computing power, technical support and training. Some of its services are now available through the NVIDIA Inception portal.

What happened. AWS released TorchNeuron in beta, free open source software that lets PyTorch code run directly on its Trainium AI chips. Developers point existing code at the chip by changing one line, and standard PyTorch features keep working.

What happened. A Microsoft executive visited Saint-Gobain in Paris to see how the building-materials maker uses Microsoft Discovery to speed up work on more energy-efficient construction products.

What happened. OpenAI is buying Orchestra Bio, a company that applies AI to the day-to-day running of research and development in the life sciences. The aim is to help drug makers move treatments through development faster.

What happened. i14 built Laplace 3B, a language model whose memory per conversation stays fixed at 29.5 MB however long the context gets. With AWS, it trained and served the model on Trainium2 chips using native PyTorch.

What happened. Amazon Bedrock customers in Canada's Central and Calgary AWS Regions can now call OpenAI's GPT-5.6, GPT-6 and GPT-6.1 Sol through cross-Region inference. Requests are processed within North America.

What happened. Amazon Bedrock now lets all Anthropic Claude models return answers in a fixed JSON structure you define. It works through both the bedrock-runtime and bedrock-mantle endpoints, with no custom parsing or retry logic.

That's the whole issue.

10 stories, about 3 min 32 sec. Nothing was left out to fill the time.

Words of the day

Accelerator
A chip built specifically for the mathematics AI needs, rather than the general-purpose work a CPU does. GPUs are the common kind; TPUs are Google's.
eager execution
A mode where each operation runs immediately as written, rather than being built into a graph first, which makes debugging easier.
torch.compile
A PyTorch feature that compiles model code into optimised kernels so it runs faster than ordinary step-by-step execution.
For today

“In the race for excellence there is no finish line.”

— Sheikh Mohammed bin Rashid Al Maktoum