Google and Meta made it easier to run PyTorch on TPUs

What happened: Google's TorchTPU backend lets PyTorch code run on Google's TPU chips with a small device change.

• Moving code takes one change, from .to("cuda") to .to("tpu"), and it supports eager execution and torch.compile.
• The release covers about 450 ATen operations and spans four GitHub repositories under the Apache 2.0 license.
• Reported figures show 52% model FLOPs utilization training Qwen 3 8B on up to 8,192 TPU v7x chips.

Why it matters: The main gain is lower porting cost if your training and serving code already runs on PyTorch. Run your own model on a TPU VM before you plan around the utilization figure, since it is self-reported.

What happened: DeepSeek published the open source code that runs one of its models, with support for Nvidia and Huawei chips.

• The only supported model is DeepSeek-V4.1-Flash, targeting Nvidia SM100 GPUs and Huawei Ascend 950 NPUs.
• Building it requires Linux, GCC 13.0 or higher, Python 3.12, and CUDA 13 or CANN 9.2.
• A full production service still needs an external KV cache, PD disaggregation, and load balancing. Reported figures include 155 trillion tokens per day per cluster and a 96% input cache hit rate.

Why it matters: You get a permissively licensed reference for serving mixture-of-experts models, which split work across specialist sub-networks, plus a second hardware path beyond Nvidia. The release is a subset of what DeepSeek runs, so benchmark before you assume it matches their hosted API.

What happened: A researcher generated 50,000 unchecked statistics pages on a new domain, and AI chatbots started citing them.

• The domain stateglobe.com was registered in March, and GPT-4.1 wrote the statistics pages in a single day.
• GA4 data shows more than 10,000 ChatGPT sessions to the site, and over 330 domains link to it.
• On October 7, the site ranked second for visibility on a few tracked prompts, behind only Statista.

Why it matters: A citation in a model answer does not confirm a number is accurate. Put source checks into any retrieval pipeline that grounds answers on web pages, before figures reach users.