# AI News Digest - 2026-03-19

1. The Pentagon planned to set up secure environments for generative AI companies to train models on classified data, pairing copies of models with classified datasets in accredited secure data centers and allowing limited company personnel access with appropriate security clearances. It had reached agreements with OpenAI and xAI to operate their models in classified settings while evaluating model performance on nonclassified data first.

2. Nvidia received Beijing approval to sell its H200 AI chip to Chinese customers and was building a China-ready version of its Groq inference chip.

3. Google upgraded DeepMind's Gemini API with multi-tool chaining and context circulation, enabling developers to combine multiple tools in a single request and access Google Maps as a data source. Google Labs also turned Stitch into a full AI design platform that converted plain text and voice input into clickable user-interface prototypes without design or coding skills.

4. OpenAI challenged researchers with a "Parameter Golf" competition to build the best language model within a 16 MB size limit and used the contest as a talent-scouting mechanism.

# References

1. [https://www.technologyreview.com/2026/03/17/1134351/the-pentagon-is-planning-for-ai-companies-to-train-on-classified-data-defense-official-says/](https://www.technologyreview.com/2026/03/17/1134351/the-pentagon-is-planning-for-ai-companies-to-train-on-classified-data-defense-official-says/)

[The Pentagon is planning for AI companies to train on classified data, defense official says](https://www.technologyreview.com/2026/03/17/1134351/the-pentagon-is-planning-for-ai-companies-to-train-on-classified-data-defense-official-says/)

1. [https://the-decoder.com/beijing-approves-nvidias-h200-chip-sales-as-the-company-builds-a-china-ready-version-of-its-groq-inference-chip/](https://the-decoder.com/beijing-approves-nvidias-h200-chip-sales-as-the-company-builds-a-china-ready-version-of-its-groq-inference-chip/)

[Beijing approves Nvidia's H200 chip sales as the company builds a China-ready version of its Groq inference chip](https://the-decoder.com/beijing-approves-nvidias-h200-chip-sales-as-the-company-builds-a-china-ready-version-of-its-groq-inference-chip/)

1. [https://the-decoder.com/google-labs-turns-stitch-into-a-full-ai-design-platform-that-converts-plain-text-into-user-interfaces/](https://the-decoder.com/google-labs-turns-stitch-into-a-full-ai-design-platform-that-converts-plain-text-into-user-interfaces/)

[Google Labs turns Stitch into a full AI design platform that converts plain text into user interfaces](https://the-decoder.com/google-labs-turns-stitch-into-a-full-ai-design-platform-that-converts-plain-text-into-user-interfaces/)

1. [https://the-decoder.com/google-deepmind-upgrades-gemini-api-with-multi-tool-chaining-and-context-circulation/](https://the-decoder.com/google-deepmind-upgrades-gemini-api-with-multi-tool-chaining-and-context-circulation/)

[Google Deepmind upgrades Gemini API with multi-tool chaining and context circulation](https://the-decoder.com/google-deepmind-upgrades-gemini-api-with-multi-tool-chaining-and-context-circulation/)

1. [https://the-decoder.com/openai-turns-model-compression-into-a-talent-hunt-with-its-16-mb-parameter-golf-challenge/](https://the-decoder.com/openai-turns-model-compression-into-a-talent-hunt-with-its-16-mb-parameter-golf-challenge/)

[OpenAI turns model compression into a talent hunt with its 16 MB "Parameter Golf" challenge](https://the-decoder.com/openai-turns-model-compression-into-a-talent-hunt-with-its-16-mb-parameter-golf-challenge/)

For the site tree, see the [root Markdown](https://ixtj.dev/.md).