Hi Synapticians!

Today, we've got quite a few interesting topics:

• NVIDIA is announcing massive investments in the US (in exchange for exemptions on exports to China),

• Palantir (a very discreet player, but huge) has been selected by NATO to automate battlefield image analysis,

• Apple claims it can now train models without using our personal data (which is great—now that they’ve solved that, maybe they can finally start doing AI),

• And Pokémon is causing chaos in model benchmarking.

But above all, we encourage you to read the last article (right after today’s splendid meme), which is dedicated to education and a French startup (cocorico!) that lets you create a didactic training course on any topic in just a few minutes. It’s really cool.

Your turn!

Top AI news

1. Nvidia shifts AI chip production to US amid trade tensions
Nvidia will manufacture its AI supercomputers entirely in the US, investing up to $500 billion in new facilities in Arizona and Texas. The move follows US tariff threats and political pressure, including a reported meeting between CEO Jensen Huang and Donald Trump. While production of Blackwell chips has begun, challenges remain: labor shortages and geopolitical risks could hinder execution. The shift reflects a broader trend of tech reshoring in response to national policy.

2. NATO deploys Palantir AI to automate battlefield intelligence
NATO has chosen Palantir’s MSS Nato AI platform to automate the analysis of battlefield data, replacing hundreds of human analysts. The system, derived from the U.S. military’s Project Maven, will be operational within 30 days. It integrates multiple data sources to support faster, more informed military decisions. Meanwhile, Palantir’s Gotham software is being considered in Germany for police use, raising legal and ethical concerns. The move highlights the growing role of AI in both defense and public safety, and the complex questions it raises about oversight and accountability.

3. Apple trains AI with synthetic data, not your emails
Apple is enhancing its AI features like email summarisation using synthetic data and differential privacy. Instead of collecting real user content, Apple generates fake emails and compares them locally on devices. Only aggregated, anonymised data is sent back. This method protects user privacy while improving AI performance. The approach is already used in Genmoji and will expand to other tools. It’s currently in beta on iOS 18.5, iPadOS 18.5, and macOS 15.5.

Bonus. Pokémon reveals the flaws in AI benchmarking practices
A viral post claimed Google’s Gemini AI outperformed Anthropic’s Claude in Pokémon, but it had help: a custom minimap. This case highlights how AI benchmarks can be skewed by implementation tricks. From Claude’s coding scaffold to Meta’s fine-tuned Llama, companies are optimizing for tests, not real-world use. The article calls for more transparency and better evaluation standards to ensure fair comparisons.

Meme of the Day

Theme of the Week

AI for Education - The Startup

Bubble Teach turns static training documents into dynamic micro-learning experiences using AI. The platform helps companies deliver personalized training, adapted to learner profiles. It’s fast, intuitive, and integrates with existing LMS tools via SCORM. Managers can track progress in real-time, and learners get bite-sized, relevant content. With a flexible pricing model, Bubble Teach makes AI-powered learning accessible to all.

Stay Connected

Feel free to contact us with any feedback or suggestions—we’d love to hear from you !