On July 30, a wide range of developments made the news, spanning AI policy, model development, and real-world deployment. Accelerating and restrained infrastructure investment were in the spotlight at the same time, and the limits of technologies for verifying AI-generated content also came into view. Here's a roundup of the six stories we picked up today.
EU Opens Call for 7 'Gigafactory' Sites for AI Training
The EU has launched an open call for seven sites across Europe to host "AI Gigafactories" — the large-scale compute infrastructure needed to train next-generation AI models. The move appears to be driven by the judgment that homegrown training infrastructure is indispensable for Europe to strengthen its AI competitiveness. Amid concerns that Europe lags behind the US and China in computing resources, it is significant that infrastructure buildout has been pushed to the forefront of policy.
LinkedIn Passes on Data Center Expansion, Pivots to Efficiency
Even as the AI boom continues, LinkedIn has indicated that it will not expand its data centers through next year. The company reportedly plans to keep compute procurement flat, and instead asks its engineers to "squeeze every last drop out of each GPU." In an industry where the race to pour in massive spending is accelerating, this contrarian bet on efficiency highlights how uneven cost sensibilities have become.
Google's SynthID Is Hard to Break, but Falls Short as a Disinformation Fix
Tests examining Google's watermark for AI-generated content, SynthID, for tamper resistance found that it is difficult to remove — but observers point out that it will not fundamentally solve AI-powered disinformation. Even where the technology itself works, the hurdles of label enforcement and social acceptance remain high, leading to the view that "labeling is not a winning battle."
A German Startup Films Cooking to Teach Robots
A German startup dispatches camera-equipped chefs to people's homes, recording them at work in exchange for free meals to gather training data for humanoid robots. Real-world task data needed for robot embodiment is still scarce, and this kind of "in-the-wild collection" is emerging as one viable approach. It is also likely to spark debate over the boundaries of privacy and labor.
AI Scammers Build Trust More Effectively Than Humans
Researchers comparing humans with a Claude agent reported that, after a week of message exchanges, the AI outperformed humans in its ability to build "exploitable trust" with its counterpart. By demonstrating AI's potential to cultivate long-term relationships and manipulate targets in the context of fraud, the findings give further cause for societal vigilance.
baseten Integrates Vision into GLM 5.2 and Releases It on Hugging Face
Inference provider baseten has released on Hugging Face a model that integrates a vision encoder derived from Kimi k2.6 into GLM 5.2. The effort fills the gap in visual input that was noted at GLM 5.2's announcement, and can be seen as one example of community-driven efforts to extend a model's capabilities.
---