Trump advisers tell AI firms they will not safety-test open-weight models, sources say ↗
The Trump administration told leading AI companies that it would not safety-test open-weight models, according to people familiar with the discussions. That leaves a rather consequential portion of risk assessment outside the government framework.
In practice, developers and users may have to shoulder more responsibility for evaluating models before release or deployment. Innovation-friendly, perhaps... but the timing is uneasy, as open models are becoming more capable at a brisk pace. (Reuters)
Open-weight AI models are catching up to the frontier. The safety gap remains. ↗
Z.ai’s GLM-5.2 is reportedly only months behind leading closed models in cyber and biological capabilities. SaferAI’s tests found that it refused none of the offensive cyber or dual-use biology tasks presented to it.
Closed models can rely on refusal training, classifiers and API controls. Downloadable weights are a different creature - once released, safeguards can be altered or removed, making the safety gap feel distinctly canyon-shaped. (TechCrunch)
Samsung Electronics launches next-generation AI memory technology ↗
Samsung unveiled a BV-NAND prototype with more than 400 layers and a new wafer-bonding design. The company says it delivers roughly 58% greater memory density than its previous generation, alongside faster read, write and input-output performance.
Samsung also outlined vertically stacked zHBM memory positioned above AI accelerators. It claims the architecture could offer more than ten times the density of conventional HBM5, triple the energy efficiency and substantially lower thermal resistance - substantial promises, to put it mildly. (Reuters)
AI data-centre race builds $1 trillion lease burden for Big Tech ↗
Microsoft, Meta, Oracle, Amazon and Alphabet have committed around $1.09 trillion to leases that have not yet begun, mostly for AI data centres. Those obligations are nearly four times their recognised lease liabilities.
The capacity could fuel the next cloud boom. Or, less cheerfully, leave companies paying for immense facilities if AI demand disappoints - a trillion-dollar coat purchased before anyone checked the weather. (Reuters)
Anthropic signs $10B deal with AI cloud startup Volta ↗
Anthropic reportedly agreed to purchase $10 billion in computing capacity from AI cloud startup Volta over six years. The deal would support a Norwegian data centre offering 133 megawatts of capacity.
Crypto-mining company Bitdeer is expected to help develop the facility, which will use Nvidia’s Vera Rubin systems. Anthropic is collecting compute partnerships like fridge magnets - except these magnets cost billions. (TechCrunch)
AI offers 'lifeline' for emerging economies, World Bank says ↗
The World Bank says AI could allow developing economies to compress a century of progress into a decade, provided they improve electricity, connectivity and digital skills. Smaller, locally adapted tools could support healthcare, education, agriculture and public services without requiring giant national models.
Its report estimates that 4.5% of jobs in low and middle-income countries are exposed to generative AI, compared with 14.2% in richer economies. Even so, the promise arrives with pointed warnings about inequality, misinformation and political repression - a lifeline with several knots tied into it. (Reuters)
FAQ
Why is the US government not safety-testing open-weight AI models?
According to the article, the Trump administration told leading AI companies that open-weight models would not be included in its government safety-testing framework. This places more responsibility on model developers, distributors and users, who may need to assess cyber, biological, misuse and deployment risks independently before releasing or adopting these systems.
Why are open-weight AI models harder to secure than closed models?
Open-weight AI models can be downloaded, modified and operated outside the original developer’s infrastructure. As a result, safeguards such as refusal training, classifiers and access controls may be weakened or removed. Closed models are usually delivered through controlled APIs, giving providers greater scope to monitor usage, restrict access and update protections after deployment.
How should organisations evaluate open-weight AI models before deployment?
A common approach is to test the model against realistic misuse scenarios, including offensive cyber tasks, dual-use biological requests and attempts to bypass safeguards. Organisations should also examine how the model will be hosted, who will be able to modify it and what monitoring systems are available. Higher-risk deployments typically require stricter access controls and sustained human oversight.
What is Samsung’s new AI memory technology designed to improve?
Samsung’s BV-NAND prototype uses more than 400 layers and a wafer-bonding design intended to increase memory density while improving read, write and input-output performance. The company also described vertically stacked zHBM memory positioned above AI accelerators. These designs are intended to address the capacity, energy and heat demands created by increasingly intensive AI workloads.
Why are Big Tech’s AI data-centre leases considered a financial risk?
Microsoft, Meta, Oracle, Amazon and Alphabet have reportedly committed about $1.09 trillion to leases that have not yet begun, much of it linked to AI data centres. The added capacity could support rapid growth in cloud and AI demand. However, the companies may still owe substantial payments if demand develops more slowly than expected or the infrastructure remains underused.
How could AI benefit emerging economies?
The World Bank argues that AI could help developing economies strengthen healthcare, education, agriculture and public services, particularly through smaller tools adapted to local needs. Those benefits depend on reliable electricity, internet access and digital skills. Governments must also manage risks such as inequality, misinformation and political repression as adoption expands.