Created automatically with AI: the texts and images in this edition are AI-generated, based on the linked original sources. Errors may occur.
Today: new AI models for language and research, tools for 3D reconstruction and programming, and privacy and greater control over AI systems.
AI-generated image
01heise online
Google unveils Gemini 4 Argon – initially for selected testers only
Gemini 4 Argon is Google’s new flagship model and will initially be tested by selected cyber defense teams and internal groups. According to heise, it processes text, images, video, and speech, but outputs only text. Its output limit is one million tokens. In a test by Artificial Analysis, Argon scored 53 points, tying with GPT-6 Astra; according to the report, it lagged behind that model and Google’s predecessor Gemini 3.1 Pro in accuracy.
Why it matters
The staged access shows that Google wants the model tested before a broader release. The reported test results vary depending on the metric.
Microsoft adds three audio models to its MAI model family
Microsoft has introduced MAI-Transcribe-2-Streaming and the speech models MAI-Voice-2.1 and MAI-Voice-2.1-Flash. According to Microsoft, the transcription model is designed to recognize 60 languages continuously; the two speech models reportedly support 23 languages and are intended for different use cases, such as real-time agents or longer audio content. The models are available through Microsoft Foundry and other platforms. To clone a voice, Microsoft requires developers to undergo verification and obtain the consent of the person concerned.
Why it matters
With these models, Microsoft is expanding its own building blocks for voice-based AI agents. The consent requirement applies to the use of a particularly sensitive feature: reproducing human voices.
Ai2 releases the open research-report model AstaBrief
The Allen Institute for AI (Ai2) is making AstaBrief 8B and its training data openly available. Given a research question and excerpts from retrieved literature, the model produces a report with citations. According to Ai2, report generation in the Asta system takes an average of 51.1 seconds in fast mode, compared with 178.5 seconds in the comparison mode using Claude. The organization notes that the comparison data comes from an earlier model generation and has not been retested against today’s leading models.
Why it matters
Open model weights and training data enable research institutions to study the approach and run it locally. The limitation in the comparative tests is important for putting the speed claim in context.
BootLoops supports AI with scientific calculations – experts remain essential
Harvard physicist Matthew Schwartz describes the open-source tool BootLoops, which uses Claude for exact scientific calculations. According to his report, 36 manuscripts across 18 fields were produced in three months. The system calculated integrals and worked on an equation from ecology, among other tasks; however, experts had to help shape the questions and interpretations for the results to be scientifically relevant. Schwartz cites premature claims of success, faulty conclusions, and high computational costs among its weaknesses.
Why it matters
The report shows a potential application for AI in research, but also the limits of automated results. Expert assessment remains central, even when calculations are performed correctly.
LEGO-Anything turns individual photos into editable 3D scenes
The LEGO-Anything research project has a programming agent generate code for a scene in Blender from a single photo, then iteratively revise it. The accompanying LEGO-Bench test contains 208 images. In the benchmark, GPT-6 Astra scored 53.4 percent on indoor scenes and 39.6 percent on outdoor images, according to the report. The tested agents could hardly assess their own geometric accuracy reliably. An accompanying plugin with specific measurements improved the results of all six models tested.
Why it matters
Executable scene code can be inspected and edited, but a working result is not necessarily an accurate reconstruction. The benchmark reveals shortcomings in spatial accuracy and self-assessment.
Claude Code gets a Mods system for custom extensions
Anthropic is adding Mods to Claude Code: JavaScript or TypeScript functions that hook into events in the tool. Developers can use them to customize the interface, intercept tool calls, or add commands, among other things. The extensions work in the command line and desktop app, and partly in the VS Code extension. Anthropic warns that Mods run with the user’s permissions and are not sandboxed. Organizations can specify which extensions are allowed to load.
Why it matters
The system gives developers more ways to adapt the AI coding assistant to their own workflows. Because Mods do not run in isolation, their provenance is also a relevant security consideration.
Apple announces tighter controls for broad file access on Mac
Apple plans to strengthen security for the macOS “Full Disk Access” permission. It can allow apps to access files, emails, messages, and browsing history. The company cites the growing risks posed by more capable and autonomous AI agents as the reason for the planned controls. According to Apple’s announcement, anyone wishing to grant an app this broad access will in future have to confirm it through a particularly explicit action. The report gives no specific timeline for the changes.
Why it matters
The change concerns a central trust question for AI agents on desktop computers: which personal data are they allowed to see? Apple wants to make consent to broad access more explicit.
Eclipse Foundation launches initiative for greater choice in AI systems
The Eclipse Foundation has launched the Sovereign AI Foundation as a vendor-neutral initiative. It is intended to help member organizations identify their dependencies on AI providers and assess open-source alternatives. Practical resources such as guides and technical blueprints are planned; the initiative will not develop its own software. According to Eclipse, 17 organizations from business, research, and the open-source community are participating at launch. “Sovereignty” does not mean giving up proprietary software, but rather gaining greater understanding, adaptability, and the ability to switch.
Why it matters
The initiative focuses on control and the ability to switch AI infrastructure, rather than on any one model. This can help organizations evaluate different providers and open tools.
Automatically summarized with AI and checked against the source texts retrieved. Original sources are linked for every story; errors may still occur. An editorial overview from Medienfuchs.
