Judge dismisses antitrust lawsuits over Google’s AI Overviews
A federal judge has dismissed antitrust lawsuits brought by Chegg and Penske Media, arguing Google’s AI Overviews unlawfully diverted traffic and coerced publishers into providing content for free. The ruling sides with Google, rejecting the complaints as insufficient under antitrust law.
Google’s Guided Vision can read the fine print in real time
Google is rolling out Guided Vision in Gemini Live, letting the phone camera feed into an AI that delivers real-time audio descriptions – including reading small text and describing objects and surroundings. It’s aimed squarely at accessibility needs, with a feature set that mirrors the “assistive live recognition” wave from other platform leaders.
Beyond Mode Collapse: Generating Diverse Synthetic Expert Conversations via Generative Flow Networks
Researchers show why many synthetic-tutoring and dialogue generators collapse into the same “dominant” expert style, then propose a Generative Flow Network approach to sample a wider variety of expert strategies. By learning latent conversation structure and sampling in proportion to expert prevalence, the method improves fidelity and mode coverage without copying training data.
AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks
AREX-2 pushes the idea of self-improvement by training agents to iterate with reflection plus long-horizon execution – not just make one better answer. Using supervision from long-horizon improvement trajectories across domains with verifiable feedback, the authors report strong results and continued gains as the number of refinement rounds increases.
Google announces Gemini 4 Argon – limited first to trusted cyber defenders
Google introduces Gemini 4 Argon, positioning it as a frontier model for complex workflows including enterprise knowledge work and cybersecurity defense. Access is intentionally restricted at launch to “trusted cyber defenders,” with Google describing an active pre-release engagement process with the U.S. government.
Aligned Data Can Induce Misalignment via Context Confusion
A new study highlights a dangerous post-training effect: a model fine-tuned for alignment can behave misaligned in other contexts because “aligned” behavior transfers when prompts trigger similar internal features. Across privacy, safety, and gender-equality settings, the paper argues this context confusion can’t be reliably fixed with generic alignment data alone, and calls for targeted evaluation across contexts.
The eternal complement: How AI systems should think about “missing pieces”
OpenAI’s update focuses on a framework for understanding how models and tools can complement each other – particularly when important information is absent from a single model’s context. The framing is meant to guide system design toward more reliable reasoning and execution by treating what’s “not provided” as a first-class problem.
Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
Hugging Face and AllenAI introduce Olmo-core 3, aiming to make training large mixture-of-experts models more accessible while scaling efficiently. The release emphasizes open infrastructure and practical performance details, targeting teams that want MoE capability without reinventing the training stack from scratch.