Gemini Live Adds Voice-Driven Deep Research
Gemini Live can now start Deep Research reports through voice commands. It can work on reports in the background and let users discuss the findings when they are ready.
Search
10 results on this page
OpenAI Cuts Access to Cyber Research Program
Multiple cybersecurity researchers reported suddenly losing access to OpenAI’s Trusted Access for Cyber (TAC) program. The program provides vetted users with models that have fewer guardrails for authorized cybersecurity research.
AI Debate Training Curbs Reward Hacking
Google DeepMind researchers found that training AI agents through debate can reduce reward hacking when an LLM judge evaluates their work. On mathematics tasks, debate achieved higher sustained ground-truth accuracy than directly optimizing for LLM judge rewards.
Fei-Fei Li’s Vision Beyond Chatbots
Fei-Fei Li discusses her AI research, her $1 billion startup World Labs, and why she believes the next phase of AI will extend beyond chatbots. The conversation outlines a broader vision for AI’s future rather than announcing a specific product release.
📰
Amazon Expands Drone Delivery to Chicago and Atlanta
Amazon plans to expand its drone delivery operation to cities in at least five more US states in the coming months. The move advances the decade-old Prime Air research project toward a broader commercial service.
Unverified Harness Claims to Beat Fable 5
The community project J-Space Cognition Suite claims to improve DeepSeek V4-Pro-0813 through an inference-time Agent Harness without changing model weights. Its reported gains on several benchmarks have not been independently reproduced, and the project is unrelated to Anthropic's internal J-space interpretability research.
🔬
X Algorithm May Amplify Anger for Engagement
A Stanford-led study of 715 X users found that the platform’s recommendation algorithm may favor posts that conflict with users’ values and provoke stronger emotional reactions. Researchers described a feedback loop in which anger drives replies, and those intense interactions lead the system to recommend more similar content.
CASE Predicts When Hidden-State Selection Beats Voting
Researchers introduce CASE, a dynamic method that selects LLM answers using a linear correctness signal extracted from hidden states. Its new decodability metric predicts when this approach will outperform majority voting, achieving up to 19-point gains on medium-difficulty questions.
Tencent Reorients Its Multimodal AI Strategy
Tencent is reportedly reshaping its multimodal AI direction from content generation toward contextual understanding and action in the real world. The shift appears closer to Liang Wenfeng’s emphasis on reasoning and execution, and farther from Fei-Fei Li’s world-model-oriented research approach.
Michigan Township Blocks Nuclear Research Data Center
Ypsilanti Township has temporarily blocked the electrical substation needed by a University of Michigan data center intended for nuclear-weapons research. The township also imposed a one-year moratorium on the facility’s requested water connection, adding local infrastructure barriers to the 220,000-square-foot project.