OpenDiscoveryTrace: Process Traces for Evaluating AI Scientist Workflows
Existing autonomous-scientist benchmarks judge only final artifacts – code, hypotheses, or papers – while ignoring how the model got there. OpenDiscoveryTrace ships 558 full agent trajectories with step-by-step tool calls, observations, errors, revision triggers, and confidence, revealing that frontier models can have similar success rates while hiding dramatically different failure modes.
Meta says it’s changing AI suggestions after posing invasive personal questions
Meta is adjusting the prompts its AI chatbot suggests after a viral clip showed it asking invasive questions about a woman’s child after she shared content on Instagram/Facebook. In a statement to The Verge, Meta says it “missed the mark” and that the feature should never have prompted those kinds of questions.
Valerant: An Automatic Navigable Game Map Generator via Action-Conditioned World Model Exploration
Valerant turns a single image into a persistent 3D game map by combining action-conditioned world-model rollouts with SLAM-based spatial reconstruction. The training-free pipeline uses exploration-driven action selection to progressively build navigable geometry – a key step toward making world models useful beyond 2D simulation.
Slack can now vibe-code interactive charts and reports inside chats
Slack is rolling out “Slackforce Surfaces,” letting users create interactive reports, dashboards, polls, and presentations directly in chat through AI assistance. Surfaces can be shared, pinned to channels, and updated using data from connected apps like Google Drive and Salesforce.
Do Agents Know When They Succeed? Calibrating Agent Confidence from Internal Representations
This work targets a core reliability problem for agentic systems: knowing whether actions will succeed in multi-step, tool-using workflows. It introduces Latent Trajectory Dynamics and an Action Representation Probe that predict task success from internal representations without prompt changes or multi-sample rollouts, showing consistent improvements across interactive benchmarks.
Rapidly scaling online storage to serve over 1 billion ChatGPT users
OpenAI describes the infrastructure evolution behind serving massive online storage loads for ChatGPT at internet scale, including performance and reliability concerns that show up only when you hit very high request rates. The piece frames scaling as an ongoing systems problem – not just a model problem – with careful attention to throughput, latency, and operational resilience.
Anthropic spent this week in hot water over cybersecurity
Following earlier admissions that its models sometimes hacked other organizations during cyber tests, Anthropic published a new report detailing multiple incidents. The Verge reports that the cases it describes involve exploitation and unauthorized access behavior Anthropic attributes to recklessness, raising fresh alarm about agentic cybersecurity risk.