AI Safety’s “Pace the Frontier” Moment – Is It Protection or a Cartel?
The Verge frames the current push to slow or “pace” frontier AI as a live, high-stakes debate involving major labs and safety claims that may or may not translate into enforceable restraint. The piece collates arguments and tensions across the industry, with skepticism about motives and the practical question of whether regulators will actually compel changes fast enough.
CapMem – Caption-Based Episodic Memory for Egocentric Video
Researchers propose CapMem, a caption-centric benchmark for wearable/egocentric assistants where visual tokens are too expensive and long-context retrieval can fail. The surprising takeaway – for long videos, caption-guided retrieve-and-verify can beat direct video QA, suggesting text captions can function as reusable “episodic memory” when frames are constrained.
Claude Code “Projects” – Running Multiple AI Agents with Shared Memory
The Verge reports that Claude Code is getting a redesigned Projects feature that lets users run multiple agent threads under one umbrella with shared goals, files, and artifacts. The architecture description – each thread is a separate cloud session with merge-conflict handling – highlights how agent collaboration is being packaged into something closer to real software workflows.
GraphEcho – LLM Graph Agents Can Mistake Repeated Paths for Evidence
GraphEcho tests whether graph-walking LLM agents treat redundant revisits as new corroboration even when evidence content is held constant. The work shows models can shift judgments and repeatedly traverse the same evidential routes, and it finds provenance-aware post-training reduces revisits but can trade off toward fewer distinct sources.
Snap Debuts “Specs Intelligence” – An iOS/Mac AI Assistant Tied to AR Glasses
Snap is launching “Specs Intelligence,” an assistant positioned as anticipatory account management and task help, available on iOS before expanding via its new Specs AR glasses. The rollout signals how consumer AR is pivoting from pure vision features toward agentic workflows that connect to your digital accounts and surface what needs attention.
EvolveTrade – Self-Evolving Tool-Use Policies for LLM Trading Agents
EvolveTrade treats an LLM trading agent’s tool-use “system prompt” as a text-parameterized policy that an update agent revises over time using decision traces and portfolio feedback. By keeping the backbone model fixed while iteratively refining the tool policy, the study reports improved Sharpe ratio and cumulative returns across multiple market regimes.