AI News from October 3, 2026

OpenAI safety employee David Robinson has resigned, saying its culture cannot manage rising AI risks and calling for stronger outside oversight.

Circuit Breaker Labs is testing AI with simulated users across languages and cultures, targeting psychological safety risks in mental health and youth apps.

A TechCrunch roundup shows how AI agents are using text messages to manage calendars, travel, family tasks, email, and other real-world workflows.

PewDiePie says Ajax is an uncensored AI model for home PCs and alleges two OpenAI bans over model distillation, spotlighting local AI and enforcement.

A reported fivefold gap in token use between AI agents and human users highlights rising inference costs, but the underlying data remains unavailable.

A reported Argo-Bench result shows the leading AI agent solved 34.8% of 210 tasks, highlighting unresolved reliability limits for autonomous systems.

Apple is reportedly preparing tighter Mac data controls after complaints involving Meta Muse, a move that could reshape how AI agents access desktop apps.

LEGO-Anything shows coding agents can build editable 3D scenes from photos, but LEGO-Bench exposes major gaps in accuracy and self-assessment.

Former AlphaGo researcher Thore Graepel argues that LLMs need auditable belief states and search, not simply longer chains of thought.

Clay CEO Kareem Amin will outline how AI is creating GTM engineering roles that turn prospect research, data, and outreach into automated growth systems.

A WP Intelligence listing highlights an Intelligence Council report on open-weight AI models and their growing role in U.S.-China competition.

Cerebras shares reportedly fell 20% to a post-IPO low as Nvidia pressure weighed on sentiment, highlighting risks for AI-chip challengers.