OpenAI x Jony Ive building a smart speaker

OpenAI's GPT-Red is an automated red-teaming model that uses self-play reinforcement learning to find AI vulnerabilities at scale, successfully making GPT-5.6 6x more robust to prompt injection attacks. This creates a safety flywheel where today's models help secure tomorrow's systems, addressing the scalability limitations of human red-teaming alone.

Supply Co. and Work Louder released a $230 specialized keyboard with RGB feedback and physical controls designed specifically for managing AI agent workflows and ChatGPT Codex interactions. The hardware represents emerging demand for purpose-built peripherals as agentic AI becomes embedded in developer workflows.

Simon Willison's Blog

Moonshot AI's Kimi K3 becomes the largest open-weight model at 2.8T parameters, achieving top-tier benchmark performance but at significantly higher pricing ($3/$15 per million tokens) than previous Chinese models. The model shows strong reasoning capabilities but consumes extensive reasoning tokens even for simple tasks.

Bloomberg

OpenAI plans to release a portable, screenless ChatGPT smart speaker in 2027 as its first major hardware device, featuring camera-based environmental awareness and mechanical elements for human-like interaction. The device is being developed with Jony Ive's design company and represents OpenAI's broader push into consumer hardware.

ReactBench

ReactBench v1 evaluates coding agents on realistic React work using both behavioral tests and 400+ deterministic quality rules, revealing that even leading models like GPT-5.6 Sol achieve only 43% success rates while introducing hundreds of production-critical bugs. This benchmark addresses the growing risk of AI-generated React code causing real-world outages and performance issues as models write more frontend code at scale.