OpenAI's Thibault Sottiaux says GPT-5.6 Sol reached state of the art on ARC-AGI-3 after two setting changes: sustained reasoning and canonical compaction across multiple context windows. Frontier performance increasingly depends on model-harness co-design.

OpenAI's Thibault Sottiaux frames the week around intelligence becoming too cheap to meter and teases another release. The signal is not capability alone, but faster progress in cost, cadence, and availability.

AI builder Peter Steinberger questions an unusually large result shared by Anthropic and argues for a basic plausibility check before turning benchmarks into victory narratives. Reproducibility and sound baselines matter more as eval competition intensifies.

Y Combinator President and CEO Garry Tan argues that the age of intelligence does not make experience obsolete. AI amplifies execution speed, which makes seasoned judgment about markets, organizations, and product tradeoffs more valuable.

Replit CEO Amjad Masad highlights Replit Design for its taste, ease, and sense of fun, not only generation speed. AI design tools are moving from can-generate demos toward products people want to use repeatedly.

Claude Blog introduces a Swift package that combines Apple's on-device models with Claude in one Foundation Models workflow. Local models handle fast private tasks, while Claude takes multi-step reasoning, code generation, web search, and data analysis, returning typed values and streamed tool results to SwiftUI.

Wired's Kevin Kelly argues that technology timelines routinely defy prediction. VR became vastly cheaper yet still awaits its LLM moment, while AI looked like an overnight success after roughly fifty years of accumulation. He treats intelligence as an undiscovered compound and calls for both open frontiers and mature governance.