Reading Notes
Links, short summaries, and notes on what is worth noticing across AI, operations, strategy, and company-building.
Archive
Thinking Machines is betting that useful AI will be customizable, specialized, and owner-operated. The argument moves beyond centralized general models toward systems that extend local knowledge and human judgment instead of replacing them.
OpenAI's new model lineup pairs stronger coding and agent capabilities with a unified ChatGPT and Codex desktop experience. The release also lowers developer costs and gives users more control over orchestrating complex work.
Cursor's usage data reveals a steep power law among AI-assisted engineers and confirms that agents spend far more tokens reading code than writing it. Context processing and caching are becoming core infrastructure concerns.
Chinese technology companies are closing the AI performance gap with domestic chips, competitive frontier models, and AI-first engineering practices. The result is an increasingly self-sufficient stack spanning silicon, models, and software development.
Uber found its strongest AI use cases by pairing engineers with domain experts and rebuilding complete workflows. The approach sharply reduced planning and finance cycle times while exposing approvals and legacy tools that no longer added value.
AI advantage is moving beyond raw model capability toward hidden reasoning states, modular post-training stacks, and continuous agent learning. Builders who understand these layers can improve performance, lower dependence on closed models, and turn production evidence into better systems.
Autonomous agents become useful only when they can identify work, act safely, and prove that real-world changes occurred. Proactive monitoring and state-aware evaluation offer a practical path beyond prompt-driven assistants and static text grading.