Some ideas for what comes next, May 2026

Details

In a wide-ranging Substack piece, Interconnects author Nathan Lambert lays out his current thinking on the AI landscape heading into mid-2026. His central thesis: the gap between open and closed frontier models is real, underappreciated, and best measured through real-world agentic task performance rather than benchmarks. Lambert's anchor example is the Claude Opus 4.5 / Claude Code moment of December 2025 — loud, obvious, and unreplicated by any open-weight model five to six months later. He estimates the open-closed gap in agentic coding tasks could persist for 12+ months, keeping Claude Code and Codex in a different product category. He is skeptical even of Google: Gemini 3.5 Flash reviews suggest it suits Google's internal products (search, YouTube) rather than serving as a coding agent competitor. If Google cannot close this gap, Lambert expects open-source labs to be even further behind. Lambert also flags several topics he has not yet written about: the 'Mythos' model, America's open-source surge, and emerging power struggles in the AI ecosystem. He frames 2026 as the first year with no breaks from accelerating disruption — more capability gains, more economic stakes, more real-world risks arriving simultaneously.