My Bets on Open Models, Mid-2026
TL;DR: Nathan Lambert (Interconnects AI) lays out 13 positions on the open vs. closed model race as of mid-2026. Core thesis: the gap is primarily an economics question, not a capability one — open models track benchmarks well but closed models have harder-to-measure robustness advantages for direct assistant use.
Key Findings
- Closed models' benchmark lead did not grow as expected given their training compute advantages — surprising to most observers.
- Chinese open-weight labs focus slightly more on benchmark scores than U.S. closed labs, aided by distillation.
- Closed model advantage: harder-to-measure robustness and general usefulness in novel situations — critical for knowledge worker assistants.
- RL-era shift: distribution to real-world use cases (e.g. Claude Code, Codex) is now a key frontier differentiator — closed labs can leverage online RL from user feedback.
- Open models will dominate repetitive automation tasks (API market, backend automation) while closed dominate direct knowledge work.
- Chinese open-weight labs expected to face funding difficulties later in 2026 — capability trajectory changes 3–9 months after.
- Local agents (OpenClaw, etc.) represent large, mostly untapped market for open models.
Related Pages
Raw source: ../../raw/rss/2026-04-15-interconnects-ai-my-bets-on-open-models-mid-2026.md