Interconnects AI · 2026-04-15 · notable
Nathan Lambert: My bets on open models, mid-2026
Ai2 post-training lead Nathan Lambert lays out his mid-2026 bets on open vs closed models: benchmarks are still closing but RLHF-driven 'robustness' and real-user interaction data keep closed labs ahead on general-purpose professional workloads.
A post-training lead at Ai2 writes down, in mid-April 2026, exactly where he thinks open and closed models will diverge for the rest of the year.
What is it?
An Interconnects essay by Nathan Lambert — senior research scientist and post-training lead at the Allen Institute for AI — laying out seven concrete predictions about the open-vs-closed-model race for the rest of 2026. It is the kind of piece that gets read and cited in lab all-hands, because Lambert is one of the very few full-time practitioners who publishes detailed thinking about post-training dynamics.
How does it work?
The essay argues that the capability gap on public benchmarks is narrowing, but benchmarks miss the 'robustness' and general-purpose quality that closed labs get from RL over real user interactions. Chinese open-weight labs currently lead on benchmark optimization but may hit economic headroom within the year. Open models will dominate repetitive automation and specialized domains; closed models retain leadership in general-purpose professional assistance. Governance pressure paradoxically raises demand for open models as a counterweight.
Why does it matter?
Most 'open vs closed' discourse is tribal or shallow. Lambert grounds his calls in specific mechanisms (RLHF pipelines, preference data, distillation, economics) and stakes a dated position — the kind of writing that makes it possible to check back in six months and see who was right. It is required reading for anyone making product or infra bets that depend on which side of the frontier moves faster.
Who is it for?
Founders, infra leads, and ML researchers making multi-quarter bets on open vs closed model stacks.
Try it
https://www.interconnects.ai/p/my-bets-on-open-models-mid-2026