Allen Institute for AI (AI2)
The Allen Institute for AI (AI2) is a Seattle-based non-profit AI research organisation founded by the late Paul G. Allen. It is known for committing to fully open AI research — publishing models, training data, and code simultaneously.
Key Contributions
- OLMo (Open Language Model) — AI2’s flagship open-source LLM series; noteworthy for releasing not just model weights but also training data and training code, enabling full reproducibility
- OLMo 3 uses Sliding Window Attention for efficient long-context handling
- Tulu 3 — AI2’s open post-training recipe; co-coined the term rlvr (Reinforcement Learning with Verifiable Rewards); one of the first public demonstrations of the RLVR pipeline
- Semantic Scholar — AI2’s academic search engine; its paper corpus (OCR-extracted PDFs) is cited as a high-quality pre-training data source
Personnel
- nathan-lambert — Post-training research lead; key author of the Tulu 3 RLVR work; RLHF Book author
Philosophy
AI2 occupies a distinct position in the AI landscape: fully open (not just open-weight), non-profit, and academically oriented. Its releases are used as baselines and research platforms by the broader community.