Allen Institute for AI (AI2)

Contribution to AI

Releasing a model's weights became ordinary practice long before releasing what it was trained on did, and the gap is where this institute did its most useful work. ELMo, in 2018, showed that a word's representation should be a function of the sentence around it, computed from the internal states of a bidirectional language model; bolted onto existing systems it improved results across six separate NLP tasks and established the pretrain-then-adapt habit that BERT inherited months later. AllenNLP turned that into infrastructure other groups could build on rather than a result they had to reimplement. A parallel line of work was about measurement: ARC and WinoGrande were designed so that surface statistical cues would not carry a model through, and Aristo's climb to over ninety per cent on the non-diagram portion of the New York State eighth-grade science exam in 2019 gave a concrete marker on a problem where the best system three years earlier had scored under sixty. OLMo in February 2024 pushed the openness argument to its limit by shipping the pretraining corpus, Dolma — three trillion tokens under an open-data licence — along with training code, intermediate checkpoints and evaluation harnesses, so that a claim about a large model could be checked rather than taken on trust. Semantic Scholar, meanwhile, made the literature itself machine-readable and free to query.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).