Ai2 at NVIDIA GTC 2026 @allenai

Uploaded March 2026 | Updated September 2026, 5 hours ago

What does truly open AI look like in practice?

At NVIDIA GTC 2026, we shared our latest work across panels, livestreams, and live demos--including Olmo Hybrid, SERA, Asta AutoDiscovery, and MolmoBot--all built around a simple idea: openness means sharing the full pipeline, not just model weights.

From research to real-world systems, this is how we’re building AI that’s reproducible, adaptable, and truly open.

Mitigating Knowledge Collapse through Epistemic Diversity

olmOCR: an open-source tool to extract clean plain text from PDFs!

BLADE: Benchmarking Language Model Agents for Data-Driven Science

Just-DREAM-about-it: Figurative Language Understanding with DREAM-FLUTE

Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking

Towards the Age of Computation and AI for High Performance Climate

When Not to Trust Language Models: Investigating Effectiveness of Parametric&Non-Parametric Memories

Open-Ended Learning Leads to Generally Capable Agents | Embodied AI Lecture Series at AI2

From LLMs to Agents: Generalizability from the Inside Out

Data-Centric Approaches to Adapting Foundation Models