The Allen Institute for AI (AI2) has introduced OLMo, a revolutionary open-source large language model designed to drive a significant transformation in AI development. Founded in 2014 by Microsoft co-founder Paul Allen, AI2 aims to drive a critical shift in AI development by offering a comprehensive solution that includes not only model code and weights but also training code, training data, and associated toolkits, all under the Apache 2.0 license.
In a landscape dominated by closed, proprietary models like OpenAI’s GPT-4, OLMo represents a breakthrough for open-source AI. Recent developments, such as the leak of a new open-source AI model rivaling GPT-4’s performance and Meta’s release of its Code Llama 70B model, underscore the growing momentum of open-source AI initiatives.
Despite its momentum, open-source AI faces criticism from some quarters, with concerns raised by researchers, regulators, and policymakers about its potential risks. However, proponents argue that OLMo’s comprehensive openness, providing access to data, code, model weights, and evaluation tools, will enable researchers to better understand and study the underlying mechanisms of language models, fostering the development of safer and more trustworthy AI systems.
Hanna Hajishirzi, OLMo project lead at AI2, emphasizes the importance of transparency in AI development, likening the lack of access to training data to conducting drug discovery without clinical trials. With OLMo, researchers gain unprecedented insight into the workings of large language models, empowering them to explore new research directions and enhance ML research and deployment.
Nathan Lambert, an ML scientist at AI2, highlights OLMo’s role in enabling novel approaches to ML research and deployment, thanks to its unparalleled openness. Jonathan Frankle, chief scientist at MosaicML and Databricks, lauds AI2’s OLMo release as a significant advancement for open science, while Hugging Face’s CTO recognizes OLMo for pushing the boundaries of open-source AI.
Yann LeCun, Meta’s chief scientist, acknowledges the pivotal role of open foundation models in driving innovation and development in generative AI. He underscores the importance of a vibrant open-source community in shaping the future of AI, echoing the sentiments of AI2’s commitment to fostering collaboration and transparency in AI research and development.