AI2 is developing a large language model optimized for science

PaLM 2. GPT-4. The list of text-generating AI practically grows by the day.

Most of these models are walled behind APIs, making it impossible for researchers to see exactly what makes them tick. But increasingly, community efforts are yielding open source AI that’s as sophisticated, if not more so, than their commercial counterparts.

The latest of these efforts is the Open Language Model, a large language model set to be released by the nonprofit Allen Institute for AI Research (AI2) sometime in 2024. Open Language Model, or OLMo for short, is being developed in collaboration with AMD and the Large Unified Modern Infrastructure consortium, which provides supercomputing power for training and education, as well as Surge AI and MosaicML (which are providing data and training code).

“The research and technology communities need access to open language models to advance this science,” Hanna Hajishirzi, the senior director of NLP research at AI2, told ProWellTech in an email interview. “With OLMo, we are working to close the gap between public and private research capabilities and knowledge by building a competitive language model.”

One might wonder — including this reporter — why AI2 felt the need to develop an open language model when there’s already several to choose from (see Bloom, Meta’s LLaMA, etc.). The way Hajishirzi sees it, while the open source releases to date have been valuable and even boundary-pushing, they’ve missed the mark in various ways.

AI2 sees OLMo as a platform, not just a model — one that’ll allow the research community to take each component AI2 creates and either use it themselves or seek to improve it. Everything AI2 makes for OLMo will be openly available, Hajishirzi says, including a public demo, training data set and API, and documented with “very limited” exceptions under “suitable” licensing.

“We’re building OLMo to create greater access for the AI research community to work directly on language models,” Hajishirzi said. “We believe the broad availability of all aspects of OLMo will enable the research community to take what we are creating and work to improve it. Our ultimate goal is to collaboratively build the best open language model in the world.”

OLMo’s other differentiator, according to Noah Smith, senior director of NLP research at AI2, is a focus on enabling the model to better leverage and understand textbooks and academic papers as opposed to, say, code. There’s been other attempts at this, like Meta’s infamous Galactica model. But Hajishirzi believes that AI2’s work in academia and the tools it’s developed for research, like Semantic Scholar, will help make OLMo “uniquely suited” for scientific and academic applications.

“We believe OLMo has the potential to be something really special in the field, especially in a landscape where many are rushing to cash in on interest in generative AI models,” Smith said. “AI2’s unique ability to act as third party experts gives us an opportunity to work not only with our own world-class expertise but collaborate with the strongest minds in the industry. As a result, we think our rigorous, documented approach will set the stage for building the next generation of safe, effective AI technologies.”

That’s a nice sentiment, to be sure. But what about the thorny ethical and legal issues around training — and releasing — generative AI? The debate’s raging around the rights of content owners (among other affected stakeholders), and countless nagging issues have yet to be settled in the courts.

To allay concerns, the OLMo team plans to work with AI2’s legal department and to-be-determined outside experts, stopping at “checkpoints” in the model-building process to reassess privacy and intellectual property rights issues.

“We hope that through an open and transparent dialogue about the model and its intended use, we can better understand how to mitigate bias, toxicity, and shine a light on outstanding research questions within the community, ultimately resulting in one of the strongest models available,” Smith said.

What about the potential for misuse? Models, which are often toxic and biased to begin with, are ripe for bad actors intent on spreading disinformation and generating malicious code.

Hajishirzi said that AI2 will use a combination of licensing, model design and selective access to the underlying components to “maximize the scientific benefits while reducing the risk of harmful use.” To guide policy, OLMo has an ethics review committee with internal and external advisors (AI2 wouldn’t say who, exactly) that’ll provide feedback throughout the model creation process.

We’ll see to what extent that makes a difference. For now, a lot’s up in the air — including most of the model’s technical specs. (AI2 did reveal that it’ll have around 70 billion parameters, parameters being the parts of the model learned from historical training data.) Training’s set to begin on LUMI’s supercomputer in Finland — the fastest supercomputer in Europe, as of January — in the coming months.

AI2 is inviting collaborators to help contribute to — and critique — the model development process. Those interested can contact the OLMo project organizers here.

Related Posts