Applications of Meta’s OPT-175B: Protein design, quantum physics, translation, and more

July 26, 2022

Since we announced OPT-175B in May, more than 4,500 individuals and institutions around the world have requested access to this groundbreaking large language model for AI research. OPT-175B was the first such model to be made freely available to the research community, providing an important new resource to accelerate work in this area of AI and ultimately help create safer, more useful, and more robust language generation systems.

As of today, we have granted access to 668 entities in 49 different countries, including about 300 universities and 80 industry research labs. These requests from external researchers have focused on tasks such as:

  • Scientific question-answering

  • Classification and prediction involving clinical warehouse data

  • Applications in mathematical logic: proofs of lemmas and theorems

  • Inspirations for new ideas in practical quantum physics research

  • Moderation of online platforms

  • Generation of quantum “games” (interactive proof systems)

  • Generative protein design

  • Novel algorithms for compression

  • Low-resource translation

  • Interpretability of other modalities via language interface

  • Bias analysis in survey interview research

We are excited by this initial response to OPT-175B and look forward to seeing how this model helps advance work in these different research areas. It’s been particularly interesting to see how other researchers have responded to our decision to release not just the model weights and code but also our notes and full logbook detailing the training process. We hope these additional resources can benefit newcomers into the field by providing a valuable behind-the-scenes look into implementation details that are not easily captured in published papers. We are also partnering with third-party organizations like the Partnership on AI so that we can help establish industry-wide norms around when and how to release these models.

We released OPT-175B (along with versions with fewer parameters) under a noncommercial license to focus on certain research use cases, and we have also taken additional precautions by gradually rolling out initial access to OPT-175B via a prioritization system. Requestors who:

  1. provide links to prior publications;

  2. have an email address matching an academic institution or industry research lab; and

  3. provide extensive details on the intended research use cases

are highly prioritized for initial access to OPT-175B. Each request has also been manually screened to verify the consistency and accuracy of information provided in the request form before the requestor receives a link to download the model.

Something Went Wrong
We're having trouble playing this video.

Our goal with this approach is to methodically expand access to the broader research community so that we can collectively help define the risks, limitations, and appropriate applications of these models. We will continue to refine this release strategy over time, as we obtain feedback from researchers, before we consider extending access beyond research use cases.

Written By

Susan Zhang

Research Engineer