OpenAI Steps Up Work on ‘G3PO’ Open-Source LLM to Compete with Meta: Report
Meta’s introduction of open-source large language model (LLM) Llama 2 is adding pressure to OpenAI’s work on its own long-rumored G3PO open-source LLM, according to a report from The Information. OpenAI has been working on G3PO for a while but has not made any official announcements or hinted at a release date, though Meta’s rollout and distribution arrangement with Microsoft, OpenAI’s closest partner, may accelerate things.
Meta made an end-run around the revenue streams of OpenAI, Google, and many other LLM developers when it opted to license Llama 2 to businesses for free. Even with the caveat brought up by CEO Mark Zuckerberg that big tech companies will still have to pay for the data, it’s a strategy with obvious appeal for brands debating among LLM and generative AI software providers. Llama 2 upgrades the power of the earlier iteration, including doubling the context window and using a 40% larger training database.
The leaked reports about an open-source model underway at OpenAI started appearing a couple of months ago, but this is the first time it’s been given a name. The lack of dates for any release suggests G3PO will start out a step behind in competing against Llama 2. On the other hand, OpenAI may feel a mix of proprietary and open-source models will mean greater success. That said, OpenAI is apparently worried that developers might flock to open-source models from a brand as recognizable as Meta more quickly than they’d like, so G3PO’s timeline might see some acceleration.
OpenAI certainly has plenty of other irons in the fire beyond the still nebulous G3PO. The company created the Superalignment group to work on controlling advanced AI, started a couple of journalism-based AI projects, and introduced a Custom Instructions feature that may presage a fully personalized ChatGPT assistant. That’s on top of a supposed app store, where developers can shop for different ways to leverage OpenAI’s models, set to roll out in the near future.