Meta's double-headed Llama
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Gabriella Turrisi/Axios
Meta policy chief Nick Clegg wants you to be impressed by the powers of its latest open source AI model, known as Llama 2 — but not so impressed that you worry about the havoc it could wreak in the wrong hands.
Why it matters: Meta has opened the new model to allow anyone to use it commercially for free — the prior version, released in February, was for research use only.
- Those using Llama 2 have to agree to an acceptable use policy, but once the code is out there, those rules could prove tough to enforce.
Driving the news: Meta sounded its trumpets Tuesday for the Llama 2 release, announcing partnerships with a variety of industry giants including Microsoft, Amazon and Qualcomm.
- Microsoft will be the preferred cloud partner, getting Llama 2 first for Azure, but Amazon also plans to offer Llama through AWS.
- Qualcomm, meanwhile, is working with Meta to ensure that Llama can run natively on phones and other devices rather than relying on the cloud.
Zoom out: Meta's open-source AI strategy aims to help the social giant catch up to the phenomenal popularity of OpenAI's ChatGPT.
Yes, but: Government experts worry that free, powerful AI models available for re-engineering could hasten the emergence of threats like genetically engineered bioweapons.
What they're saying: Clegg, Meta's global affairs president, used an Axios interview to talk down Llama's capabilities.
- While acknowledging “you can't predict or litigate for all downstream uses," of AI, he argued Llama 2 could not be categorized alongside "frontier models" which are defined as highly capable models in risky fields.
- Meta is not aiming to create “all-singing, all-dancing artificial general intelligence," Clegg said, calling its Llama models "much, much dumber than that. They're just a textual predictive pattern recognition system.”
Between the lines: Misinformation is universally viewed by experts as a top AI risk.
- Because Llama 2 is designed to run on devices as well as in the cloud, Meta may have a hard time holding users to its acceptable use policy.
- And Meta-owned Facebook and Instagram — like the rest of the world — are going to have to deal with whatever Llama spits out.
The big picture: Llama 2 is the latest of a number of projects that Meta has released via something akin to open source.
- Llama 2's release isn't technically open source, per the Open Standards Initiative, but the fact that anyone can use it and access much of its code is likely to insure rapid adoption and innovation around the model.
- Meta has tended to favor open sourcing key technical projects, including PyTorch and its AI work, with over 600 of its models now available for use on the HuggingFace platform.
- Clegg noted the company does not automatically open source models, and wrote last week that there are cases where closed models are necessary for safety.
What they're saying: Efforts to test how Llama 2 reacted to demands for sensitive material in fields such as nuclear, biological and chemical weapons led to “very marginal” issues, Clegg said, insisting these “could easily be mitigated.”
- Meta will submit its Llama models to the DefCon hackathon in August for further stress testing.
