Expanding its efforts in the generative AI sector, Meta in partnership with Microsoft announced the availability of its open-source large language model Llama 2, which is essentially a family of language models that are used to power generative AI tools, on July 18. In a blogpost, Meta stated that Llama 2 will be available free of charge for research and commercial use.
Llama 2 can now be accessed in the Microsoft Azure AI catalog, through Amazon Web Services (AWS), and Hugging Face among other providers. Additionally, according to Meta, the language model is also “optimised to run locally on Windows”.
In a research paper titled ‘Llama 2: Open Foundation and Fine-Tuned Chat Models’, the team detailed the methods employed to train the language models used to build Llama 2, techniques for safety evaluation, and discussions around limitations as well as responsible use of such models.
How’s it different from Llama? In February this year, Meta had released LLaMA—Large Language Model Meta AI, under a non-commercial license focused on “research use cases”. While access to Llama was restricted and granted to academic researchers on a case-to-case basis, Llama 2 is an open-source resource for developers and businesses to build their own tools.
Article continues below ⬇, you might also want to read:
- Kenyan Workers Call For Probe Into Disturbing Work Conditions In AI Content Moderation For OpenAI
- Five Talking Points: Open AI’s Plan To Launch An App Store For AI Models
On capabilities of Llama 2: In its research paper, Meta has claimed “These models have demonstrated their competitiveness with existing open-source chat models, as well as competency that is equivalent to some proprietary models on evaluation sets we examined, although they still lag behind other models like GPT-4.”
As AI commentators seem to pitch Meta’s language models against Open AI’s GPT-4 which powers ChatGPT and Microsoft Bing, AI scientist at NVIDIA Jim Fan noted on Twitter, “Llama-2 is NOT yet at GPT-3.5 level, mainly because of its weak coding abilities.” Having said that, experts, including Fan, have lauded Meta’s efforts in providing significant amount of detail on its language models in the white paper.
You'll soon see lots of "Llama just dethroned ChatGPT" or "OpenAI is so done" posts on Twitter. Before your timeline gets flooded, I'll share my notes:
▸ Llama-2 likely costs $20M+ to train. Meta has done an incredible service to the community by releasing the model with a… pic.twitter.com/MrABHrmACv
— Jim Fan (@DrJimFan) July 18, 2023
On training data: Non-transparency when it comes to data used for training large language models remains a point of contention here as well. Abeba Birhane, Senior Research Fellow at Mozilla Trustworthy AI, pointed out on Twitter that Meta’s research paper does not clearly define the sources that were used to train the models.
why is the dataset section bare? what "publicly available sources"? and which cites containing high volumes of personal information? pic.twitter.com/kXYqV0SWxF
— Abeba Birhane (@Abebab) July 18, 2023
STAY ON TOP OF TECH NEWS: Our daily newsletter with the top story of the day from MediaNama, delivered to your inbox before 9 AM. Click here to sign up today!