After weeks of speculation and anticipation, Meta made it’s open source LLaMA 2 Foundation LLM AI model official today, available via a distribution partnership with Microsoft and its Azure Cloud service:
“Meta and Microsoft will release a new version of Meta Platforms’ artificial-intelligence language model, LLaMA2, in a partnership between the tech giants that will make the software available to companies for the first time.”
“In making LLaMA 2 available to Azure customers, Microsoft is showing a willingness to broaden beyond ChatGPT maker OpenAI as its AI platform of choice. Despite the fact that Microsoft has committed billions to OpenAI, other cloud-computing providers like Amazon and Google are increasingly market-ing themselves as neutral platforms where developers can pick among an array of generative AI models from different companies.”
“Meta is releasing LLaMA 2 as “open source” software, which typically is made widely available for use, modification and sharing by the public. The earlier version put the company at the center of a vibrant, if uncontrolled, surge in AI software development.”
Importantly, Meta is also doing the following:
“We’re now ready to open source the next version of Llama 2 and are making it available free of charge for research and commercial use. We’re including model weights and starting code for the pretrained model and conversational fine-tuned versions too.”
The model weights are an important element for a lot of developers and businesses in using LLaMA 2 for their own AI applications and services. LLaMA 2 will be available in three sizes with 7, 13 and 70 billion parameters. That compares to a 175 billion parameters for OpenAI’s GPT 3 LLM AI model, and reported 1.8 trillion parameters for GPT 4. So in that context, these Foundation models vary across the size spectrum. But even in these sizes, they have powerful research and commercial uses.
Some additional technical details from the company on LLaMA2 were as follows:
“The technical advancements — 40% more data, 2x context window, grouped-query attention, and all of the improvements in our fine-tuning, RLHF, etc.”
We are also seeing some details around the possible open source licenses and conditions, as this Runtime report indicates:
“Meta did not specify the terms of its license in its announcement, but Axios reported that the license was "a customized partial open source license" and Databricks advised customers "there are some restrictions."
“Among other things, if your company serves more than 700 million monthly active users — a category that certainly includes Google but also Facebook competitors like TikTok and Snap — you'll have to request a license from Meta to use LLaMA 2.”
“Those restrictions are not sitting well with the open-source community, as it appears Meta is borrowing the goodwill associated with "open source" while forgoing the use of a license approved by the Open Source Initiative, which is generally considered the minimum requirement to appropriate the term.”
“Also, "Meta is not releasing information about the data set that it used to train LLaMA 2 and cannot guarantee that it didn’t include copyrighted works or personal data," according to MIT Technology Review.”
Importantly, Meta builds on its game-changing LLaMA models introduced in March, and its tradition of serving the open source community for over 15 years. That further leverages and accelerates developers ability to take Meta’s open source AI user interfaces and tools like React and Pytorch, which have become a huge tail-wind for open source AI developers worldwide.
And working details of the model this time won’t have to be leaked for intense developer activity to be accelerated around open source AI. LLaMA 2 comes with weights accessible, so new products and services can be built on top of the AI Tech stack. Meta will also distribute LLaMA 2 via cloud services beyond Microsoft Azure, via service like Amazon AWS, Hugging Face, and other AI software distribution hubs.
And importantly, further development of AI software and services on top of the models, can be done by smaller companies and businesses of all sizes worldwide. This leverages open source models and tools to accelerate the trend of AI eating software, and indeed, remakes what’s possible with AI software and hardware overall. Full speed ahead just got faster.
As we’ve discussed earlier, open source AI has been a strong enough force for engineers at Google and OpenAI/Microsoft to be concerned about from a competitive perspective, as aired in a Google engineer’s leaked memo a few months ago regarding ‘Google’ (and OpenAI’s ‘We have no Moat’). And even though the new head of Google’s global AI head Demis Hassabis says he is not worried, given Google’s AI strategy,
And today’s Meta announcements continue to build on Meta’s AI opportunities on top of the infrastructure tech stack with its Apps, Ads, and services, and of course itslatest Twitter competing Threads, with over a 100 million users.
Leveraging its massive daily Data, images, and video assets, in reinforcement learning loops, Meta is doing to have a huge say in where AI goes over the next three years and beyond. Meta is likely one of the largest customer already for Nvidia’s much in demand AI GPU chips in this AI Gold rush, for the foreseeable future.
All this makes Meta an AI Force to be reckoned with by all industry participants. With or without a cage match between Zuckerberg and Musk, and Elon’s plans for his own Foundation LLM AI company xAI.
And a counterforce to the Foundation LLMAI model strategies deployed by OpenAI/Microsoft, Google, Amazon and many others, who are counting on business models off the model leveraged by API and data and software services integrated into its core sets of models, in the cloud and at the Edge.
This last bit is particularly important since LLaMA 2 small enough to run on local devices closer to users, and not depend primarily on the cloud for reinforcement learning, inference and training loops. In fact, Meta in its announcement highlighted that they’re working with Qualcomm to ensure LLaMA 2 can run natively on phones and other devices rather than being Cloud dependent.
One can see companies like Apple experimenting with LLM Ads in this direction on Apple Silicon in the hands of over 2 billion users using Macs, iPads and iPhones. With dozens of commercial open source LLM AI models on the smaller side available, Apple too has a lot of choices ahead. It’s not out of the question for OpenAI to potentially decide to open source their smaller parameter sized GPT3 and/or GPT 3.5 models for such applications via Partners. Counter-moves like that are possible vs what Meta is doing with OpenAI’s equity partner Microsoft.
Meta is throwing sand into those gears indeed and there are multiple chess games afoot concurrently. In another metaphorical context, the global AI gladiator games got more interesting. Stay tuned.