A few months ago, everyone wondered who would win the AI arms race. Microsoft aligned with OpenAI. Google launched Bard. Meta started working on its own large language model, LLaMA. Other companies started thinking about launching AI platforms, and curious users pitted the models against each other.
But a recent deal suggests we may see a growing number of partnerships, too, not just head-to-head competition. Earlier this week, Meta offered its LLaMA large language model for free under an open license and brought it to Microsoft’s Azure platform. The decision highlighted the benefits of interoperability in AI, and as more companies join the field, it likely won’t be the last of its kind.
The well-known LLMs to date have been relatively isolated and offered in a more controlled environment where users need permission to build with the model or use the data. OpenAI continues to train GPT, releasing GPT-4 in March and provide developers with access to the payment API to the latest version of your model. Apple is developing its own LLM, called Ajax, though details are scant; it is not yet publicly available and its open source status is unknown. Bard, Google’s LLM, is not open source at all.
Initially, LLaMA was not publicly available and only accessible through Meta, and Meta has yet to disclose its training data. But LLaMA was always intended to be open source and was created to “further democratize access” to AI. This week, Meta fulfilled, at least in part, that promise. Users of closed systems must pay a license fee to access the model where it is hosted and distribute applications using that same model. The way that Meta opened up LLaMA, by making it available to Azure users and unlicensed to some degree, eliminates that drawback.
Meta opening up LLaMa and bringing it to Azure makes business sense, especially if Meta believes in open AI development. It is a first step to allow people to access more LLM models on platforms and compare the results. A greater variety of LLM frameworks to choose from also highlights the question of how each model can work together. And LLM developers want people to use their models, so having them available on a wide range of platforms brings them closer to more users.
Even the most competitive Big Tech companies do business with each other. Meta is no stranger to working with Microsoft: Meta brought Microsoft’s Teams product to Workplace by Meta, which already runs the Office 365 suite.
Openness has its risks. Ilya Sutskever, co-founder and chief scientist at OpenAI, a more open organization when it was founded in 2015, said the edge regrets sharing research for fear of competition and security. Opening up data sets makes it easier to sue for copyright infringement, for example, because people can see what sources were scraped for data to train models.
But having more LLM frameworks to choose from could be good news for AI interoperability advocates.
Since LLMs are, by default, different from each other, developers often have to choose which model to build apps with. There is no good way for systems to talk.
Walled gardens don’t surprise most users of modern technology, but proponents of AI interoperability argue that the only way AI can grow and evolve is not through closed silos, but through open structures that can communicate with each other. Even Microsoft believes in interoperable AI; joined other tech companies to join the Open neural network exchangea group that wants to promote an industry standard for AI interoperability so that developers can “find the right combinations of tools.”
Allowing AI systems to work together could lead to better results for things like search queries. Companies that can train models on different data sets could provide better and more comprehensive service and, if a model is wrong, potentially avoid catastrophic overreliance on one data source. And being able to develop for LLaMa and OpenAI’s GPT models in one place could reduce development costs and timelines.
For now, the fact that LLaMa is available on Azure does not mean that applications built with LLaMa can suddenly communicate with those running on OpenAI’s GPT models. No one has created that bridge yet. Also, not everyone agrees that LLaMa meets all the requirements for open source software, especially since it does not use a license approved by the Open Source Initiative. It also limits who can commercially use LLaMa without paying a fee. for his community license agreementdevelopers who have more than 700 million monthly active users “must request a license from Meta.”
But this is a good step in the right direction for open source and interop, if only to allow developers easier access between models. There is room for healthy competition, but if companies really want AI to evolve, working together is the best option.
