Meta Enters The AI Race With An “Open” Approach
What Is Llama 2 and what does it mean for the competitive landscape of generative AI
On the heels of a successful launch of Threads, Meta released another important product last week that has garnered much less attention by comparison, but one that could prove to have a huge impact on consumer technology down the road. The product in question? Llama 2, the newest version of Meta’s large language model.
So far, the development of consumer-facing generative AI has been a two-horse race between Google and Microsoft, the latter of which benefited greatly from its early investments in and close partnership with OpenAI, the company behind ChatGPT, whose viral adoption kicked off this current round of AI frenzy.
With the arrival of Llama 2, the competitive landscape of generative AI might start to shift from primarily “Microsoft vs. Google” to a showdown between two different development models — Open Source vs. Closed. Up until now, the competition has been primarily between Microsoft and Google, both favoring a closed model. However, the open-source nature of Llama 2 could attract new players to the field and contribute to a more diverse and competitive market. This could accelerate the development of consumer-facing AI applications and ultimately benefit consumers.
What Makes Llama 2 Special
Llama stands for Large Language Model Meta AI. As its name suggests, it is a large language model (LLM) created by Meta; the first edition was officially released in February 2023 not as a public chatbot (a la ChatGPT), but as an open-source package that anyone in the AI community can request access to. A week after, however, the entire model was leaked online and spread across various AI developer communities. But as a “raw” LLM model, Llama has not been “fine-tuned” for conversations like ChatGPT or Bing has, which means that average users wouldn’t be able to use it even if they downloaded a leak. Nevertheless, the leak sparked some questions about Meta’s role in the generative AI space.
Fast forward to last week, on July 18, Meta announced Llama 2 in partnership with Microsoft. According to Meta, Llama 2 was trained on 40 percent more data when compared to Llama 1, which includes information from “publicly available online data sources.” AI researchers have found that when compared to GPT-4, the latest LLM by OpenAI, Llama 2 excels at writing texts and is more user-friendly, whereas GPT-4 is more adaptable and can handle more difficult tasks such as coding. Half of ChatGPT 3.5’s size, this new model is portable to smartphones and, unlike LLaMa 1, fully fine-tuned for conversations.
More importantly, Llama 2 is being released on an “open source” model, meaning that it’s free for all research and commercial use — as long as any AI applications built upon it do not exceed 700 million monthly active users, (a clear guardrail preventing other big tech companies from using it for free.) Given this limitation in the fine print of its licensing agreement, Llama 2 is technically not ”open source” in the purest sense of that term, but for the vast majority of developers tapping into this LLM courtesy of Meta, it is a free-to-use resource by all practical intent and purposes.
While Llama is far from the only open source LLM available — check here for a full list — it might very well be the most useful for AI developers. Developing an LLM can be an expensive endeavor, especially for AI training and fine-tuning with human feedback. The wide availability of Llama 2 could spur a lot more experimentation, without developers needing to pay OpenAI or Google for access to their respective LLM, thus undermining a key part of the current business models for the market leaders in the generative AI space. Of course, it’d still cost a pretty penny to buy a slew of GPUs required to run Llama-based AI applications at scale, but that’s nothing compared to the licensing fees that OpenAI is charging for access to its GPT-4 model.
What This Mean for the AI Competitive Landscape
The launch of Llama 2 is a significant event in the competitive landscape of generative AI. As the first major open-source LLM developed by a “big tech” firm, it could have a major impact on the way that generative AI models are developed and deployed.
Up until now, the generative AI market has been dominated by closed-source models from companies like OpenAI and Google. While these models can be quite powerful, they are also very expensive to develop and maintain, which has made it difficult for smaller companies and independent developers to get involved in the field.
In theory, Llama 2 alters this equation. It is a powerful LLM that is available to anyone to use, which means that smaller companies and independent developers can now compete with the big players in the field.
The reality, of course, is a lot more complicated than that.
First of all, Llama 2 was announced as part of a Microsoft event. Although it is being released on a practically open source model, (and Meta has said it will be integrated into everyone from AWS to Hugging Face to even Alibaba — just not Google!), Meta’s LLM has been strongly associated with Microsoft, who, despite its close partnership with OpenAI, seems to be hedging its bet on the open vs. closed model debate. By the virtue of association, however, Microsoft gets to promote its Azure cloud service as where most developers can access Llama 2, while Meta gets a legitimacy boost by associating its LLM with one of the leaders in the generative AI market.
By all accounts, OpenAI seems to be less interested in making a killing financially by commercializing its LLMs and more interested in developing the ultimate artificial general intelligence (AGI). Of course, regular users won’t care which AI tool is powered by which LLM, as long as it works. For Microsoft, being the primarily consumer-facing interface where mainstream users would encounter generative AI-powered features and creative tools is a great position to be in. After all, this is a company already planning to charge $30 a month for its AI Co-pilot feature in Windows and Office Suite.
Secondly, there is regulatory uncertainty over which model governments around the world would prefer to cultivate. Many believe that the arrival of Llama 2 destroys the competitive moat for closed models, which may end up slowing down the release of future updates of LLMs. In May, the Semianalysis blog published a memo from an anonymous Google engineer warning that the open-source LLMs would eventually overtake anything closed models could do. The paper was aptly titled: “We Have No Moat, And Neither Does OpenAI.” OpenAI has been arguing, perhaps somewhat self-servingly, for the closed model (ironically given the name), advocating for carefully controlled access to LLMs so as to minimize risk both of mis-use in terms of AI-generated mass spam or misinformation, and the broader risk of a sudden arrival of technological singularity.
On the other hand, those who advocate for an open model argue that you cannot stop the potential mis-use of generative AI now anyway since open source LLMs have already proliferated the market, and there is no guarantee that a closed development model would effectively minimize the impact of the singularity, whose feasibility is still up for debate. And if we were to open it up to all developers, the AI market would develop more rapidly, and everybody will get an increasing revenue stream as the total pie of consumer-facing AI keeps growing. It will be interesting to see how various governments around the world react to this debate and pick a side. The policies around this space, or the lack thereof, would certainly push the proverbial thumb on either side of the scale.
Another thing to consider here is Meta’s reputation over data management, which has historically not been particularly great, and whether that would cause resistance for some businesses to work with Meta’s AI model. While many brands do rely on its unified ad suite to reach the billions of users of Meta platforms, they may be less willing to trust Meta with their own first-party, proprietary data, which would be a necessary step to customize Llama into a generative AI tool that suits their particular purposes, whether internal or consumer-facing.
Lastly, there is also the question of how long Meta can afford to foot the bill of continuously updating a LLM that it will, for the most part, give away for free. In its most recent earnings report, Meta forecast a 20% revenue growth in Q3, returning to pre-pandemic and pre-ATT levels. This could potentially create some cover for its expensive investments in AI and metaverse. But if it wants to beat the closed models, it will need to form a strong anti-Google Ai alliance with the other big tech firms, aka Microsoft and Amazon at the moment. (Apple is reportedly testing its own “Apple GPT,” presumably based on a closed model because, well, it’s Apple.)
So far, Microsoft’s Azure has not been quite pulling its weight. Still, Microsoft remains hopeful. The company acknowledged in its latest earnings call that “even with strong demand and a leadership position, growth from our AI services will be gradual as Azure AI scales and our copilots reach general availability dates.” The arrival of Llama 2 could ignite some new interests in Azure, but since it is available on AWS, still the leading cloud service in the market, that may not end up helping Microsoft all that much. Nevertheless, Meta will likely continue to develop future iterations of Llama as long as it can get enough licensing fees from the anti-Google alliance it is forming with Microsoft and others.
