On May 4, 2023, bloggers shared an alleged leaked memo from a Google staff member discussing their AI strategy. Although Google hasn't confirmed it, the document appears authentic. Here's a summary and its impact on your generative AI strategy.
The author contends that both Google and OpenAI will lose the large language model race to the open-source community. They argue that Google should follow the successful strategy employed with Android and Chrome by supporting the open-source community with its full weight and engineering resources.
The article delves into why and how the open-source community will outpace both Google and OpenAI. This opinion is debatable, as GPT-4 remains far superior to any open-source alternatives. However, the author points out that open-source alternatives are improving rapidly. A similar situation occurred in the generative image space with OpenAI's DALL-E being quickly matched by open-source solutions like Stability AI.
The same scenario is predicted to unfold with large language models. With hundreds of thousands of engineers working on new approaches, it becomes difficult for a single company to keep up. For instance, it was discovered that smaller models trained on better data can perform just as well as larger models trained on poor data in some scenarios. This revelation has substantial economic implications since training a massive language model can cost tens of millions, while training a smaller model may only require $100. The author also argues that the open-source community will execute faster by making incremental improvements to models instead of retraining a new one from scratch.
The open-source community is also free to work without institutional constraints. Google and OpenAI must answer to the media, the public, and potentially lawmakers about their proprietary technology development and deployment. However, the guy living in his mom’s basement and eating Pringles all day doesn’t have these constraints. He can create and release projects without such restrictions, focusing on solving problems without worrying about social consequences.
These points are debatable, and it's not always true that open-source software wins. For example, Linux, a free open-source operating system, never gained traction in the market. However, the open-source community has a key advantage: large tech companies willing to support open-source large language models for their economic benefit. Meta is a prime example, as Mark Zuckerberg has expressed that open-source technologies benefit his company, which profits from advertising. Better large language models, regardless of their origin, offer an economic advantage in ad sales.
The same applies to Apple, Netflix, Spotify, and thousands of other companies. So, even if the open-source community requires substantial capital investment to compete with OpenAI/Microsoft, it's likely that one or more of these companies will be eager to provide support.
So what should you do?
Your generative AI strategy should prepare for a future with multiple large language models in your ecosystem. We have been advising our clients to design initial applications that can easily swap models. Unfortunately many businesses are building their first systems without considering the need for such architectural flexibility. For example, a highly optimized prompt engineering solution may not work with a different or even improved model from the same company. Similarly, the embeddings generated by GPT-4 may not be compatible with a different large language model, and your architecture should account for the need to regenerate them as technology advances.
In summary, it's essential to view large language models not as individual products, but as a collection of software functions. You'll have many operating in your environment, all driving significant automation at scale.