It’s been widely believed that large companies with proprietary models will come to dominate the AI space, but a leaked document purportedly from a Google researcher says that that might not be the case.
Google and OpenAI have no moats in the AI field, and will likely be outstripped by open source alternatives, a Google researcher reportedly claimed in an internal document that was subsequently leaked. The document says that while OpenAI and Google spent large amounts of time and large sums to money to come up with state of the art models, cheaper open source alternatives were catching up very quickly. These models, by virtue of being cheaper, customizable and quicker to release, will likely see greater adoption than large customized models created by companies like OpenAI and Google.
The remarkably well-thought out document says that the leak of Facebook’s LLaMA model set of a chain reaction of events that made LLMs widely accessible and simple to train. LLaMA’s weights were leaked within a week of its release, and the open source community pounced on them, tinkering with the model, improving it, and adding more functionality. The advances that the community made were rapid. Within weeks of LLaMA’s leak, researchers were able to iterate and create models that were nearly as good as Google’s Bard, and within touching distance of OpenAI’s ChatGPT.
The document says that like with image generation, text generation could also likely end up being dominated by the open source community. OpenAI had released their image generation model named Dall-E, but it now remains largely forgotten in comparison with open source alternatives like Stable Diffusion. The document says that something similar could happen with the text generation space.
It’s an interesting claim, but it’s important to note that this isn’t the opinion of Google, but that of a single employee within the company. It’s still a compelling argument that could predict how the AI space will shape up in the coming years. Here’s the document in its entirety.