Google LLC and OpenAI LP, the developer behind ChatGPT, are facing tough competition from open-source developers working in the field of generative artificial intelligence (AI), which could potentially outdo them, warned Google’s senior engineer Luke Sernau in a leaked document.
In early April, Sernau’s document titled ‘We Have No Moat and Neither Does OpenAI’ was circulated by Google staff among internal servers and was eventually leaked on Thursday by consulting firm SemiAnalysis. In it, Sernau says that Google and OpenAI have been pitted against each other for control of the lucrative AI market, but the real challenge could be in the form of an outsider – the open source community.
Generative AI enables users to generate new content from prompts supplied in plain language. ChatGPT, one of OpenAI’s earlier projects, drew a lot of attention from the public, eventually resulting in millions of users and scores of headlines about it.
Sernau says that OpenAI and Google do have some leading edge in terms of generating higher quality results with their large language models (LLMs) such as GPT-4 and Bard, though this gap has been narrowing quickly. Furthermore, open source models – which tend to be more lightweight, cost-effective and easier to deploy – are on pace to catch up and even surpass their larger competitors.
Due to its expensive and heavy nature, large language models are quite restrictive in terms of portability and require a large amount of computing power and time to generate model results. On the other hand, smaller models built with open source technologies can tackle similar tasks with higher accuracy, but require significantly fewer resources to train and deploy.
Popular open source projects such as Meta Platforms Inc.’s Large Language Model Meta AI (LLaMA) had been leaked to the public in March and since then, other makers quickly built upon the framework, creating small-scale models such as Alpaca-13B and Vicuna-13B that can produce comparable outcomes in terms of accuracy.
Sernau says that this has opened up a new world of possibilities as far as training and experimentation are concerned, reducing the cost of creating a model to a single user and their laptop. People have also grown wary against paying for expensive and restrictive models when smaller open source models can deliver the same results for free.
To keep up with this trend, Seranu suggested that Google should work together with the open source community, where it has seen the most success. Since the leak of LLaMA, Meta has gathered a huge community of builders and developers, and may rise as the leader if it can effectively incorporate the research generated by that community.