[Tech Talk] Why Did Google Say It Has No 'AI Moat'?

Google, OpenAI, and Others' 'Closed AI' Strategy
Hiding Training Methods to Gain Technological Edge
Open Source Rises in Response
Community May Overwhelm Big Tech

"We do not have a moat, and OpenAI does not either. The uncomfortable truth is that the company that will win the AI competition is neither us nor OpenAI. A third force is emerging, and that name is open source."

This is the content of an internal Google message leaked by the US semiconductor analysis group 'Semianalysis' on the 4th (local time). The term 'moat' refers to decisive technological capabilities that maintain a company's super-gap, and Google AI engineers have essentially admitted that they do not hold the future leadership of the AI industry.

The message mentioned 'open source' as the third force that will surpass Google and OpenAI, the developer of ChatGPT. Open source is a strategy that freely opens all information related to AI models to users, encouraging them to develop freely. So why do Google developers regard open source AI with caution?

Training know-how is AI technology... Big Tech hides trade secrets

Sam Altman, CEO of OpenAI
Photo by EPA Yonhap News

The work of creating and running neural network AI is broadly divided into 'training' and 'inference.' Training is the process of training AI with vast datasets. Inference, on the other hand, is performing various tasks based on the trained AI.

Famous AI companies such as OpenAI, Google, and Google's AI research company 'DeepMind' possess top-tier technology in both training and inference fields. Among these, training is what determines an AI company's technological prowess.

Simply creating an AI with a massive neural network is not very meaningful. The performance of AI varies greatly depending on how well it is 'trained.'

This is where the 'skill' of AI companies diverges. In fact, AI training is closer to craftsmanship than science. How to compose the dataset for training AI, how many times to repeat training (called 'epochs'), and other factors all become variables that determine AI's performance after training ends.

No matter how large the AI model is, if the quality of the data is low, it will produce poor results, and training too little or even too much can damage the AI. Various other factors in the training process also affect the AI's 'growth.' The way to create good AI is to proceed step by step through multiple trials and errors and analyze performance changes.

Therefore, current AI companies are extremely reluctant to 'disclose' their models externally. This is because all the trials and errors they have experienced during AI training are their own know-how and technological strength. OpenAI, which started as a non-profit AI research organization, has recently shifted to a 'closed strategy.' This is likely a move to prevent competing companies from copying their know-how.

Open source 'open strategy' challenges Big Tech's closed strategy

The image generation AI 'Stable Diffusion' developed by Stability AI is a leading example of an open-source AI model. [Image source=Stable Diffusion]

Ironically, this 'closed strategy' risks being overtaken by another AI that advocates an extreme 'open strategy.' Open source, which openly discloses everything from model components and detailed architecture to training methods from the start, is a representative example.

Meta AI, which was considered somewhat behind Google and OpenAI, succeeded in a rapid chase with an open source strategy. By opening the large language generation model Llama as open source, it showed quick progress. Also, the famous UK AI startup 'StabilityAI' is improving model performance by advocating an open source strategy.

The strength of open source lies in forming a 'user community.' Users will modify or improve open source models according to their needs and share their improvements with others. This method can accelerate the performance improvement of open source AI.

The unique 'culture' of the AI industry may also fuel the development of open source. In an interview with the British weekly 'The Economist' in January, David Ha, an AI researcher from Google Brain, pointed out, "The reason why none of the big tech companies currently claim to be the leader in AI technology is because of the lab research culture," adding, "All machine learning researchers hang out together."

This means that AI researchers from big tech to startups already have a culture of secretly sharing tips and know-how with each other. This is because the AI industry itself is still in its early stages, and many talents are employed alternately by various companies.

Major AI companies such as DeepMind, Google, and Meta are gathered at King's Cross in London, UK.
[Image source=Yonhap News]

Currently, in global AI industry hubs such as San Francisco in the US and London in the UK, most AI companies are concentrated within a single city. Famous researchers often pass through Google, Meta, Microsoft, OpenAI, etc., before leaving to start their own startups. In such a situation, 'corporate secrets' cannot be properly maintained.

However, big tech companies have vast researcher networks and the latest supercomputer infrastructure that open source lacks. If AI development depends on 'how much trial and error is done,' big tech has sufficient advantages. Ultimately, the competition will be between the massive user community and the enormous capital power of big tech. This is why Google has admitted that it still does not have a 'moat.'