Google’s $4 Billion AI Superstars: The Researchers Who Created the ‘T’ in ChatGPT
In a twist of fate, one of Google’s most significant inventions was born from a casual lunchtime conversation. In 2017, a group of researchers at Alphabet’s Mountain View headquarters discussed how to improve computers’ text generation efficiency. Little did they know that their experiment would lead to a groundbreaking AI advancement. In a research paper titled Attention is All You Need, the team introduced the Transformer, a system that revolutionized the generation of human-like text, images, and other forms of data. OpenAI later adopted this technology to develop ChatGPT and other tools.
Surprisingly, Google didn’t immediately utilize the new technology. It remained dormant as the company focused on converting their cutting-edge research into practical services. In the meantime, OpenAI capitalized on Google’s invention, posing a significant threat to the search giant. Even the researchers who co-authored the 2017 paper left Google and went on to establish their own ventures, including Cohere, which specializes in enterprise software, and Character.ai, founded by Noam Shazeer, an AI legend at Google. Today, their combined businesses are valued at approximately $4.1 billion.
The last remaining author at Google, Llion Jones, recently confirmed his departure to start his own company. Witnessing the success of the technology he helped create has been surreal for him. Although he may not be a household name, Jones takes pride in being part of the team responsible for the T in ChatGPT.
So where did Google go wrong? One key issue is scale. With over 7,100 people working on AI out of a total workforce of around 140,000, Google faced challenges in decision-making and strategic direction. Researchers often had to navigate multiple layers of management for approval, hindering progress. Furthermore, Google Brain, a prominent AI division, lacked a clear vision, leaving researchers focused on personal goals rather than collaborative advancement.
Another hurdle Google faced was its high bar for turning ideas into viable products. Only billion-dollar business prospects received sufficient attention, neglecting potential opportunities with iterative processes. In contrast, the Transformer authors dared to take risks and challenge the status quo. Their innovative thinking and willingness to venture into uncharted territories paved the way for their breakthrough.
The team recognized the limitations of the existing approach, which processed words sequentially, and harnessed the power of parallel processing. By removing the recurrent aspect of the neural networks used at the time, they unlocked the ability to process multiple words simultaneously. This new architecture proved superior when translating complex sentences and improved Google Translate’s accuracy. However, Google took a considerable amount of time to incorporate this technique into its translation tool and language model, BERT.
Over the years, the authors watched their ideas flourish outside of Google. OpenAI integrated their techniques into ChatGPT, DALL-E, and Midjourney developed image tools. DeepMind even utilized their methodology for their protein folding system, AlphaFold. It became evident that the most exciting AI innovations were happening beyond Google’s walls.
While some argue that Google’s cautious approach to deploying AI services is responsible for the slower progress, it is crucial to differentiate between strategic prudence and unnecessary inertia caused by organizational size. Currently, the most thrilling advancements emerge from nimble startups, although they are often absorbed by large tech players. As the AI race unfolds, it is these giants who stand to benefit financially.
In the end, Google might have the last laugh, but its journey has been far from remarkable. It is a testament to the ever-increasing influence of small, agile startups in the AI landscape.
