Google’s Lost Opportunity: How the Company’s Own Invention Fueled the AI Revolution
It seems fitting that one of Google’s most important inventions — one that would come back to haunt the company — was initially devised over lunch. Back in 2017, researchers at Alphabet’s Mountain View, California, headquarters were discussing ways to make computers generate text more efficiently. Within five months, they had conducted experiments and, unaware of the groundbreaking nature of their discovery, documented their findings in a research paper titled “Attention is All You Need.” This paper would fundamentally change the landscape of AI.
The paper’s eight authors had developed the Transformer, a system capable of enabling machines to generate humanlike text, images, DNA sequences, and various other forms of data with unprecedented efficiency. This research paper would go on to be cited over 80,000 times by other researchers, and the AI architecture they designed would lay the foundation for technologies like OpenAI’s ChatGPT (where the "T" stands for Transformer), image-generating tools like Midjourney, and more.
While it was standard practice for tech companies like Google to open-source new techniques to gather feedback, attract talent, and foster a community of supporters, Google itself didn’t immediately adopt this groundbreaking technology. The system remained dormant for years as the company struggled to translate its cutting-edge research into usable services. Meanwhile, OpenAI leveraged Google’s innovation to launch the most formidable threat to the search giant in recent history. Despite Google’s cultivated talent and innovation, rival firms were the ones who capitalized on its significant discovery.
The researchers who co-authored the 2017 paper didn’t anticipate a long-term future at Google either. In fact, all of them have since left the company. Some have moved on to start their own ventures, including Cohere, a company developing enterprise software, and Character.ai, founded by Noam Shazeer, the longest-serving Googler in the group, who was considered an AI legend within the company. Collectively, these startups are currently valued at approximately $4.1 billion (roughly Rs. 33,640 crore), according to valuations from research firm Pitchbook and price-tracking site CoinMarketCap. They are now considered AI royalty in Silicon Valley.
The last of the eight authors to remain at Google, Llion Jones, announced this week that he was leaving to pursue his own entrepreneurial endeavor. He described the experience of watching the technology he co-created take the world by storm over the past year as surreal. “It’s only recently that I’ve felt … famous?” Jones says. “No one knows my face or my name, but it takes five seconds to explain: ‘I was on the team that created the ‘T’ in ChatGPT.’”
It seems strange that Jones achieved celebrity status due to actions outside of Google. Where did the company go wrong?
Google’s Size and Internal Roadblocks
One obvious factor is scale. Google boasts an army of 7,133 AI employees, representing a significant portion of its overall workforce of approximately 140,000, according to Glass.ai, an AI firm that conducted a LinkedIn profile analysis to identify AI employees at Big Tech firms earlier this year for Bloomberg Opinion. In comparison, OpenAI, which ignited an AI arms race, achieved this with a much smaller team — about 150 AI researchers out of a total staff of around 375 in 2023.
Google’s immense size created a complex bureaucratic process where scientists and engineers had to navigate multiple layers of management to gain approval for their ideas, particularly during the Transformer’s development, according to several former scientists and engineers. Researchers at Google Brain, one of the company’s primary AI divisions, also lacked a clear strategic direction, leading many to prioritize career advancement and their visibility on research publications.
Furthermore, the bar for transforming ideas into new products was exceptionally high. “Google doesn’t move unless [an idea is] a billion-dollar business,” remarks Illia Polosukhin, who was 25 when he first collaborated with fellow researchers Ashish Vaswani and Jakob Uszkoreit at the Google canteen. However, building a billion-dollar business requires constant iteration and an acceptance of numerous dead ends, something that Google did not always tolerate.
A Victim of Its Own Success
In a way, Google became a victim of its own success. It boasted renowned AI scientists like Geoffrey Hinton, and in 2017, was already using cutting-edge AI techniques to process text. The prevailing mindset among many researchers was “If it ain’t broke, don’t fix it.”
However, this is where the Transformer authors had an edge: Polosukhin was preparing to leave Google and was more willing than most to take risks (he has since founded a blockchain company). Vaswani, destined to become the lead author of their paper, was eager to tackle a large-scale project (he and Niki Parmar went on to establish the enterprise software firm Essential.ai). And Uszkoreit had a penchant for challenging the status quo in AI research — his motto was, if it ain’t broke, break it (he has since co-founded a biotechnology company called Inceptive Nucleics).
Breaking the Mold: “Attention” and the Transformer
In 2016, Uszkoreit had explored the concept of "attention" in AI, where a computer identifies the most crucial information within a dataset. A year later, over lunch, the trio discussed applying this principle to translate words more efficiently. Google Translate at the time was cumbersome, particularly when dealing with non-Latin languages. “Chinese to Russian was terrible,” Polosukhin recalls.
The issue was that recurrent neural networks processed words sequentially, a process that was slow and didn’t fully utilize chips capable of processing numerous tasks simultaneously. The CPU in your home computer probably has four "cores," which process and execute instructions, but those used in servers for AI processing have thousands of cores. This translates to an AI model having the ability to "read" multiple words in a sentence concurrently. However, no one had fully leveraged this capability.
Overcoming Skepticism and Achieving Breakthroughs
Uszkoreit would walk around the Google office sketching diagrams of the new architecture on whiteboards, often met with incredulity. His team aimed to eliminate the "recurrent" aspect of the recurrent neural networks used at the time, which “sounded mad,” according to Jones. But as a few other researchers, including Parmar, Aidan Gomez, and Lukasz Kaiser, joined the group, they began observing improvements.
Consider this example: In the sentence, “The animal didn’t cross the street because it was too tired,” the word "it" refers to the animal. However, an AI system would struggle if the sentence changed to, “because it was too wide,” as "it" would become more ambiguous. However, the new system was able to overcome this challenge. Jones remembers witnessing this process. “I thought, ‘This is special,’” he says.
Uszkoreit, fluent in German, also noticed that the new technique could translate English into German far more accurately than Google Translate had ever achieved.
A Missed Opportunity: Google’s Hesitation
Despite this success, Google took a considerable amount of time to apply the technique to its free translation tool or to its language model BERT, and the company never deployed it in a chatbot that anyone could access. It wasn’t until the launch of ChatGPT in late 2022 that Google felt compelled to swiftly release a competitor called Bard in March 2023.
Over the years, the authors observed their ideas being applied to a diverse range of tasks by others, from OpenAI’s early versions of ChatGPT to DALL-E, and from Midjourney’s image tool to DeepMind’s protein-folding system AlphaFold. It was undeniable that the most exciting advancements were occurring outside of Mountain View.
While one could argue that Google has been cautious in deploying AI services, slowness doesn’t always equate to prudence. It can also signify inertia and bureaucratic bloat. Today, some of the most innovative AI breakthroughs are emerging from small, agile startups. It’s unfortunate that many of these startups will be swallowed by large tech behemoths, who are poised to reap the most significant financial rewards in the AI race despite playing catch-up.
Google may ultimately have the last laugh, but its journey will likely be unremarkable in many ways.
© 2023 Bloomberg LP
