Google’s Gemini may not help it reclaim its lost 'AI crown'

Gemini is Latin for twin, and so it is an apt name. It also hints at Google’s struggle to create a Large Language Model (LLM) rivalling OpenAI’s GPT4. It sometimes surprises people that Google ‘invented’ generative AI and LLMs—more precisely the Transformer architecture which underpins this technology. The concept behind Transformers was revealed in a 2017 paper, ‘Attention is all you need,’ written largely by researchers at Google Brain. Google also owns the deep learning powerhouse DeepMind, making it the leader in Artificial Intelligence (AI). But it chose not to take the transformer discovery forward, and so it was OpenAI which picked up the ball and carried it to the touchline. The reason could be the Innovator’s Dilemma; Google feared cannibalization of its Search business or a reputational risk from LLMs. It also could be fraternal rivalry between the two mighty AI twins: Jeff Dean’s Google Brain and Demis Hassabis’ DeepMind. Eventually, Google merged the two AI units, made Hassabis the leader, and set out to build Gemini.

The project has emerged to be the single largest effort by Google in recent times, aiming to take back the AI crown from OpenAI and other assorted competitors. Not to mention arch-competitor Microsoft, which is delivering enterprise products at lightning speed and scale, leveraging its tight relationship with OpenAI. Google’s earlier release Bard underwhelmed critics and left it with egg on its face. So, Google has put everything it has into Gemini, making it its largest product effort in recent times.

Launch day demos and announcements were impressive. Gemini is multimodal from the word go, working with pictures, videos, text and voice seamlessly. Its nano version is built for mobile use and can work on-device without depending too much on the cloud. Google announced its integration in the next version of its Pixel mobile phone. Gemini Pro is already integrated into Bard. Gemini also might have reasoning and planning capabilities, which could be built into advanced personal assistants. The Ultra version generated some buzz, with Google claiming that it beats GPT4 in some benchmark tests and was better at generating computer code and summarizing news articles. Some of this was debunked the next day itself and Google admitted that some parts of its launch video were ‘staged’ (some of the demos were pre-recorded). As it had used a few atypical measures to claim superior performance over GPT4, its boast was left unverified.

Finally, Gemini Ultra will only be available next year, by which time GPT4.5 or 5 might be out, moving the goalposts further for Google yet again. However, I wouldn’t write Google off. It has enormous AI strengths—a phalanx of top talent, computers and money across its merged entities, plus its web reach through Chrome, Search and YouTube. The timing is great, given the internal turmoil at OpenAI and apprehensions on retaining its key talent.

Another aspect was how the Gemini announcement seemed like one more BigTech launch. As every tech company rushes to launch its own LLM, Matteo Wong of the Atlantic calls it “Generative AI’s iPhone moment," where “generative AI is more about competition than revolution." It seems like a déjà vu moment where, as Wong says, it is an open question “whether the generative-AI race prompts genuine societal transformation or simply provides a new profit model" for Big Tech firms (bit.ly/41sHyED).

While Gemini’s Latin origin is ‘twin,’ it has a Babylonian history too. In their astronomy, the stars Pollux and Castor were known as the Great Twins. Their Babylonian names meant ‘Mighty King’ and ‘The One arisen from the Underworld.’ With Gemini, GPT4 and other GenAI models, it remains to be seen what they will be: mighty kings of technology working for the benefit of society, or the destroyer of worlds.