Google pins its hopes on Gemini to leapfrog GPT-4

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

It has taken a year, but Google has finally delivered a coherent response to the surprise challenge to its dominance in artificial intelligence that came with the launch of ChatGPT.

This week’s release of Gemini, a family of large language models, will give it a stronger platform to fight back against both OpenAI, the company behind ChatGPT, and Microsoft, which has used OpenAI’s models to supercharge all its software and cloud services this year.  

The question now is whether Gemini can make a meaningful difference to Google’s existing services — and, perhaps even more important, whether it can become a foundation for a new range of services that carry AI much deeper into everyday life.

With the three “flavours” of Gemini announced this week, Google is finally stamping its mark on a technology that its own researchers did much to pioneer, but which OpenAI’s ChatGPT carried into the mainstream. The Pro version, for instance, is positioned squarely against OpenAI’s GPT-3.5, the model behind the free version of ChatGPT and the workhorse for many of the first generative AI applications from other companies that have hit the market this year.

The smaller Gemini Nano is matched against systems such as the smallest version of LLaMa 2, Facebook’s open-source model, making it capable of being run on a mobile device. Apple, as always, is taking a considered approach before bringing generative AI to the iPhone, but the appearance of Gemini on Google’s latest Pixel handset is a sign that it can’t afford to wait too long.

It is the top-of-the-line Gemini Ultra, due out early next year, that carries Google’s main hopes of matching or leapfrogging OpenAI’s GPT-4 in the race to turn generative AI into a more useful everyday tool. The company fell behind this year, but has some clear advantages that could help bring Gemini to a big market in 2024.

One is distribution. Google said this week, for instance, that Gemini will be added to Chrome, which has more than 60 per cent of the browser market, giving billions of web users instant access to tools that are able to do things such as analyse the content of web pages.

As Google flexes its existing market power like this to boost its AI ambitions, competition regulators will be watching closely.

Another advantage for Google is the uncertainty around OpenAI. After the shock sacking and reinstatement of chief executive Sam Altman last month, the many businesses that have built their own generative AI plans on top of OpenAI’s models will be looking to hedge their bets.

The search company will also be hoping that its Bard chatbot will do a better job of rivalling ChatGPT now that it has a better language model behind it. But its best hope of regaining an edge may lie in being the first to come up with the next breakthrough services powered by generative AI. Some of the capabilities claimed for Gemini point to where Google thinks these might lie.

It has made much, for instance, of the fact that Gemini was designed from the outset to be “multimodal” — that is, able to understand not just text but also images, video and audio. According to Google, that makes it better suited than models such as GPT-4 to deal with the sort of everyday situations that rely on senses such as sight and hearing.

That may be a step towards AI systems that are better able to operate in the real world. But it is too soon to tell what applications this could make possible, or whether Google really has achieved the technical superiority it claims.

Another avenue for development lies in what Google claims are Gemini’s reasoning and planning capabilities. These are the kind of skills that could prepare the ground for personal assistants able to tackle complex problems and set a plan of action.

If such assistants are linked to other internet services, they could also become agents, taking action on their users’ behalf. Imagine a shopping agent, for instance, that not only hunts out the products you want but goes ahead and pays for them as well.

This is already shaping up to be one of the key AI battles of 2024 and beyond. OpenAI took a first step in this direction last month when it said its users would be able to build rudimentary agents on top of its models, then offer them for sale on an OpenAI app store. That could point to the next big AI breakthrough beyond ChatGPT — and this time, Google has no intention of being left behind.

richard.waters@ft.com