Google strives to catch up with OpenAI through its new Gemini triplet models

By Juha Saarinen

The competition for the artificial intelligence crown is heating up, with Alphabet-owned Google releasing the Gemini deep learning model.

Gemini is aimed squarely against Microsoft-backed OpenAI’s Generative Pretrained Transformer (GPT), which caught Google on the backfoot when it was released to the public through the conversational ChatGPT service.

ChatGPT took the world by storm and became the fastest-ever growing platform, reaching over 100 million users in just two months. Microsoft has seized the AI opportunity quickly, and has invested several billions of dollars in OpenAI since 2019.

The investment is partly through providing compute capacity on Microsoft’s Azure cloud platform. Microsoft is looking to recoup its investment through integrating ChatGPT and OpenAI’s image generating model DALL-E with its office productivity software, and its Bing search engine.

Enterprises worldwide that already use Microsoft applications have flocked to Copilot trials and interest in AI is running hot currently.

This week’s Gemini launch points to Google taking several leaves out of Microsoft’s AI book especially when it comes to targeting enterprise customers.

A careful read of Google’s announcement however, suggests the Internet giant still has a long way to go before hauling in OpenAI’s lead.

Google falling behind in the AI race

With OpenAI hogging the headlines, most recently over serious internal ructions that saw chief executive Sam Altman being fired and rehired last month, Google has struggled to come up with an effective response to ChatGPT.

It finally launched its own publicly available conversational AI, Bard, in February this year. Bard did not impress the world, to put it mildly. After a live streamed YouTube launch in February, Google’s share price dropped by some 8%, shaving around US$100 billion off the Internet giant’s market capitalisation.

Since its launch, Bard has had its underlying technology, the language models with massive amounts of digitised source material that the AI processes for its responses. It has now been upgraded to Gemini, Google said.

What is Gemini?

Despite the name, Gemini is in fact a triplet of AI models. The smallest, Gemini Nano, is aimed at mobile devices, whereas the Gemini Pro is said to be Google’s “best model for scaling across a wide range of tasks.”

Gemini Ultra runs in Google’s massive data centres, and it is the model that the Internet company says beats GPT-4 in 30 out of 32 academic benchmarks used for large language model research and development.

Like the subscription-based GPT-4, Gemini is multimodal. This means that unlike earlier GenAI LLMs, the AI can process more than just text. It can handle images, audio and video input as well.

Google also said Gemini, and the largest Ultra model in particular, are able to generate high-quality computer code in popular programming languages.

Here’s the rub: a direct comparison between Gemini Ultra and GPT-4 is not yet possible.

The Ultra and Nano models are not yet available, Google said. Ultra is said to become available for select customers next year.

For Gemini Pro, Google makes the more modest claim of beating the current free version 3.5 of OpenAI’s GPT in six out of eight benchmarks. This suggests Gemini Pro is someway behind GPT-4. which is available for a monthly subscription of US$20 a month, and which includes the DALL-E image generating AI as well.

Next year, Google intends to launch Bard Advanced. This will provide access to Gemini Ultra, but Google did not say when the launch would take place, or whether access to Ultra would be free or cost money.

Google is also bringing Gemini to its Pixel 8 Pro smartphone, which is not for sale in New Zealand, and integrating the AI into its products such as Search, Ads and Chrome web browser.

Its Duet AI assistant will also be equipped with Gemini, Google said. This strategy is similar to Microsoft.

GenAIs are infamous for occasionally producing completely made-up output, which the industry calls “hallucinating”. Amazon Web Services launched its Microsoft Copilot AI competitor, Q, in November. Soon after, documents from AWS staffers were leaked to The Platformer site, detailing how Q hallucinates badly and leaks sensitive information provided to it.

Google said it has sought to address the content safety issue by running the AI through comprehensive evaluations. This includes testing for bias, toxic responses, and using adversarial techniques to ferret out safety issues in the AI.

Users who currently want to try out Gemini can access the mid-tier Pro version through Google Bard which interest.co.nz used to generate the output for this article. It is available in 170 economies, including New Zealand.