Google Launches Gemini Next-Generation Large Language Model

Google has finally unveiled its fabled next-generation large language model, Gemini, claiming it’s the largest and “most capable” model the company has ever built.

After months of silence, the company dropped a blog post showcasing the model, revealing Gemini comes in three sizes - Ultra, Pro and Nano, enabling it to run on mobile devices up to data centers.

Google announced that Gemini will start powering solutions today – with its Bard chatbot powered by Gemini Pro, the initial version of the model, but only in English. The company claims Gemini Pro will improve Bard’s reasoning and understanding abilities.

In early 2024, Google plans to launch Bard Advanced. No further details were given, but Google said it would provide access to its “advanced models and capabilities — starting with Gemini Ultra.”

Google’s new Pixel 8 Pro handset will be the first mobile device capable of running Gemini Nano. Starting next year, it will have access to new features including Summarize in the Recorder app and a developer preview of Smart Reply in Gboard coming first to WhatsApp.

Gemini is also coming to Google Ads, Chrome and Duet AI. In the coming months, Gemini will power features in more of its products and services like Ads, Chrome and Duet AI. With more hints at what’s to come, Google said it’s been experimenting with the model in search, with Gemini making its Search Generative Experience (SGE) faster.

Related:Google Delays ‘Gemini’ Launch to 2024

No reference to Gemini powering Makersuite features appeared in the blog, however. A leak back in October showed that Google’s web development platform was set to benefit from Gemini’s generative abilities.

Google CEO Sundar Pichai first revealing Gemini at I/O 2023

Access Google Gemini

Starting Dec. 13, developers and enterprise customers can access Gemini Pro via the Gemini API in Vertex AI or Google AI Studio, its free web-based developer tool.

Gemini Ultra will be made available to select groups before being given to enterprise customers in early 2024.

Android developers wanting to build Gemini-powered apps on-device can sign up for an early preview of Nano via Android AICore.

Move Over ChatGPT

Gemini was first announced at the company’s I/O event back in May but has been kept largely under wraps ever since. It’s among the so-called “frontier models,” foundational large language models that were the focal point of the recent AI Safety Summit.

Gemini was meant to be unveiled at private company events this week, however, these were scrapped by Google CEO Sundar Pichai over the model’s poor handling of some non-English queries. Private demonstrations for unnamed business partners were given, however.

Related:DeepMind's New AI can Learn Tasks Directly from Humans

Senior leaders at Google reportedly want Gemini to compete with OpenAI’s ChatGPT, with Pichai keen for the initial version of the model to be released as soon as possible to compete.

Another concern for Google is that Microsoft's AI Copilot powering its productivity software is becoming a significant business. ChatGPT and other OpenAI tech are powering various Microsoft services, with Google keen to note fall behind.

New TPU Chips and Hypercomputer

Google said it’s been training the large language model on its in-house TPU chips. Alongside the Gemini announcement, the company also announced new AI accelerators Cloud TPU v5p.

The new chips train large language models 2.8 times faster than the previous generation TPU v4. 

The new chips are even more powerful than the Cloud TPU v5e, which Google only made generally available following its Next event back in August. Google Cloud said the latest line however is its “most powerful yet.”

In addition, AI Hypercomputer from Google Cloud, a new supercomputer architecture designed for improving AI training and tuning, was also unveiled.

Google Cloud argues that traditional computing infrastructures “are no longer adequate for the exponential demands of generative AI.”

Related:Google DeepMind's Six Levels of AGI

The new AI Hypercomputer works as an integrated system of AI-optimized hardware, software, ML frameworks and consumption models. Powered by the new TPU v5ps, it’s designed to train large language models, with the combination of hardware and software powering an integrated and secure computing system.

“At Google, we’ve done decades of research and development on this very problem, culminating in AI Hypercomputer, a system of technologies optimized to work in concert to enable modern AI workloads,” a company blog post reads.

To access either Cloud TPU v5p or the new AI Hypercomputer, enterprises will have to contact their respective Google Cloud account manager.