Google launches Gemini 1.5 generative AI model: what it can do

Google Gemini is only about two months old, but the search giant has already launched its successor, its latest large language model (LLM) to date – Gemini 1.5. This version is currently only available to enterprises and developers, with a general rollout to consumers expected soon.

Google CEO Sundar Pichai said that Gemini 1.5 demonstrated “significant improvements in multiple dimensions” while using less computing to achieve the same results as Gemini 1.0 Ultra, the most advanced legal system. Master’s degree) of comparable quality.

In addition, Pichai added that the new generation has achieved breakthroughs in long-context understanding, and now Google has been able to “increase the amount of information our models can process – continuously running up to 1 million tokens for the longest The context window doesn’t have any large base models yet.”

Gemini 1.5: What’s new?

First, the Gemini 1.5 model is equipped with a new Mixture-of-Experts (MoE) architecture, making it more efficient and easier to service.

Initially, Google released the 1.5 Pro version for early testing, which performed similarly to 1.0 Ultra. Gemini 1.5 Pro will offer a standard 128,000-token context window, but a small number of developers and enterprises can experiment with context windows of up to 1 million tokens.

Google also emphasized that because it is built on the Transformer and MoE architecture, Gemini 1.5 is more efficient because the MoE model is divided into smaller “expert” neural networks.

Additionally, understanding context is another key area of ​​improvement in Gemini 1.5. “The 1.5 Pro can handle large amounts of information at once, including 1 hour of video, 11 hours of audio, a codebase of over 30,000 lines of code, or over 7,00,000 words. In our research, we also successfully tested up to 1000 10,000 tokens,” Google said.

In short, if you give Gemini 1.5 a large amount of information, such as a large novel, a research article, or, as Google mentioned, the 402 pages of Apollo 11 mission records, and ask it to summarize, it can make it happen. You can then also ask detailed questions based on what Gemini 1.5 understands from it.

Performance improves by leaps and bounds

Compared to the Gemini 1.0 Pro, the new 1.5 Pro model outperformed the Gemini 1.0 Pro in 87% of the benchmark tests used to develop Google LLM, and performed similarly to the Gemini 1.0 Ultra.

Another big change is the ability to demonstrate “situated learning” skills. This means the Gemini 1.5 Pro can “learn new skills from information given in long prompts without the need for additional fine-tuning.”