Google Just Showed Off a Powerful New Update to Gemini
If it seems like Google is bombarding us with Gemini ads, that’s because it is . Last week, the company renamed its Bard AI bot Gemini and introduced Gemini Advanced, its first paid AI subscription tier. Following this announcement comes Gemini 1.5 , the next iteration of Google’s artificial intelligence model.
What is Gemini 1.5?
Google says Gemini 1.5 is built on the Transformer and Mixture-of-Experts (MoE) architecture. Compared to traditional Transformer architecture, which is essentially one big neural network, Google says MoE models can be divided into multiple “expert” neural networks. This way, the model only activates the paths that are needed for a given task, which Google says is much more efficient than the standard feature.
The first version of Gemini 1.5 available to testers is Gemini 1.5 Pro. “Pro” is Google’s name for LLMs (Large Language Models) that work best with the widest range of tasks and devices. Think of it as GPT-3.5, a model designed to be used in the most possible applications, not the most powerful ones.
While 1.5 Pro has a standard context window for 128,000 tokens ( the same as GPT-4 ), Google says it is currently testing a context window for up to one million tokens, as well as 10 million , up from 32,000 in 1.0 Pro. The more tokens the model can run, the larger your prompt can be and therefore the larger the request the model can handle at once. According to Google, 1.5 Pro can handle up to one hour of video, 11 hours of audio, and codebases with more than 30,000 lines of code (or more than 700,000 words). (It can also handle single queries with over 100,000 lines of code.)
The 1.5 Pro is reported to be better at understanding large amounts of information and can answer complex and specific questions. In the demo video, Google shows an example of loading 402-page transcripts from the Apollo 11 mission into 1.5 Pro. He then shared a sketch of a walking shoe with an arrow pointing to the ground and asked the model what the image was about. The model was able to determine that this sketch depicted the moment when Neil Armstrong set foot on the moon and said his famous quote, all according to the drawing. It can also analyze content like a silent film, describing what happened and highlighting small moments that most viewers might miss.
Google says 1.5 Pro beats 1.0 Pro in 87% of the benchmark tests it uses. It also performs “on the same level” as the 1.0 Ultra, Gemini Advanced model. The company is also pleased with its “contextual learning,” in which the model can take information from the current data set and apply it in new ways without needing additional prompts.
How to try Gemini 1.5
Google is running trials for Gemini 1.5 Pro through AI Studio and Vertex AI and has a waitlist for interested developers. If you are a really interested developer, you can sign up for the waitlist here .
Once Gemini 1.5 becomes widely available through Google’s Gemini services, you will be able to use a window of 125,000 tokens for free. If you want access to the million token window, it will cost you money. (Price unknown at this time.)