It may seem like Google is constantly telling us about Gemini, and that's because they are. The company changed the name of its AI bot, Bard, to Gemini last week and added its first paid AI subscription level, Gemini Advanced. Gemini 1.5, the next version of Google's AI model, was just released, right after that news.
What is Gemini 1.5?
Google says that Gemini 1.5 is based on a design called Mixed-of-Experts (MoE) and Transformer. Google says that MoE models can be broken up into several "expert" neural networks, while a standard Transformer design is just one big neural network. That way, Google says, the model will only turn on the routes that are needed for the job at hand. This is much more efficient than the standard function.Gemini 1.5 Pro is the first version of Gemini 1.5 that users will be able to get. Google calls its "Pro" LLMs (large language models) the ones that work best on the most jobs and devices. The GPT-3.5 model is an example of this. It is meant to be used in as many situations as possible, not just the most powerful ones.
The normal context window for 1.5 Pro is 128,000 tokens, which is the same as GPT-4. However, Google is trying a context window with up to one million tokens and 10 million tokens, compared to 32,000 tokens for 1.0 Pro. Your prompt can be bigger if the model can run more tokens. This means that the model can handle bigger requests at once. Google says that 1.5 Pro can handle codebases with over 30,000 lines of code (or over 700,000 words) and up to an hour of video at the same time. It can also deal with single questions that have over 100,000 lines of code.
It is said that 1.5 Pro is better at learning a lot of information and being able to answer detailed and complicated questions about it. Google shows how to feed the 402-page transcripts from the Apollo 11 trip to 1.5 Pro in a demo film. On the next page, it showed a sketch of a boot walking with an arrow pointing down. It asked the model what the picture was about. From the picture, the model was able to tell that it showed the moment Neil Armstrong stepped foot on the moon and said his famous quote. It can also look at material like a silent movie, explain what happened, and point out small details that most people might miss.
Google says that 1.5 Pro does better than 1.0 Pro on 87% of the tests it runs. It also works "at a broadly similar level" to 1.0 Ultra, which is what Gemini Advanced is powered by. People at the company are also pleased with its "in-context learning," which lets the model use what it already knows in new ways without any extra help.
How to get Gemini 1.5
Google is testing Gemini 1.5 Pro with AI Studio and Vertex AI and has a list of developers who want to be on the waitlist. Interested developers can put their names on the waitlist here.When Google's Gemini services make Gemini 1.5 available to more people, you'll be able to use its 125,000 token window for free. It costs money to get into the one million token spot. It's not known what the price is yet.
0 Comments