Google’s Star AI Model, Gemini 1.5, Is Getting A Few Changes And Improvements
Gemini 1.5 came out in February, which wasn’t long after the model’s release about two months earlier. This was the first sign of how serious the company was taking the development of Gemini. With the Gemini model released in December, developers could also make use of the Google API on their platforms or in their applications.
At Google I/O yesterday, the company announced that they’ve made several code improvements to Gemini 1.5 Pro, which means that it should work a lot better when helping you translate, code, reason, etc. The best part is that these updates have rolled out right away.
Gemini is also getting some new developer-facing features, such as “video frame extraction” and “parallel function calling”. I’m no dev, but the first one seems rather self-explanatory and the second should likely improve efficiency by not needing to wait for one function to be called at a time. In June, the Gemini API is also expected to support “context caching”, which should make it remember the context of sessions a lot better.
Introducing Gemini 1.5 Flash
Part of the Gemini announcement included the launch of Gemini 1.5 Flash, which is a lightweight model of Gemini that is optimized for speed. According to Google, it is going to be best for “narrower and high-frequency tasks where the speed of the model’s response time matters the most”.
Gemini 1.5 Flash, just like Gemini 1.5 Pro, will be multimodal, which means that you’ll be able to use text, images, audio, and video as parts of your requests.
Google Gemma, The Open-Source AI Model, Is Having New Models Added
Now, you might not have heard much about Google Gemma as a consumer, but developers are more likely to have. Google Gemma is sort of like Gemini, but if you’re looking for an open-source AI model that you can actually dive into, modify, and bend to your will, then Gemma is a much better option for you.
The Gemma family got two new members yesterday. The star of the show is Gemma 2, which is expected to be released in June. Google says that it does better work than models twice as big as it, and it’ll work optimally on GPUs and TPUs.
For specific use cases, Google also has certain pre-trained versions of Gemma, such as CodeGemma and RecurrentGemma. However, now they’ve added PaliGemma to the mix, which is a vision-language model that can help you identify and describe photos, making it perfect for image captioning and labeling.
