Google Gemini 1.5: Flash 5x Faster than GPT-4 for Olympiad Math
3 Major Upgrades in Google Gemini 1.5: Speed, Efficiency, and Multi-Modal Mastery
Google announced that Gemini 1.5 is a generational leap over Claude 3.0 and GPT-4 Turbo.
In February, Google launched the multi-modal model Gemini 1.5. It improved performance and speed through engineering and infrastructure optimizations, and MoE architecture. It offers longer context, stronger reasoning, and better handling of multi-modal content.
This Friday, Google DeepMind released the 153-page technical report on Gemini 1.5, including updates on the Flash version.
In this report, Google introduces the Gemini 1.5 series. These next-gen, high-efficiency models can handle detailed info from millions of tokens. They can process long documents and hours of video.
The series includes two new models:
1. The updated Gemini 1.5 Pro has better features and benchmarks than the February version.
2. The Gemini 1.5 Flash is a lighter version, designed for efficiency with little loss in performance.
The report from this week’s Google I/O says Gemini 1.5 Flash is a Transformer decoder model. It has the same 2M+ context and multi-modal features as Gemini 1.5 Pro. It uses Tensor Processing Units (TPUs) efficiently and has lower latency. Gemini 1.5 Flash can compute attention and feedforward components in parallel and has better online extraction. It uses advanced preprocessing methods for better training quality.
The report looks at how fast different AI systems make text in various languages. It checks English, Chinese, Japanese, and French.
For these languages, Gemini 1.5 Flash was the quickest. It could handle 10,000 letters the fastest.
The report compares different Gemini AI models. It looks at how well they do on tests involving coding, math, science, and logical thinking. The results for the newer models are from after they were trained with instructions.
Comparison of Gemini 1.5 Pro with Gemini 1.0 Pro and Ultra in video understanding benchmarks.
Comparison of Gemini 1.5 Pro with USM, Whisper, Gemini 1.0 Pro, and Gemini 1.0 Ultra in audio understanding tasks.
The Gemini 1.5 models did very well on tasks involving long content. They improved long-document and long-video question answering, and long-context speech recognition. They matched or beat the Gemini 1.0 Ultra on many tests.
Google said Gemini 1.5’s performance improved a lot from February to May.
Comparing Gemini 1.5 Pro in May to the February version, the new model improved in reasoning, coding, visual, and video tasks. Audio and translation stayed the same. For FLEURS, lower scores are better.
Oriol Vinyals works at Google DeepMind. He said Gemini 1.5 Pro is better than 1.0 Ultra. Gemini 1.5 Flash is as good as 1.0Â Ultra.