Over the past week, the AI industry has been buzzing with activity, as both Google and Meta have rolled out new versions of their AI models, garnering significant attention. Firstly, Google announced updates to its Gemini series on Tuesday, introducing two new production-ready models — Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002.
This update has significantly enhanced the overall quality of the models, particularly in mathematics, long context processing, and visual tasks. Google claims a 7% performance boost on the MMLU-Pro benchmark, with a staggering 20% improvement in mathematical tasks. While benchmark tests are known to have limited significance in the AI field, these results are still quite exciting.
In addition to the performance improvements, Google has also drastically reduced the usage costs for Gemini1.5Pro, with input and output token costs falling by 64% and 52%, respectively. This move makes Gemini more cost-effective for developers.
Furthermore, the update has also increased the request processing speed for both Gemini1.5Flash and Pro, with the former now supporting 2000 requests per minute and the latter 1000 requests per minute. These improvements will undoubtedly make it easier for developers to build applications.
On Wednesday, Meta wasn't idle either, launching Llama3.2, a significant update to its open-source heavyweight AI model. This update includes large language models with visual capabilities, ranging from 1.1 billion to 9 billion parameters, along with lightweight text models designed for mobile devices with 100 million and 300 million parameters.
Meta claims that these visual models can rival the leading closed-source models in the market in terms of image recognition and visual understanding. Additionally, some AI researchers have tested the new models, showing that these smaller models perform exceptionally well on many text tasks.
Then, on Thursday, Google DeepMind officially announced a major project — AlphaChip. This project, based on research from 2020, aims to design chip layouts through reinforcement learning. Google states that AlphaChip has achieved "superhuman chip layouts" in its latest three generations of Tensor Processing Units (TPUs), reducing the time to generate high-quality chip layouts from weeks or months to just hours.
Notably, Google has also shared the pre-trained model of AlphaChip on GitHub for public use, allowing other chip design companies to utilize this technology, with companies like MediaTek already adopting it.
Key Points:
📈 ** Google releases new version of Gemini models, enhancing overall performance and significantly reducing costs.**
🤖 **Meta introduces Llama3.2, small language models with visual capabilities, performing exceptionally well.**
💡 **Google's AlphaChip accelerates chip design, significantly improving design efficiency and sharing technology.**