US technology giant Google continues to launch attacks on OpenAI and Meta in the field of big language models.
On the evening of February 21st, Google announced that the new generation of free and commercially available large language model Gemma is open for use worldwide. This model is regarded by Google as its "most advanced open model".
This is a major move made by the company in the field of open AI big models. Tris Warkentin, Director of Product Management at Google DeepMind, stated that open models are a new opportunity for Google to collaborate with communities and people outside of Google to create new opportunities in AI development.
Gemma is named after the Latin word "gemstone" and is only used to process text information. Its basic technical architecture is consistent with Google's strongest AI model Gemini, but its parameter size is relatively small, with only two versions of 2 billion and 7 billion parameters, and both Gemma models have pre trained and instruction fine-tuning versions.
A smaller parameter size helps Gemma achieve wider deployment. Google introduced that Gemma supports mainstream AI frameworks and can also run on environments such as laptops, desktops, the Internet of Things, mobile devices, and the cloud.
The evaluation results released by the company show that Gemma outperforms the Llama 2 model in many external benchmark tests such as mathematics, coding, reasoning proficiency, and knowledge testing. Llama 2 is the latest generation open source big model released by Meta, which includes models with 7 billion, 13 billion, and 70 billion parameters.
It is worth noting that Google emphasizes that Gemma is an open model rather than "open source", which means that Google will not share multiple technical details of Gemma, including its source code, training data, etc. On the application side, Google claims that its terms of use allow all organizations to responsibly engage in commercial and distribution.
Open Gemma or partial response to criticism in the field of open source big models. Previously, Google and OpenAI were criticized by the outside world for adhering to technological isolation, and both chose to use isolation in their latest and most advanced models, which was considered detrimental to technological progress.
Regarding this, Zhang Junlin, the head of new technology research and development on Sina Weibo, commented that Gemma represents a shift in Google's big model strategy - balancing open source and closed source, with open source focusing on the most powerful small-scale models, hoping to defeat Meta and Mistral (European AI company launched Mistral 7B open source AI model); Closed source focuses on large-scale models with the best performance, and hopes to catch up with OpenAI as soon as possible.
In the AI community, Meta's Llama 2 has always been one of the most powerful open source big models, and the model information and source code support free commercial use, thus gaining a large number of AI developer support.
Google clearly hopes to attract more developers into the Google cloud ecosystem through Gemma. On the one hand, Gemma has optimized Google's self-developed cloud AI chip TPU, claiming that it can achieve better performance. Meanwhile, new users of Google Cloud will also receive $300 in cloud credits to study Gemma.
In addition, Gemma will be able to run on Nvidia chips and be optimized through collaboration between both parties to accelerate the inference performance of the model in cloud data centers and PC side. If Gemma is used on AI PCs equipped with Nvidia GPUs to drive local chatbot software and integrate with Nvidia's multiple AI tools.
The big model battle among large technology companies such as OpenAI, Google, and Meta is becoming increasingly fierce.
Google launched the AI dialogue robot Bard in March 2023 and the latest closed source big language model PaLM2 in May last year. Last week, the company officially announced the "next-generation AI big model" Gemini 1.5, stating that it has surpassed OpenAI's GPT-4 Turbo in many aspects. Meta is passionate about open source models, and its Llama 2 is the most well-known.
In recent days, OpenAI has once again ignited the AI industry with the release of the Sora video model, further distancing itself from other large model companies. Google's ultimate goal of catching up with OpenAI will still be filled with many uncertainties.