Uncategorized
28 February 2024
Google’s new flagship AI model, “Gemini,” is set to directly compete with GPT-4, boasting five times the computational power of GPT-4. Trained on Google’s TPUv5 chips, this device has the capacity to process tasks simultaneously with a massive 16,384 chips. The dataset used to train this model is approximately 65 trillion tokens in size. Moreover, it can generate both text and images. The training also included content from YouTube and utilized advanced training techniques similar to those used in “AlphaGo.” Google plans to publicly launch the Gemini model in December 2023.
Gemini is a series of large language models that combine GPT-4 with training techniques such as reinforced learning and tree search from AlphaGo, potentially dethroning ChatGPT as the most dominant generative AI solution on the planet.
The news follows several months after Google merged its Brain and DeepMind AI labs to form a new research team called Google DeepMind and only a few months after the launch of Bard and the next-generation PaLM 2 LLM.
Researchers estimate that the generative AI market will be worth $1.3 trillion by 2032, indicating that Google is likely to continue investing in the field to maintain its leadership position in AI development.
Although many expect the launch of Google Gemini in the fall of 2023, little is known about the model’s capabilities.
In May, Sundar Pichai, CEO of Google and Alphabet, published a blog post offering a high-level overview of LLM, stating:
“Gemini is built from scratch to be highly efficient in multimodal, tool, and API integrations, and is constructed to enable future innovations such as memory and planning.”
Pichai also noted, “Although it’s still early, we are already seeing impressive multimodal capabilities that were not seen in previous models. After security refinements and rigorous testing, Gemini, like PaLM 2, will be available for sale in various sizes and capabilities.”
Since then, aside from Google DeepMind CEO Demis Hassabis’s comment in an interview with Wired that Gemini would “combine some of the strengths of AlphaGo-type systems with the remarkable language capabilities of large models,” little official information has been released about the launch.
Android Police also claimed that an anonymous source mentioned Gemini’s ability to generate text and contextual images and receive training on resources such as YouTube video transcripts.
One of the biggest debates surrounding the launch of Gemini is whether the enigmatic language model has what it takes to dethrone ChatGPT, which has reached over 100 million active users this year.
At first glance, Gemini’s ability to generate text and images gives it a serious advantage over GPT-4 in terms of the range of content it can produce.
However, perhaps the most dangerous difference is Google’s vast array of proprietary training data. Google Gemini can process data from services including Google Search, YouTube, Google Books, and Google Scholar.
The use of these proprietary data in training Gemini models could significantly enhance the complexity of insights and inferences it can draw from a dataset, especially if initial reports of Gemini being trained on twice as many tokens as GPT-4 are accurate.
Furthermore, this year’s partnership between Google DeepMind and Brain teams cannot be underestimated, as it pits OpenAI against a team of top-tier AI researchers, including Google co-founder Sergey Brin and senior DeepMind AI.
This team, experienced in creating AI programs capable of learning from feedback and improving problem-solving over time, has a deep understanding of applying techniques such as reinforced learning and tree search, demonstrated in teaching AlphaGo to defeat a Go champion.
Gemini’s multimodal capabilities, enhanced usage in parenting, ability to generate text and images, and Google’s proprietary data appear to be all the components needed for Gemini to outperform GPT-4.
The question now is: What will OpenAI do in response?
Leave a Reply