Google’s Gemini artificial intelligence model can be five times more powerful than the OpenAI GPT-4 language model. It should combine the capabilities of GPT-4 and artificial intelligence programs such as AlphaGo.

Gemini: A Strategic Response to Google ChatGPT and GPT-4 OpenAI?

The growing race for artificial intelligence has prompted Alphabet, the parent company of Google, to merge two AI research groups, Google Brain and DeepMind, into a single entity called Google DeepMind. Alphabet’s stated goal is to centralize its efforts and optimize investments. In addition, these two teams have impressive records. DeepMind is the creator of AlphaGo, the artificial intelligence that beat the 2016 world go champion, AlphaFold, which predicts protein structure, and more. Google Brain has contributed to Google Translate and notably created TensorFlow as well as the LaMDA model on which the AI ​​Bard chatbot is based.

From this merger, Project Gemini was born, a series of large language models that combine the power of GPT-4 with learning methods taken from AlphaGo, such as reinforcement learning and tree search. It has the potential to overthrow ChatGPT as the most dominant generative AI solution on the planet. Gemini seems to be a strategic answer to Google ChatGPT. The launch of ChatGPT shook Google to its core. The threat was such that the research giant had to go red code and start investing to catch up with OpenAI in the field of generative AI.

These efforts led to the release of not only Bard, but Gemini as well. While Bard struggles to win over users, Gemini aims to challenge ChatGPT’s dominance in the highly competitive generative AI market. To do this, Gemini will need to leverage its multi-modal capabilities and potential access to Google’s vast in-house training data from various services. Analysts estimate that the generative AI market could be worth $1.3 trillion by 2032, so it’s clear that Google is investing heavily in this space to maintain its leadership position in AI development.

Although many expect the Gemini to hit the market in the fall of 2023, little is known about the capabilities of the model. Last May, Sundar Pichai, CEO of Google and Alphabet, published a blog post providing an overview of the model, highlighting its multi-modal capabilities: Gemini was built from the ground up to be multi-modal, highly efficient in tool and API integration, and built to enable future innovation. such as memory and scheduling. While this is just the beginning, we are already seeing some impressive multi-modal features that weren’t present in previous models.

Once finalized and thoroughly tested for safety, the Gemini will be available in a variety of sizes and capacities, just like the PaLM 2, he added. Since then, little has been known about the official release other than Demis Hassabis, CEO of Google DeepMind, who spoke to Wired, noting that “Gemini will combine some of the strengths of systems like AlphaGo with the amazing language capabilities of larger models.” . Other sources also said that Gemini will be able to generate contextual text and images and will learn from sources such as YouTube video transcripts. But Google hasn’t confirmed anything.

Analysts say the merger of the DeepMind and Google Brain teams should not be underestimated as it pits OpenAI against a team of world-class artificial intelligence researchers, including Google co-founder Sergey Brin and DeepMind senior AI scientist and machine learning expert Paul Barham. Together, these researchers form an experienced team that knows how to apply techniques such as reinforcement learning and tree search to create artificial intelligence programs that can collect feedback and improve problem solving over time.

Does Gemini have the ability to outperform ChatGPT and GPT-4?

One of the biggest conversations around the release of Gemini is whether Google’s mysterious model could topple ChatGPT, which has surpassed 100 million monthly active users this year. At first glance, Gemini’s ability to generate text and images gives it a major advantage over GPT4 when it comes to the range of content it can generate. However, perhaps the most formidable difference between the two is the vast array of Google’s own training data. According to several sources, Gemini can process data from different services.

These include, among others, Google Search, YouTube, Google Books, and Google Scholar. (It’s not clear if Google has the right to use this data to train its AI systems.) Analysts suggest that using this proprietary data to train Gemini models could result in a distinct advantage in terms of the complexity of the information and the inferences they can draw from the set. data. This is especially true if early reports that the Gemini model is being trained on twice as many tokens as GPT4 are correct. Other sources speak of five twenty times as many transactions.

In an article published on Monday, technology journalists Dylan Patel and Daniel Nishball said Google will operate at a speed that will increase its total floating-point operations per second (flops) by five times. the end of the year. The path to a 20x increase is clear by the end of next year, given ongoing infrastructure construction. Whether Google has the courage to make these models public without harming its creativity or its existing business model is another question.

Gemini’s raison d’être for Google is to improve its own products, challenge Microsoft, and ultimately advance medical science. The final release date and capabilities of Gemini are not yet known. The article also points out that Google has the processing power to do the job. According to the article, Google is part of a group of companies that have over 20,000 A100/H100 GPUs for development. OpenAI, Anthropic, Inflection, X, and Meta will be the other organizations in a select group of companies with the computing power to compete in the AI ​​race.

The authors of the article also claim that by the end of next year, some of the aforementioned companies, as well as many Chinese companies, will have over 100,000 units of these GPUs. However, no official source is yet able to confirm these claims or these figures. None of these companies disclose the computing power they have available to develop their AI systems. In announcing these results last week, Nvidia acknowledged it was having difficulty meeting demand for its H100 graphics accelerators, but did not disclose its customers.

Overall, Gemini’s multimodal capabilities, the use of reinforcement learning, text and image generation capabilities, and Google’s own data are all the ingredients Gemini needs to outperform GPT-4. Training data is the main differentiating factor. In the end, which organization wins the race for the largest language models will largely depend on which organization trains its models on the largest and richest dataset. Now the question is what will OpenAI do to keep up with this competition.

And you ?

What is your opinion on this matter?

What do you think of the Google Gemini project?

What do you think we should expect from this project?

Can Google handle the Microsoft-OpenAI partnership?

Can Google regain its lead in the AI ​​race? For what ?

see also

How Google DeepMind plans to revolutionize the field of generative artificial intelligence and replace ChatGPT with the Gemini artificial intelligence system, which combines the strengths of AlphaGo with the power of large language models.

Sergey Brin, co-founder of Google, is signing his comeback to help develop ChatGPT competitor Gemini. Former Alphabet President Involved in AI Strategy Development and Recruitment

Key OpenAI staff leave to join Google, working on AI system that could outperform ChatGPT