Google is preparing to release Gemini, the most advanced large language model (LLM) its AI work has produced so far. With Gemini, the search giant hopes to pass OpenAI’s GPT-4 on multimodal learning and on the ability to carry out more human-like tasks, and to move a step closer to human-level AI.
The general public has not seen how Gemini performs yet, but Google has given certain companies access to try the new LLM.
What Is Google Gemini?
Widely read as Google’s answer to GPT-4, Gemini is described by Demis Hassabis, CEO and co-founder of Google DeepMind, the Google subsidiary working on AI research and the creator of Gemini, this way: “Gemini is not merely a combination of scale, it is a combination of innovations.”
The LLM builds on DeepMind’s earlier multimodal models such as Flamingo and PaLM 2, both to build its own conversational AI and to push natural language processing further. We thought the GPT-3 model was large at over 170 billion parameters. Now we expect Gemini to pass that and become the largest language model built so far.
Google Gemini Features
Beyond being the largest language model, recent reports point to several Google Gemini features worth waiting for. Some of them:
• A family of models: Hassabis said we should think of Gemini as “ individual models at different sizes ”, which may mean the new platform can serve tasks of any size depending on the use case.
• Multimodal learning : on top of Google DeepMind’s earlier multimodal models, Gemini’s AI infrastructure carries “ several improvements ” that let it learn from and produce at least text and images.
• Problem solving and reasoning : Google’s new LLM will also be able to filter out the faulty information typically found in Gemini’s rivals, ChatGPT among them.
• Fact checking and memory : Google Search has been built into Gemini’s AI infrastructure as a tool, for better accuracy in what it generates. Gemini will also use “episodic memory banks” to store and retrieve data, letting it develop and widen its knowledge base as it learns.
These features give us a look at how Gemini will perform against other AI personal assistants and applications such as ChatGPT and Bard. We can expect better conversational AI infrastructure and more accurate, more relevant results for users.
Google’s Best AI Work
Google’s recent progress in AI should be no surprise. Even before GPT-3 arrived, Google was quietly building its own AI infrastructure to meet what users needed.
In 2018 Google introduced Duplex, a voice based AI assistant that could make phone calls and restaurant reservations on the user’s behalf. Beyond that, in October 2019 Google launched BERT to improve natural language understanding (NLU) in search queries.
Google has also taken several steps in recent months to develop its existing large language models and its conversational AI capabilities:
• Search Generative Experience (SGE) : an AI search assistant approach that summarizes content for the user while still sending traffic to the publisher’s website.
• Google Bard : Google’s conversational AI, using a chat interface to answer questions, summarize subjects and even write poems.
• Responsible AI with Anthropic : Google invested $300 million to double down on its mission toward safer, more ethical use of generative AI.
Gemini looks like the peak of Google’s efforts and is the company’s largest language model to date. That gain in scale is not for show. According to Hassabis, they brought the error rate down from 10% to just 1%, which is an enormous improvement.
In the end, in our web design work and our digital marketing processes we will have an alternative to OpenAI as well. Which gives us a chance to look again at the role of AI in digital marketing.
Gemini vs ChatGPT-4
The Türk SEM View
Anyone interested in AI and in Google should welcome this Gemini news. In the end we will have a conversational AI technology going toe to toe with OpenAI.
For now, though, all you can do is sit and wait while a handful of selected developers find their way around Google Gemini. You can still reach some of Google’s experimental features through Google Labs. To try the ones below, all you have to do is join the waiting list.
• Project IDX
• Magic Compose
• Search Generative Experience
• Duet AI
• MusicLM
We Asked Bard: “Which Is Better, Gemini or Bard?”
Gemini and Bard are two large language models (LLMs) developed by Google AI. Both can generate text, translate languages, write different kinds of creative content and answer your questions informatively. There are some important differences, though.
Gemini is an LLM trained on a dataset of text and code. That lets Gemini produce more creative and more engaging text. Gemini can generate different creative text formats: poems, stories, code, scripts, pieces of music, emails, letters and so on.
Bard is an LLM trained on a dataset of text and real world information. That lets Bard supply more accurate and more informative material about the real world. Bard can give you information on historical events, scientific subjects or current affairs.
Broadly, Gemini is better at producing more creative and engaging text. Bard is better at supplying more accurate and informative material about the real world.
We put the same question to ChatGPT-4 as well, but GPT-4 had no idea what was going on and compared the star sign Gemini with the concept of a bard.
I would expect Google to be a little more original in how it names things.
