What features are expected in Google's Gemini?

2023-08-17

Google's highly anticipated foundational model for deep learning, Gemini, is expected to be released next month. Demis Hassabis recently announced that engineers at DeepMind are using AlphaGo's technology to develop Gemini, which will become a competitor in the field of artificial intelligence and was previewed at Google I/O. Hassabis stated that Gemini will be more capable than OpenAI's GPT-4.

Hassabis said, "From a high-level perspective, you can think of Gemini as combining the strengths of the AlphaGo-type system with the remarkable language abilities of large models. We also have some new innovations that are very interesting." He added.

In April of this year, Google merged the Google Brain and DeepMind teams into one entity - Google DeepMind. Pichai unexpectedly made this merger with the aim of improving efficiency by combining Google's seemingly endless computing resources with DeepMind's precise research, building more capable systems, and pushing the boundaries of artificial intelligence competition.

Prior to this, both entities had developed their own responses to ChatGPT. While DeepMind launched the Goodall project with an undisclosed model called Chipmunk, Google introduced Bard based on the Google Brain model. Despite the competition between the teams, DeepMind abandoned the Goodall project and shifted to collaborating on the development of Gemini.

However, it is worth noting that PaLM and PaLM 2 were not created by DeepMind. Therefore, Gemini is expected to be DeepMind's first potential commercialized model, rather than being confined to research like interesting models such as Gato.

Although still in the early stages of development, Google reports that Gemini has made significant progress in multimodal capabilities, surpassing previous models. Notably, Gemini's design starts from the ground up and adopts a multifaceted design approach. This design not only prioritizes multimodality, allowing it to process and understand various forms of data, but also emphasizes efficiency in tool and API integration. Gemini's architecture may also facilitate future innovations, particularly in memory and planning.

This progress has a significant impact, indicating that Gemini can better understand and process various types of data. While GPT-4 excels at understanding and generating conversational text, Gemini will surpass this capability by proficiently handling various inputs including text, images, and videos. Gemini will also be able to generate outputs in the form of text, video, audio, music, and images. Additionally, it possesses reasoning capabilities that can facilitate translation in multiple languages and input formats.

Furthermore, discussions among Google employees revolve around utilizing Gemini to achieve various functionalities. This includes analyzing charts, generating graphics with accompanying text descriptions, and performing tasks such as operating software through text or voice commands.

Driving Enterprise Services

Google hopes that Gemini will drive a range of services. These applications include Bard, a chatbot competing with OpenAI's ChatGPT, and enterprise-oriented platforms such as Google Docs and Slides. To achieve this goal, Google aims to charge application developers through its Google Cloud server rental department for access to Gemini. Currently, Google Cloud offers access to lower-level artificial intelligence models designed by Google through Vertex AI. By combining these new features, Google aims to narrow the gap with Microsoft, which is already far ahead in integrating new AI capabilities into the Office 365 suite. Microsoft also provides OpenAI's models to its application users.

Developing New Medical Use Cases

Google has been committed to combining its artificial intelligence models with developing medical use cases. They have been testing an AI tool called Med-PaLM 2, which can answer medical questions. The product is being tested by renowned medical institutions such as the Mayo Clinic Research Hospital.

With the power of Gemini, these efforts could be amplified and used for medical chatbots or assisting in surgical procedures or medical processes.

Building Robots

In addition, Google may also leverage the experience of building DeepMind's "general" system, Gato, which has undergone multimodal and multitask training and completed 604 tasks including image captioning, dialogue, stacking blocks with a robotic arm, playing games, and navigating 3D environments.

What sets Gato apart is the diversity of tasks and training methods, employing a Transformer neural network and various data forms such as text, images, and actions. During deployment, Gato marks prompts and observations to generate actions in sequence.

Similarly, the recently introduced RT-2 is based on the Transformer architecture and trained on network text and images, enabling it to directly generate robot actions.

Like a language model, it guides the robot's behavior by learning from network data. This innovation is built upon visual language models (VLMs) such as PaLI-X and PaLM-E, which use action labels in the output to effectively control the robot's behavior.

Google DeepMind's recent release of RT-2, as the successor to the Robotics Transformer model, also marks an important step forward in the field of robotics. RT-2 is based on the Transformer architecture and trained on network text and images, enabling it to directly generate robot actions.

This innovation draws on the development of visual language models (VLMs) such as PaLI-X and PaLM-E, which use action labels in the output to effectively control the robot's behavior. Similar to a language model, it guides the robot's behavior by learning from network data.

While DeepMind's Gato is seen as a step towards achieving general artificial intelligence (AGI) as it can perform diverse tasks, Gemini may represent practical progress towards similar intelligence.

Potential Suppression of OpenAI's GPT-4

The collaboration between Google Brain and DeepMind could pose challenges for OpenAI and other competitors. Additionally, individuals like former Google President Sergey Brin have joined forces to enhance their artificial intelligence capabilities.

OpenAI CEO Sam Altman believes that video training is the next frontier, however, Google has the advantage and possesses the world's largest video library - YouTube.

Gemini is being trained using YouTube videos, making it the first multimodal model trained not only on text (or text plus images in the case of GPT-4) but also on videos. This could give Gemini capabilities beyond GPT-4. Furthermore, let's not forget that it has almost unrestricted access to the entire internet, a claim recently made by Google through changes to its privacy policy.

Moreover, reports suggest that Gemini has twice the number of training tokens compared to GPT-4 and ten times that of PaLM 2, making it more intelligent and less prone to hallucinations. Furthermore, with the friction between OpenAI and Microsoft, Google may be poised to surpass OpenAI and become the first company to achieve AGI or a similar AGI model.

RECENT AI NEWS

RECENT AI TOOLS