Google has introduced Gemini 2.0, the latest version of its multimodal AI model, signaling a shift toward what the company is calling “the agentic era.” The upgraded model promises not only to outperform previous iterations on standard benchmarks but also introduces more proactive, or agentic, functions. The company announced that “Project Astra,” its experimental assistant, would receive updates that allow it to use Google Search, Lens, and Maps, and that “Project Mariner,” a Chrome extension, would enable Gemini 2.0 to navigate a user’s web browser to complete tasks autonomously.

The updates mark “an ambitious leap toward AI systems that can independently complete complex tasks and introducing native image generation and multilingual audio capabilities — features that position the tech giant for direct competition with OpenAI and Anthropic in an increasingly heated race for AI dominance,” reports VentureBeat.

Gemini 2.0 Flash, the experimental version of the new model released last week, improves on Gemini 1.5 Flash by supporting multimodal output, in addition to the previously supported multimodal input. The company claims that its latency is greatly reduced, operating at approximately twice the speed of Gemini 1.5. It can natively call tools like Google Search, perform code execution, and interface with third-party tools.

“The practical application of AI agents is a research area full of exciting possibilities,” the company explains in a blog announcement. “We’re exploring this new frontier with a series of prototypes that can help people achieve tasks and get things done.”

Google emphasized responsibility and safety in developing Gemini 2.0 and its various agentic capabilities. The company is taking an “exploratory and gradual approach to development,” and working with internal and external parties to perform risk assessments and safety evaluations.

With Project Astra, the company is building various safeguards that are meant to keep users’ data private and secure. Project Mariner is similarly focused on safety, with an element of its training being the ability to distinguish user instruction from third-party attempts at prompt injection that could have malicious intent.

Google also announced Jules, an experimental tool designed to assist coders, and teased the potential of its Genie 2 model, which can generate playable 3D worlds.

“We really see 2025 as the true start of the agent-based era, and Gemini 2.0 is the foundation of that,” Dennis Hassabis, CEO of Google DeepMind, told The Verge.

“If Gemini 1.0 was about organizing and understanding information, Gemini 2.0 is about making it much more useful,” according to Google and Alphabet CEO Sundar Pichai. “I can’t wait to see what this next era brings.”

Related:
Google Reveals Gemini 2, AI Agents, and a Prototype Personal Assistant, Wired, 12/11/24
Watch Google DeepMind’s Genie 2 Generate Playable 3D Worlds, TechCrunch, 12/11/24
Google’s New Trillium AI Chip Delivers 4x Speed and Powers Gemini 2.0, VentureBeat, 12/11/24