Google, not to be outdone by OpenAI’s 12 days of announcements, dumped a massive set of releases, headlined by the launch of Gemini 2.0, its next-generation AI model. Starting with an experimental version called Gemini 2.0 Flash, this model generates text, images, and speech while processing various inputs, including text, images, audio, and video. It outperforms its predecessor, Gemini 1.5 Pro, with twice the speed on key benchmarks. Currently available through Google’s developer platforms, some features are limited to early access partners until January 2025. Sundar Pichai, Google’s CEO, emphasized Gemini’s potential to support “agentic” AI systems capable of thinking ahead and acting on users’ behalf.
Key highlights include:
- Multimodal Live API: This new tool supports real-time audio and video streaming, enabling developers to build more interactive applications.
- Deep Research: Exclusive to Gemini Advanced (priced at $20/month), this feature assists with comprehensive research by creating multi-step plans, refining analyses, and generating detailed reports.
- Project Mariner: Formerly known as Jarvis, this rebranded AI-powered Chrome extension navigates websites like a human user. Currently available to select testers, it demonstrated tasks such as finding contact information, albeit slowly. Google describes it as a research prototype for understanding and reasoning across web elements.
- Project Astra: An experimental AI assistant leveraging Gemini 2.0, Astra captures and summarizes up to ten minutes of video via an Android app or prototype glasses. Improvements include enhanced language support and reduced lag for more conversational interactions. However, Astra still struggles with noisy environments and cannot access personal data like emails or photos. It remains in testing, with a waitlist for additional users.
- Jules: Google’s new AI coding assistant helps developers debug and automate tasks within GitHub workflows. Jules generates multi-step plans, modifies files, and prepares pull requests for Python and JavaScript. Trusted testers are using it now, with a broader release in early 2025.
Google is also scaling its infrastructure, deploying over 100,000 custom chips to support these advancements.
Why do we care?
The trend is to multi-modal models – yet notable that many businesses are seeing significant value in specialized, smaller language models. Keep that in mind.
Projects like Astra and Mariner are promising but unfinished. Astra’s inability to handle noisy environments and Mariner’s slow performance suggest these tools are not production ready. The insight is the push to agents. These are the early days, and I’d be cautious about customer implementations, yet also highly experimental in discovering capabilities. And I say that knowing it’s a hard line to walk with limited resources.
Use Gemini’s announcements as an opportunity to educate clients about emerging AI capabilities and their limitations, positioning yourself as a trusted advisor rather than just an implementer.