Published on December 6, 2023, 11:24 am

Google has unveiled its highly anticipated generative AI model called Gemini. Although the current version, Gemini Pro, is a lightweight version of the more powerful model set to launch next year, it still showcases impressive capabilities.

Gemini is not just one AI model but a family of models. Gemini Nano comes in two sizes, targeting low- and high-memory devices. The easiest way to experience Gemini Pro is through Google’s ChatGPT competitor, Bard, which now runs on a fine-tuned version of Gemini Pro. According to Sissie Hsiao, GM of Google Assistant and Bard, this version delivers improved reasoning, planning, and understanding capabilities compared to its predecessor.

Gemini Pro will be available for enterprise customers using Vertex AI, Google’s machine learning platform. It will also be accessible through Google’s Generative AI Studio developer suite. Additionally, in the coming months, Gemini will integrate into various Google products such as Duet AI, Chrome and Ads, and Search as part of the Search Generative Experience.

Another variant called Gemini Nano will soon launch as a preview via Google’s AICore app. This model will power features like summarization in the Recorder app and suggested replies for messaging apps on Android 14 devices.

Gemini Ultra is another exciting aspect of this lineup. It is trained to be “natively multimodal,” meaning it can comprehend information from various modalities such as text, images, audio, and code. It surpasses OpenAI’s GPT-4 with Vision in terms of understanding multiple contexts beyond just words and images.

However impressive these models may sound, there are certain limitations to consider. For instance, researchers have discovered that generative AI models tend to invent facts or demonstrate biases when prompted in certain ways. It remains unclear whether Gemini Ultra faces similar challenges.

The issue of training data also arises here. While Collins revealed that at least some data was sourced from public web sources and filtered for quality, Google did not address queries about data collection and potential licensing issues.

Google’s training methodology also has environmental implications. Models the size of GPT-4 reportedly emit substantial carbon dioxide during training. Although Google claims to have made Gemini more efficient and cost-effective, it did not provide specific details on these aspects.

Gemini shows promise in various benchmark tests, outperforming GPT-4 in certain areas. However, it is essential to examine the results critically and await further evaluations to assess its true capabilities accurately.

Overall, while Gemini Pro and Gemini Ultra hold significant potential, their launch may have fallen short of expectations. The limited information shared during the briefing left room for doubt and unanswered questions regarding performance, biases, and monetization strategies. Nevertheless, Google’s progress in generative AI with products like Bard, PaLM 2, and Imagen should be acknowledged despite some developmental challenges along the way.


Comments are closed.