On Thursday, Google introduced Gemini 1.5 Pro, boasting “dramatically enhanced performance” compared to its predecessor. This announcement follows the recent rollout of Gemini 1.0 Ultra and the rebranding of the Bard chatbot to Gemini, aligning with the platform's upgraded capabilities.
In a blog post, Google CEO Sundar Pichai and Google DeepMind CEO Demis Hassabis emphasized their commitment to ethical AI development while showcasing the rapid advancements in their models. “Our teams continue pushing the frontiers of our latest models with safety at the core,” stated Pichai. This focus on safety addresses concerns from AI skeptics and regulators, while also highlighting the urgency to enhance performance for developers, customers, and investors, especially in the wake of OpenAI’s success with ChatGPT.
Gemini 1.5 Pro matches the performance of Gemini 1.0 Ultra but does so with improved efficiency and reduced computational demands. It features multimodal capabilities, enabling the processing of text, images, videos, audio, and code within a single query. As AI models evolve, they offer greater versatility similar to OpenAI's integration of DALL-E 3 with ChatGPT.
With the capacity to handle up to one million tokens—equivalent to over 700,000 words, an hour of video, 11 hours of audio, and codebases with more than 30,000 lines—Gemini 1.5 Pro offers significant enhancements. Google has successfully tested versions of the model supporting up to 10 million tokens. The model demonstrates high accuracy in handling queries with larger token counts and excelled in the Needle In a Haystack evaluation, where it found embedded information 99% of the time in text blocks of one million tokens.
Gemini 1.5 Pro can analyze details from the extensive Apollo 11 mission transcripts and evaluate plot points from a 44-minute silent film featuring Buster Keaton. Its innovative long context window sets it apart from other large-scale models, prompting ongoing developments in evaluation benchmarks, according to Hassabis.
The launch includes capabilities for 128,000 tokens, matching the upper limit of OpenAI’s publicly announced GPT-4 models. Google intends to introduce new pricing tiers that accommodate higher token limits in the future.
A standout feature is Gemini 1.5 Pro's ability to learn new skills through long prompts without needing additional fine-tuning. In the Machine Translation from One Book benchmark, the model successfully learned to translate English to Kalamang—a language with fewer than 200 speakers, achieving proficiency similar to that of a human learner.
In a noteworthy detail for developers, Gemini 1.5 Pro excels at problem-solving across substantial code blocks. When given prompts exceeding 100,000 lines of code, it can reason effectively, suggest modifications, and explain code functions.
On the ethical front, Google continues to employ the same responsible deployment strategies as with the Gemini 1.0 models. These include red-teaming techniques, where a group of ethical developers tests for potential harms, as well as rigorous scrutiny of content safety and representational issues. Continuous development of new ethical and safety protocols is a key focus for the company.
Gemini 1.5 is currently available in early access for developers and enterprise customers, with plans for broader availability. Meanwhile, Gemini 1.0 remains accessible to consumers, along with a Pro variant available for a monthly subscription of $20.
Ranking
