Hello IP World! Gemini 1.5 is your new innovation Wingman

谷歌 has recently released a new AI model called Google Gemini 1.5. This breakthrough release from Google, not only introduces advanced AI acceleration but also extends the context length to an impressive 1 million tokens. The substantial increase in context length holds tremendous implications for AI workflows, allowing the model to process and understand intricate relationships within extensive datasets. This expanded capability paves the way for more sophisticated language understanding, context-aware predictions, and nuanced decision-making.

However, it's important to note that access to Gemini 1.5 is currently restricted, with interested users having to join a waitlist for eventual access. This exclusivity adds an element of anticipation, underlining the significance of this AI model in the eyes of the tech community. As eager anticipation builds, it is clear that Gemini 1.5 is set to be a game-changer, reshaping the landscape of AI applications once it becomes generally available.

Integration of “Mixture of Experts”

Gemini 1.5, the latest iteration of the AI model, introduces substantial improvements over its predecessor. One notable enhancement is its utilization of the “mixture of experts” (MoE) model, a specialized technique that enhances efficiency and speed. This model incorporates a constellation of "expert" neural networks, which, depending on the input, selectively activates the most relevant pathways, significantly boosting efficiency. Unlike executing the entire model with each query, Gemini 1.5’s MoE selectively leverages these expert pathways, resulting in faster response times and improved performance.

This innovative model signifies a leap forward in the evolution of AI, pushing the boundaries of what was previously attainable and creating new opportunities for integrating artificial intelligence across diverse applications. The implications of this development are poised to profoundly impact the efficiency and effectiveness of AI-driven solutions across various industries.

Context Length Explained: How Big is Big?

In our recently published article, Gen AI & IP: The Next Power Couple, we explored key concepts such as context length and tokens. To recap,

• Think of a token as roughly three-quarters of an English word. While GPT-4 boasts a context length of 128,000 tokens, its optimal performance drops after processing just 16,000 tokens.
• In practical terms, when using GPT-4, it's recommended to stick to around 16,000 tokens or roughly 12,000 words. To put it in perspective for IP and legal language, this aligns with the average length of a Granted Patent.
• Now, with Gemini 1.5, there's a game-changer. It supports an expansive context of up to 1 million tokens – that's about 750,000 words, roughly equivalent to the entire text length of 60 patents.

Real-World Use Cases: Pushing the Boundaries

The introduction of a 1 million context length in Gemini 1.5 marks a revolutionary shift. It eliminates the need to segment your text for AI processing into multiple files, followed by a sequential analysis of each file. With Gemini 1.5, you can streamline the process by directly inputting multiple files in a single call to the model, simplifying and expediting the analysis of your data.

• Analysing text

Ethan Mollick, an educator at Wharton specializing in AI/LLM Studies, conducted an experiment by submitting over 20 of his own academic papers, totalling 1000 pages in PDF format, as a single query to Gemini 1.5. Remarkably, within less than a minute, Gemini efficiently identified the common themes addressed across the extensive collection of papers.

• Analysing long videos

The proficiency of understanding lengthy videos was exemplified during the launch of Google's Gemini 1.5 Pro. In a demonstration, developers showcased the model's capability by using a 44-minute silent film as a prompt. They further tested its accuracy with multimodal prompts. In a similar vein, when presented with a lengthy video of the entire NBA dunk contest and queried about the dunk with the highest score, the Gemini 1.5 Pro model successfully pinpointed the optimal 50 dunks based on its adeptness in comprehending extended video contexts.

• And, the list keeps on going

1. Gemini 1.5 Pro can complete language translations within minutes, including the ability to translate entire newsletters from English to languages with limited speakers, such as Saterland Frisian, spoken by only over 2000 people.
2. With its advanced capabilities, Gemini 1.5 Pro can skilfully interpret even the most intricate tables and statistics found in lengthy reports within PDF files.
3. Sully Omar demonstrated the remarkable capabilities of Gemini 1.5 Pro by uploading an entire codebase directly from GitHub, including all issues and the Vercel AI SDK. Not only did the model comprehend the codebase, but it also identified the most urgent issue and implemented a fix.

Vector Databases Still Relevant?

The use of either a Large Context Length or Retrieval Augmented Generation (RAG), also known as a Vector Database, is a prevalent technique in connecting AI models to a knowledge base. In the ongoing discourse within the AI community, the debate revolves around which approach is more effective. Let's explore the distinctions:

Let’s understand this:

Surely, large context length doesn’t kill RAG or vector storage techniques – they have their applicability, but the large context length is enabling the next iteration of AI products.

Sagacious IP’s AI Portfolio gets a boost

At Sagacious IP, we take pride in our portfolio of over 20 AI tools designed for diverse IP services and workflows. The upcoming incorporation of Google Gemini 1.5 into our AI arsenal is poised to elevate the performance of several tools. For personalized AI solutions tailored to your unique IP workflows, reach out to us.

Additionally, we're currently engaged in a comprehensive GenAI landscape study. Stay tuned for the release by following our LinkedIn channel to stay abreast of the latest developments.