Demystifying AI Jargon: A Semeris Guide

Published 12. June 2023.

As AI continues to evolve and shape the future, the language surrounding it can sometimes prove a bit tricky to navigate. At Semeris, we're here to help. We believe that a mutual understanding of key AI concepts is crucial for our collaboration. So, let's demystify some essential AI terms together, making this exciting field more accessible and less intimidating.

Large Language Models (LLMs): LLMs, such as GPT-4 (OpenAI), LLaMA (Meta), and BLOOM (BigScience), are AI models engineered to understand, generate, and manipulate human language. Astonishingly, their primary method of learning is by predicting the next word in a sentence, using vast amounts of written language data, which even includes legal text. Despite this seemingly simple task, they manage to absorb the intricacies and nuances of human language, enabling them to perform a wide variety of tasks. These models power many of the cutting-edge innovations in AI, ranging from simple chatbots to sophisticated tools for text analysis.

Solving Tasks: When we talk about an AI "solving tasks," we refer to its ability to achieve a diverse array of objectives. This might involve something straightforward like answering a trivia question, but it could also encompass far more complex endeavors. For example, an AI can summarize a dense legal document, detect named entities in a sea of text, or even transform legal clauses into actual software code. The real magic of Large Language Models lies in their sheer versatility – they're capable of tackling tasks that span from the simplest language-based activities to high-level cognitive challenges that were once exclusive to human expertise.

Zero-Shot Learning: This is an ability of AI to tackle tasks it hasn't encountered during training. Say you provide a complex contract clause to an AI model and ask for a summary. Despite never encountering this exact clause, the model draws upon its vast training data to generate a concise summary. This ability for on-the-spot adaptation is a hallmark of zero-shot learning in LLMs.

Few-Shot Learning: Imagine you're teaching an AI to recognize contract renewal clauses. You'd start by showing it a few different examples of these clauses. Then, when you present a new, unseen clause, the AI uses the given examples to identify it as a renewal clause. This teaching-by-examples approach is what we refer to as few-shot learning.

Ground Truth: This term has its roots in cartography, where 'ground truth' referred to information collected on location (the ground) rather than inferred. In AI, 'ground truth' denotes the confirmed, reliable data we use as a baseline for training and evaluation of AI models. At Semeris, our expert analysts specialize in creating 'ground truth' datasets, such as human-verified entities in legal documents, ensuring our AI models have a solid foundation of truth to learn from.

Hallucination: The term 'hallucination' in AI refers to when an AI model produces output that seems plausible but isn't supported by the input or its training data. This could happen due to conflicting information within the training data or simply because the AI veers off-course. At times, these missteps can be quite substantial, producing outputs that are significantly off the mark. At Semeris, we are experts at minimizing these instances, ensuring you receive dependable and accurate AI outputs.

Prompt Engineering: A "prompt" is essentially an input given to an AI system to guide its response, akin to framing a question in a Q&A game. The craft of designing these inputs is what we refer to as prompt engineering. In the context of legal documents, a prompt could also include a specific clause or paragraph that we ask the AI to analyze. Picking that clause automatically is a task in itself for the Semeris Docs platform. Here at Semeris, we're constantly refining our prompt engineering techniques, closely following the latest whitepapers in the field to ensure the responses we generate are as informative and relevant as possible.

At Semeris, we're applying AI technologies to create practical, impactful solutions. We're making legal language easier to digest, enriching our library with data points verified by our expert analysts – consider them our 'AI wranglers', subtly guiding the AI towards the right path.

Simply put, Semeris is here to cut through the noise and help you make the most of this groundbreaking tech. If you've got AI ideas to explore or jargon that needs busting, we're always here to lend a hand.