Beyond Queries

Exploring the Human-Like Reasoning of Large Language Models

Despite the media attention given to AI, the mechanisms underlying technologies such as Large Language Models (LLMs) remain difficult to understand which is also limiting their use. Recently, I came across an explanation that compares LLMs to human thought, shedding light on their true capabilities. This analogy may help clarify and emphasize that they are more than just an intelligent way to query information.

Large Language Models map the statistical association between a set of ideas.

We inherently understand the statistical likelihood that a certain character combination will form a recognizable word. When we come across an unfamiliar word, we might still perceive it as plausible if its character arrangement seems familiar. Take "Blentive" as an example - it resembles the statistical pattern of a legitimate English word, even though it is as much a non-word as "zqxjk." This statistical likelihood extends from individual letters to words, sentences, and paragraphs.

Large language models are trained to statistically understand the relationships between linguistic fragments, called tokens (GPT Tokenizer), on a massive scale, building a network in which fragments closer to each other have a higher probability of being related than more distant fragments. This approach effectively constructs a digital representation of information along with a mathematical understanding of its relationships, allowing us to make comparisons, derive information, and invites for reasoning.

By design, this mirrors the human brain and our thinking. When presented with statements, ideas, or questions (inputs), we unconsciously deconstruct these inputs and try to identify information that has a high correlation within our existing knowledge or we look for information from which we can derive or approximate from. The similarity is almost comical, because when we cannot process a specific piece of information, we tend to use more distant knowledge and incorporate less connected concepts that have a higher probability of being wrong (in short, we make something up), making us mirror LLMs when they begin to hallucinate. (Hallucination)

Every thought, sentence, or idea connects to a network of related concepts. This web of interconnected information helps us shape our thoughts. In a similar way, Large Language Models (LLMs) tap into this digital network to comprehend inputs, pinpoint relevant starting points within their framework, and generate a response.

Understanding this aspect of AI highlights the importance of two critical variables: Temperature and Top-P. Unlike our brains, which dynamically adjust to focus on relevant information, today's AI requires explicit settings to decide what information to consider. Temperature and Top-P essentially set the boundaries for creativity and context.

Temperature controls the randomness of predictions; lower temperatures result in more predictable and conservative text, while higher temperatures encourage creativity and diversity of responses. Top-P, on the other hand, helps the model determine the range of knowledge to consider by focusing only on the most likely.

Imagine a dark room filled with objects, each object representing a potential word. The Temperature setting in an LLM adjusts the spread of a flashlight's beam, narrowing at low temperatures to focus on common, predictable words, and widening at high temperatures to explore a diverse, less predictable vocabulary. In addition, the "Top-P" setting filters the objects illuminated by the flashlight based on their relevance, selecting a subset of words that are most relevant to the context.

Equipped with the LLM capabilities described above and aware of their parameters, these models can thoughtfully weigh, compare and reason on different ideas and concepts. For example, an LLM can analyse different business strategies and evaluate their potential benefits and risks by synthesizing information.

As we integrate LLMs more deeply into our decision-making, it's critical to recognize their potential beyond simple queening. These tools greatly enhance our ability to analyse, test, and reason about complex scenarios. By tapping into their true capabilities, LLMs can become a partners in formulating strategies and improving business results.

How are you using LLM's as of today ? Share your thoughts in the comments, I look forward to your ideas and feedback.

Constantin

#ArtificialIntelligence #GenerativeAI #AIethics #DigitalTransformation #MachineLearning #TechInnovation #DataPrivacy #AITechnology #FutureOfWork