What: LLMs come with inherent limitations that leaders need to understand.

Why it is important: The transformative nature of LLMs and GenAI is too important for the mechanisms to be ignored by decision-makers.

Large language models (LLMs) like ChatGPT are transforming businesses with their ability to generate detailed, humanlike responses in plain language and code. ChatGPT rapidly gained 100 million users, and in the first half of 2023, AI startups received over $40 billion in investments, indicating significant interest in LLM applications.

Despite their capabilities, LLMs can be misleading due to their humanlike text outputs, which can be mistaken for human intelligence. Understanding their mechanics and limitations is crucial for effective application in businesses. LLMs predict text sequences using machine learning, with advancements like neural network transformers allowing for more context-aware responses. The model’s size, training data quality and volume, and context window size are key to its performance.

However, LLMs have significant limitations:

1. Reasoning: They struggle with complex logical reasoning, often producing errors in multistep logical tasks.
2. Knowledge/Expertise: LLMs are limited by their training data and can generate outdated, incorrect, or "hallucinated" information.
3. Understanding: They may not fully grasp the context or details of prompts, leading to incoherent or disjointed responses.
4. Planning/Execution: LLMs can suggest impractical or naive actions due to limited understanding and reasoning capabilities.

In business, these limitations can lead to unreliable outputs. Effective application requires complementing LLMs with human oversight and other technologies. Keeping humans in the loop is essential for validating AI-generated content and translating complex business problems into effective prompts.

To enhance LLMs, researchers are working on augmenting them with reasoning engines, domain-specific training, and reinforcement learning from expert feedback. Businesses should stay informed about these advancements and carefully consider the context in which LLMs are deployed, especially in high-stakes or logic-intensive scenarios. Proper understanding and application of LLMs, combined with complementary technologies and human oversight, can maximize their potential while mitigating risks.

The Working Limitations of Large Language Models