Lesson 3B: Capabilities & limitations | AI Fluency: Framework & Foundations Course

Generative AI: Capabilities and Limitations of LLMs

Key Concepts:

• Large Language Models (LLMs): AI systems like Claude skilled in language processing.
• Training Data: The data used to train LLMs, which determines their knowledge base.
• Knowledge Cutoff Date: The date after which an LLM has no inherent knowledge.
• Hallucination: An LLM confidently stating something incorrect.
• Context Window: The amount of information an LLM can process at one time.
• Non-deterministic: The characteristic of LLMs producing slightly different outputs for the same input.
• Temperature: A setting that controls the randomness of LLM output.
• Retrieval Augmented Generation: A technique connecting LLMs to external knowledge sources.
• AI Fluency: Understanding what AI can and cannot do.

What LLMs Do Well

• Versatile Language Skills: LLMs demonstrate remarkable language abilities, including:
  • Crafting emails in a specific voice.
  • Condensing lengthy reports into summaries.
  • Translating between languages.
  • Explaining complex topics across diverse fields (e.g., microbiology, marketing strategy).
• Task Switching: LLMs can seamlessly transition between different tasks without retraining. For example, the same model can write poetry and analyze business trends.
• Conversational Memory: LLMs can maintain context and remember previous parts of a conversation, similar to human conversation partners.
• External Tool Integration: Modern LLMs can connect to external tools and information sources, such as:
  • Web search.
  • File processing.
  • Other applications. This expands their capabilities significantly.

Limitations of LLMs

• Training Data Boundaries:
  • Knowledge Cutoff Date: LLMs have a knowledge cutoff date, meaning they lack inherent knowledge of events after that date. For example, a model with a cutoff date of November 2024 is not trained on data after that date.
  • Models need tools like web search to learn more about recent developments.
  • Inaccuracies in Training Data: LLMs can learn and reproduce inaccuracies present in their training data.
  • Hallucinations: LLMs can generate incorrect information that sounds plausible. This is due to generating responses based on statistical patterns rather than retrieving verified facts.
• Context Window Limitations:
  • LLMs have a limited context window, restricting the amount of information they can process at once.
  • Exceeding the context window results in the AI forgetting information, typically on a first-in, first-out basis.
  • This limits the ability to process large documents or remember entire conversations.
• Non-Deterministic Output:
  • LLMs are inherently non-deterministic, meaning they can produce slightly different outputs for the same input.
  • This variability stems from probabilistic decisions about text generation based on training data patterns.
  • The "temperature" setting can control this randomness. Lower temperature means less randomness.
• Reasoning Limitations:
  • LLMs have historically struggled with complex reasoning tasks, especially those involving mathematical or logical problems with multiple steps.
  • Newer "reasoning" or "extended thinking" models are showing progress in this area.
• Limited Access to Data and Tools:
  • Even with external tool access, LLMs may lack access to specific data sources or specialized tools required for certain tasks.
  • If a model lacks access to necessary data or tools, it cannot provide accurate or complete answers.

Addressing Limitations and Future Evolution

• Retrieval Augmented Generation (RAG): This technique connects LLMs to external knowledge and data sources to improve accuracy and reduce hallucinations.
• Ongoing Research: Researchers are actively working to improve LLMs' reasoning capabilities and expand their access to tools.
• Complementary Strengths: The most effective applications of AI leverage the strengths of both humans and AI.
  • Humans provide critical thinking, judgment, creativity, and ethical oversight.
  • AI offers speed, scale, pattern recognition, and the ability to process vast amounts of information.
• Continued Learning: Staying abreast of AI advancements and experimenting with new possibilities is crucial.

Conclusion

Understanding the capabilities and limitations of generative AI, particularly LLMs, is essential for "AI fluency." This knowledge enables effective integration of these systems into work and daily life. The key is to leverage the complementary strengths of humans and AI, recognizing that AI's limitations will continue to evolve. Direct experience and experimentation are valuable for developing an intuitive understanding of what generative AI can and cannot do.