When to use generative AI vs. traditional AI vs. no AI

Real Terms for AI: When Not to Use Generative AI

Key Concepts:

• Generative AI: AI models capable of generating new content (text, images, code, etc.).
• Traditional AI/Machine Learning: Established AI techniques focused on specific tasks like classification, clustering, and prediction using pre-trained or custom-trained models.
• LLMs (Large Language Models): A type of generative AI specializing in language processing.
• Agents: AI systems designed to perform tasks autonomously, often combining multiple AI techniques and tools.
• Function Calling: A technique allowing LLMs to interact with external tools and APIs.
• OCR (Optical Character Recognition): Technology for extracting text from images.
• Sentiment Analysis: Determining the emotional tone of text.
• Entity Extraction: Identifying and categorizing key elements within text.

1. Strengths of Generative AI

The episode begins by acknowledging the power of generative AI, highlighting its core strength: generation. Specifically, it excels at:

• Text Generation: Creating various forms of text, from articles to chatbot responses.
• Summarization: Condensing large amounts of text into concise summaries.
• Image Generation: Producing new images from text prompts.
• Code Generation: Writing code in various programming languages.
• Complex Reasoning & Agent Use Cases: Tasks requiring multi-step reasoning, combining data from multiple sources, utilizing function calling, and building AI agents (as discussed in previous episodes). The ability to leverage memory is also a key strength.

2. When to Opt for Traditional AI/Machine Learning

The central argument of the episode is that generative AI isn’t always the best solution. Traditional AI and machine learning are often more efficient and appropriate when:

• Pre-trained Models Exist: If a pre-trained model already effectively addresses the use case, leveraging it is the most practical approach.
  • Example: Google Cloud Vision API for image labeling, feature detection (faces, landmarks), and OCR. This avoids the need to generate image understanding.
• Specific, Well-Defined Tasks: Many problems don’t require the generative capabilities of LLMs and can be solved more efficiently with specialized models.
  • Examples: Sentiment analysis, entity extraction. Traditional models are specifically trained for these tasks.
• Clustering and Classification: Sorting and labeling data based on similarities.
  • Example: Categorizing customer service requests (returns, information, billing) in a pet shop.

3. Training Your Own Models & Generative AI for Data Creation

When a suitable pre-trained model isn’t available, training a custom model is an option. This requires:

• Labeled Training Data: A dataset of examples where the desired output is already known (e.g., customer requests labeled as “return,” “billing,” etc.).

If labeled training data is lacking, generative AI can be used to create synthetic training data. This is presented as a “super cool” application of generative AI – using it for generation to facilitate other AI techniques.

4. The "If Statement" Rule: Simplicity & Regularity

A key principle is to prioritize simplicity. If a task can be accomplished with basic code:

• If Statements, Switch Statements, Regular Expressions: These should be preferred over complex AI solutions.

This applies when:

• Data is Well-Formed & Regular: Data coming directly from databases or other structured systems.
• Tasks are Straightforward: Simple data manipulation or routing.
  • Example: Extracting order numbers from standardized confirmation emails using a regular expression.
  • Example: Routing phone calls to extensions based on user input using standard code.

5. Use Case Examples & Decision Making

The episode provides several examples to illustrate when to use which approach:

| Use Case | Recommended Technique | Reasoning | |---|---|---| | Reading boss’s emails for sentiment | Traditional AI (Sentiment Analysis) | Existing language models are optimized for this task. | | Summarizing boss’s emails | Generative AI | Requires generating a new, concise summary. | | Spam Detection | Traditional AI/Machine Learning | Classification problem with abundant training data. | | Finding cat pictures | Traditional AI (Image Classification) | Classification task; pre-trained models are available. | | Responding to cat food delivery inquiries | Generative AI (potentially combined with code) | Requires generating personalized responses. |

6. Combining Techniques in AI Agents

The discussion emphasizes that AI applications, particularly agents, often benefit from combining multiple techniques:

• Generative AI for Text Generation: Creating personalized emails or responses.
• Traditional AI for Specific Tasks: Sentiment analysis to gauge customer frustration.
• Plain Code for Logic & Routing: Determining shipping methods or routing phone calls.
• Regular Expressions for Data Extraction: Parsing structured data.

7. Data, Research Findings, and Statistics

While no specific statistics are presented, the episode implicitly acknowledges the abundance of training data available for tasks like spam detection, supporting the use of traditional machine learning.

8. Notable Quotes

• Jason: “AI is just code.” (Emphasizing the fundamental nature of AI as implemented through programming.)
• Aza: “When an if statement or a switch statement or a regular expression would work, use that. It's just code.” (Highlighting the importance of simplicity and avoiding over-engineering.)

Conclusion

The episode delivers a crucial message: generative AI is a powerful tool, but it’s not a universal solution. Developers should carefully evaluate the specific requirements of each task and choose the most efficient and appropriate technique – whether it’s generative AI, traditional AI/machine learning, or even simple code. The most effective AI applications often involve a combination of these approaches, orchestrated within an agent framework. The linked resources (flowchart/decision tree and example techniques) provide practical guidance for making informed decisions.