Code Execution with Gemini | Intro to Tools

Key Concepts

• Large Language Models (LLMs): AI models that process and generate human-like text by predicting the next word in a sequence based on learned patterns.
• Code Execution Tool: A feature that allows LLMs to generate and run code (e.g., Python) to perform computations, data analysis, and other tasks.
• Gemini Flash: A specific LLM model mentioned in the transcript.
• Google AI Studio: A platform for developing and prototyping with AI models.
• Prime Numbers: Numbers greater than 1 that have only two divisors: 1 and themselves.
• CSV (Comma Separated Values): A file format used to store tabular data.
• Pandas: A popular Python library for data manipulation and analysis.
• Data Frame: A two-dimensional labeled data structure with columns of potentially different types, commonly used in data analysis.

Limitations of LLMs in Mathematical Computation

Large Language Models (LLMs), despite their advancements, fundamentally operate by predicting the most probable next word in a sequence. This mechanism, while effective for language generation, makes them inefficient and inaccurate for precise numerical computations. They are not designed as calculators.

Enhancing LLM Capabilities with Tools: The Code Execution Tool

To overcome the limitations of LLMs in mathematical and computational tasks, they can be connected to external tools. The code execution tool is highlighted as a method to significantly improve Gemini's effectiveness in areas involving math, data analysis, computation, and code-based reasoning.

Example 1: Calculating the Sum of Prime Numbers

Initial Attempt without Code Execution

• Model: Gemini Flash (latest)
• Prompt: "calculate the sum of the first 60 prime numbers."
• Expected Answer: 7,699
• Actual Output: 8,368
• Observation: The LLM produced an incorrect answer. This is attributed to its reliance on the "next token mechanism" for problem-solving, which is not optimal for mathematical accuracy. The model's output was a long string of text attempting to derive the answer.

Improved Attempt with Code Execution

• Process: The same prompt was used, but with the code execution tool enabled in Google AI Studio.
• Mechanism: Instead of directly calculating, the LLM generated Python code. This code included a function to determine if a number is prime and then summed the first 60 prime numbers.
• Execution: The generated Python code was executed on the API backend.
• Actual Output: The correct answer, 7,699, was returned.
• Benefit: This approach not only yields the correct mathematical result but also provides the generated Python code. This allows users to review the logic and understand how the answer was derived, offering transparency and auditability.

Example 2: Practical Data Analysis with CSV Genius App

Scenario: Analyzing Tropical Fruit Stand Data

• Application: A custom app called "CSV Genius" built using Vibe Coding in AI Studio.
• Functionality: Allows users to upload CSV files and ask natural language questions about the data.
• Data Source: A CSV file containing transaction data from a tropical fruit stand, including columns like transaction ID, date, product sold, quantity, and price.
• Goal: To enable users to perform data analysis tasks, similar to what might be done with the Pandas library, without requiring explicit coding knowledge.

Query and Result

• Prompt: "what fruit did I sell the most of (quantity)."
• LLM's Task: To analyze the uploaded CSV data and identify the fruit with the highest quantity sold.
• Result: The app reported that mangoes were sold the most, with 63 units.

Understanding the Derivation: Generated Code

• Feature: A "generated code toggle" within the app allows users to view the code the LLM produced to arrive at the answer.
• Generated Code Analysis:
  1. The LLM loaded the uploaded CSV data into a data frame.
  2. It then filtered the data, specifically for "fresh fruit" (though the transcript doesn't detail the exact filtering logic for this).
  3. Finally, it processed the data to identify and print the fruit with the highest quantity sold.
• Transparency: This generated code snippet demonstrates that Gemini wrote and executed Python code to perform the data analysis, confirming the origin of the "mangoes" result.

Technical Implementation in AI Studio

• AI Chat Creation: Within the AI chats creation method in the code section of AI Studio, the following configurations were observed:
  • Model: Gemini 2.5 Flash was selected.
  • Tool: The code execution tool was explicitly passed, enabling the model to call and utilize this functionality.

Conclusion and Takeaways

The transcript emphasizes that while LLMs excel at language and code generation, they are not inherently suited for precise mathematical computation. By integrating LLMs with tools like the code execution tool, their capabilities can be significantly extended. This allows LLMs to leverage their strengths in generating code to perform complex calculations and data analysis, providing accurate results and transparent, auditable processes. The CSV Genius app serves as a practical example of how this integration can democratize data analysis, making it accessible to users without deep technical expertise.