AI4Legislation held its 3rd webinar last Sunday, June 8 2025, featuring an LLM expert from the East Coast: Mr. Leon Guo! Leon discussed various aspects of AI language models, including their capabilities, limitations, and applications in coding and data analysis. We also explored AI tools, data scraping techniques, and the potential of AI agents, concluding with discussions on project ideas and resources for model selection.

AI in Programming: Opportunities and Challenges

The meeting began with Chunhua promoting an upcoming competition and introducing the AI for Legislation Tech Seminar. Cynthia hosted the seminar, introducing speakers and managing the presentation flow. Leon, an expert in large language models, discussed the evolving role of computer science and the impact of AI on programming. He highlighted how AI tools like GitHub Copilot have revolutionized coding by reducing the need for manual debugging and API documentation reading. However, Leon also noted limitations of AI, such as high operational costs, slow processing speeds, and the potential for hallucinations and errors. The session concluded with an open invitation for questions, which attendees were encouraged to submit via chat.

AI Model Cost and Performance Comparison

Leon presented a comparison of AI language models, focusing on cost and performance. He highlighted that the latest Gemini model is more affordable than OpenAI’s GPT-3, with input tokens costing $1.25 to $2 and output tokens $10, compared to OpenAI’s $10 input and $40 output tokens. Leon also demonstrated a visualization tool he created to compare different AI models across companies, showing that 2025 is a significant year for AI advancements, particularly from Google. Finally, Leon discussed the use of Python for data analysis, noting its advantages over traditional BI software like Power BI, particularly in handling large datasets and performing mathematical operations accurately.

Data Scraping and Visualization Tools

Leon presented on data scraping and visualization tools, highlighting Python libraries like requests, Selenium, and Scrapy Spider for web data collection. He introduced several public datasets, including Ledger Scan API and US government data, and demonstrated how to use Google Gemini for data visualization prototyping. Leon also discussed the capabilities of Grok, emphasizing its real-time data features and comparison to other tools like ChatGPT and Claude.

AI Tools for Developers

Leon demonstrated the capabilities of different AI tools, particularly focusing on GitHub Copilot’s debugging and code generation features, which he noted costs $10 per month and offers more advanced functionality than ChatGPT. He explained that different AI models serve different levels of developers, from beginners to professionals, and emphasized that learning Python is still valuable for problem-solving, though he suggested a practical approach of starting with specific problems rather than a formal course. Chunhua agreed that Python knowledge is beneficial for verifying AI outputs and guiding the process, while the conversation ended with Jeff expressing appreciation for the presentation despite not being in the tech field.

AI Agent Customization and Limitations

Leon explained the capabilities and limitations of AI agents, particularly focusing on large language models and their ability to handle specialized tasks. He clarified that while current models have a knowledge cutoff date and cannot learn from ongoing user interactions, they can be customized for specific industries or fields through tools like MCP (Model Context Protocol) and fine-tuning. Jeff inquired about the feasibility of building small, specialized models for specific tasks, and Leon discussed the costs and challenges involved, including the risk of obsolescence due to advancements in AI technology. Jason asked about the possibility of non-computer-savvy individuals creating agents for repetitive data input tasks, to which Leon suggested using retrieval-based agents or platforms like Replit for smaller, more focused problems.

McP: Evolution and Business Opportunities

Leon explained the functionality of MCP, an interface between language models and APIs, and demonstrated how it can be used to create tools for non-programmers. He also discussed the evolution from GPT to MCP, noting that while GPT was easier to use, MCP is more powerful. The group explored potential business opportunities around non-code MCP development tools and AI cursor functionality for presentations. Leon shared a benchmarking website (beta.lm-arena) to help users choose models for different tasks, and SVCA requested Leon to share his slides for future reference.

Summary prepared by Zoom AI Companion