MCPs are Boring (or: Why we are losing the Sparkle of LLMs) - Manuel Odendahl Introduction and Background 00:01
Manuel Odendahl introduces himself as a software engineer with 25 years of experience and discusses his journey with language models (LMs) since 2022.
He highlights his focus on backend coding and tool calling, emphasizing the capabilities of LLMs in generating structured code.
The Magic of Tool Calling 01:13
LLMs can produce language that effectively calls APIs and functions without needing to validate schemas.
The process involves the LLM determining which tools to use based on the context provided, such as checking the weather in San Francisco.
Challenges with Multiple Tools 03:23
When LLMs are given too many tools, they may struggle to decide which to use, leading to incorrect tool calls or schema misunderstandings.
The inefficiency of needing to repeat arguments already present in the context is highlighted, leading to increased token usage and costs.
Limitations of MCPs 07:30
MCPs (Machine Communication Protocols) are seen as limited and boring, failing to leverage the full potential of LLMs.
Suggestions for improvement include passing chat history and metadata to tool calls to enhance efficiency.
Enhancing User Interaction 10:12
Proposed improvements include allowing users to edit tool call inputs and outputs through a user-friendly interface, rather than relying solely on text windows.
The ability to create dynamic UI elements for user adjustments is emphasized.
The Power of Code Generation 12:30
LLMs excel at generating code and can create complex queries and functions, making them powerful tools for engineering tasks.
Manuel discusses how LLMs can create useful libraries and functions from simple queries, enhancing their utility.
The Future of LLMs and Tool Calling 18:13
The speaker advocates for the use of eval functions instead of traditional MCPs, suggesting they yield better performance and save resources.
He emphasizes the importance of creating reusable libraries and functions rather than limiting to static tool calls.
Conclusion and Call to Action 26:23
Manuel concludes that LLMs are capable of infinite creativity and encourages engineers to leverage LLMs to create versatile tools and solutions.
He urges the audience to think beyond basic tool calling and rediscover the "sparkle" of LLMs, inspiring a more innovative approach to their usage.