O3 Can Use Tools While Thinking — But So Could GPT-4?

Key Concepts:

• 03 (Hyped new model): A new AI model being promoted by influencers.
• Tool Calling: The ability of a model to use external tools within its reasoning process.
• GPT-4: OpenAI's previous generation model, released in 2023.
• Code Interpreter: A feature of GPT-4 that allows it to write and execute Python code.
• Tokens: The basic units of text that language models process.

Main Argument:

The speaker questions the novelty of 03's "tool calling" ability, arguing that GPT-4, with its Code Interpreter feature released in 2023, already possessed a similar capability.

Comparison of 03 and GPT-4's Code Interpreter:

• GPT-4 (2023): The speaker describes how GPT-4, using the Code Interpreter, could analyze CSV files, answer questions, and iteratively write and execute Python code to achieve the desired result. If the code failed, GPT-4 would attempt to fix it and retry. If successful, it would analyze the results and potentially write more code.
• 03 (New Model): Influencers are promoting 03 for its ability to call tools within its thinking process. The speaker argues that this is not fundamentally different from GPT-4's Code Interpreter functionality.
• Token Stream Analogy: The speaker explains that both models operate on a stream of tokens. In both cases, some tokens trigger tool calls, which return data that is then converted into tokens and influences subsequent token generation.

The Speaker's Confusion and Question:

The speaker expresses confusion about how 03's tool calling is a novel feature, given GPT-4's existing capabilities. The speaker acknowledges that 03 might be better trained to perform tool calling more effectively, but maintains that the underlying concept is not new.

Call to Action:

The speaker invites viewers to share their thoughts and perspectives in the comments, acknowledging the possibility of missing something or being incorrect in their assessment.

Conclusion:

The speaker is skeptical of the hype surrounding 03's tool calling ability, suggesting it is not a groundbreaking innovation but rather an improvement upon existing functionality present in GPT-4 since 2023. The core argument is that both models achieve similar results through the manipulation of tokens and the execution of external tools.