You show up to a party, and someone asks you how you can see why an AI reached a specific conclusion and you answer:

• Reasoning Trace – A detailed, step-by-step explanation of how an AI model arrives at a conclusion. It helps in understanding the model’s decision-making process, improving transparency and debugging.

Then, because apparently you’re in SF, someone asks about how to refine reasoning even further and you mention:

• Chain-of-Thought (CoT) – A technique where AI models generate intermediate reasoning steps to improve complex problem-solving, similar to how humans think through problems step by step.

At a different party, geez people just love talking about AI, you’re asked about tokens and you answer with some additional context around limits:

• TPM Limits (Tokens Per Minute Limits) – A restriction set by AI providers to limit how many tokens (words, characters) can be processed per minute, helping manage server load and prevent excessive use.
• MCP (Maximum Context Processing) – A term that may refer to the maximum amount of context an AI model can handle in a single input. Some models have limits on the number of tokens (words, characters) they can process at once.

The next thing you know you’re talking about efficiency gains.:

• Structured Outputs – AI-generated responses that follow a specific format, such as JSON, tables, or key-value pairs, making it easier for systems to process and integrate the results.
• Prompt Caching – Storing frequently used prompts and their responses to reduce computation time and improve efficiency in AI applications, especially for repeated queries.

And just like that, you realize you’re AI’s human in the loop.