Matt Stauffer sits down with Sam Rose, developer educator at ngrok, to unpack what’s actually happening inside the AI models we use every day. Sam breaks down ideas like parameters, gradient descent, embeddings, and vector search in a way that makes them intuitive, using simple analogies such as converting Celsius to Fahrenheit. Along the way, they explore Ethan Mollick’s concept of the “jagged frontier” of AI capability and discuss why benchmarks like SWE-bench don’t always tell the full story.

The conversation wraps with a thoughtful look at the ethical tensions around modern AI: intellectual property concerns, the massive financial momentum behind these systems, and what it means to keep using these tools thoughtfully even when you’re still wrestling with their implications.

Matt Stauffer on Twitter
Tighten Website
Sam Website
Sam on Twitter
Prompt Caching Blog Post
Co-Intelligence Book
Benchmark Blog Post
Distributed Representations of Words and Phrases and their Compositionality
Stephen Welch on Twitter
Welch Labs on YouTube
Julia Turc on YouTube
Build a Large Language Model (From Scratch) Book
Nathan Lambert
Aaron Francis Website
Faster.dev
Mitchell Hashimoto
Armin Ronacher
My AI Skeptic Friends Are All Nuts Blog Post

-----

Editing and transcription sponsored by Tighten.

Beyond “Spicy Autocomplete”: What AI Models Actually Do