What does it take to be an AI whisperer?

What does it take to truly understand how AI models think? Anthropic researcher Amanda Askell shares what it means to be an “LLM whisperer.”

Amanda Askellguest

Dec 14, 20250mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Amanda Askell on empirically learning to “whisper” to LLMs

Effective LLM “whispering” starts with a willingness to interact with models frequently and study many outputs to build intuition for their behavior.
The work is fundamentally empirical: you learn by experimenting with prompts and observing how responses change.
Clear communication of your concern or intent in the prompt is central, because small wording choices can lead to unexpected model interpretations.
When the model responds unexpectedly, you can either ask it to explain its reasoning or diagnose which part of your prompt caused the misunderstanding.
Askell finds this iterative process intellectually compelling because it reveals surprising depth in how models respond to different inputs.

IDEAS WORTH REMEMBERING

5 ideas

Treat prompting as an empirical craft, not a purely theoretical one.

Askell emphasizes that progress comes from repeated trials—writing prompts, observing outputs, and refining based on what you see rather than relying on abstract rules alone.

Build “shape of the model” intuition through volume and attention.

Regular exposure to output after output helps you recognize patterns in how the model reacts to different phrasing, constraints, and contexts.

Write prompts as clearly as possible to reduce misinterpretation.

She frames her approach as explaining an issue or concern to the model with maximum clarity, since ambiguity or missing context can easily derail results.

Unexpected outputs are diagnostic signals, not just failures.

When the model does something surprising, Askell recommends either asking the model why it responded that way or investigating which prompt element led to confusion.

Iterate by pinpointing the misunderstanding trigger in your input.

A key skill is tracing the output back to specific wording or assumptions in the prompt, then rewriting to remove the failure mode you uncovered.

WORDS WORTH SAVING

4 quotes

One thing is just, like, a willingness to interact with the models a lot and to, like, really look at output after output, and to use this to get a sense of, like, the shape of the models and how they respond to different things.

— Amanda Askell

To be willing to experiment. It's actually just, like, a very empirical domain, and maybe that's, like, the thing that people don't often get.

— Amanda Askell

I try and explain, like, some issue or concern or, or thought that I'm having to the model as clearly as possible, and then if it does something kind of unexpected, um, you know, you can either ask it why, or you can try and figure out what in, in the thing that you said caused it to kind of misunderstand you.

— Amanda Askell

I find the work extremely interesting. I think it highlights really interesting depths to the models

— Amanda Askell

High-volume model interactionEmpirical prompt experimentationPrompt clarity and intent expressionDiagnosing unexpected outputsAsking the model “why”Developing intuition for model behaviorModel depth and interpretability

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.