* The LLM's prompt contains instructions to explore the prompt space and think about how it can wrangle the data to do it's task.
  * It is like how data scientists working on a fresh CSV dump of a housing prices dataset will print out random things into a Jupyter notebook to _understand_ what they are dealing with.
  * While exploring, the LLM can also create new variables inside the Python runtime that contain important transformations of the data!
  * Remember, Python variables persist across different REPL execution calls. **I keep coming back to the Jupyter Notebook example because you must make this connection.** Each time the LLM writes a block of code and executes is equivalent to us humans writing a block of code and executing a cell!


Here is an example of RLM analyzing transcripts from Lex Fridman podcasts:
[![AVB](https://pbs.twimg.com/profile_images/2015375309147611136/WKvfQ-oV_normal.jpg)](https://x.com/neural_avb/status/2023387617891582018)
[AVB](https://x.com/neural_avb/status/2023387617891582018)
[@neural_avb](https://x.com/neural_avb/status/2023387617891582018)
·[Follow](https://x.com/intent/follow?screen_name=neural_avb)
[](https://x.com/neural_avb/status/2023387617891582018)
New RLM trajectory that blew my mind! I will use this one as the main example in the YT tutorial. I passed in a CSV containing transcripts of 320 episodes of the Lex Fridman podcast and asked it to find what his first 10 ML guests had to say about AGI. The context had[Show more](https://x.com/neural_avb/status/2023387617891582018)
[Watch on X](https://x.com/neural_avb/status/2023387617891582018)
[1:22 PM · Feb 16, 2026](https://x.com/neural_avb/status/2023387617891582018)[](https://help.x.com/en/x-for-websites-ads-info-and-privacy)
[569](https://x.com/intent/like?tweet_id=2023387617891582018)[Reply](https://x.com/intent/tweet?in_reply_to=2023387617891582018)Copy link
[Read 33 replies](https://x.com/neural_avb/status/2023387617891582018)
Example explorations or transformations of context can be:
  * LLM extracts an underlying CSV structure and puts the data into a pandas dataframe to process easier later
  * The LLM extracts specific sections from a markdown file and creates a dictionary of subchapter_title -> subchapter texts
  * The LLM issues regexes or find statements to search for keywords within the context (basic keyword search)
  * The exploration stage is all about distilling the complete prompt into smaller, useful variables.


For our Problem 1, though, the task is straightforward, so the LLM's exploratory task is rather easy.
markdown