I liked your question so much I ironically asked Gemini about this.
Out of the 5 answers it gave, those ones seemed interesting :
>The Hardware Whisperer (Requires Simple Hardware):
>Setup: Requires a cheap microcontroller like a Raspberry Pi Pico or an ESP32/ESP8266, plus maybe an LED or a button.
>Task: "Make the LED blink out 'SOS' in Morse code when you press the button." Or: "Read the value from this simple sensor (e.g., temperature) and print it to the serial monitor only if it's above a certain threshold."
>Why it works: LLMs can generate MicroPython or Arduino C++ code easily. However, the user must deal with the physical wiring, installing drivers/firmware tools, uploading the code, and debugging why it's not working (Is the code wrong? Is the wiring loose? Is the board getting power?). This physical interaction layer is opaque to the LLM.
>Setup: Create a very simple text-based game (e.g., navigate a maze, guess a number with clues, a basic simulation) that runs locally. The game's state changes with each command.
>Task: "Reach the end of the maze / Win the game." The rules might not be fully explained, requiring experimentation.
>Why it works: The LLM can't play the game directly because it doesn't have access to the running process or its internal state. The user must interact with it, observe the output, and decide on the next input. They might use an LLM to suggest strategies or even write a script to automate playing the game (if the interaction pattern is simple enough), which itself teaches valuable skills.
Out of the 5 answers it gave, those ones seemed interesting :
>The Hardware Whisperer (Requires Simple Hardware): >Setup: Requires a cheap microcontroller like a Raspberry Pi Pico or an ESP32/ESP8266, plus maybe an LED or a button. >Task: "Make the LED blink out 'SOS' in Morse code when you press the button." Or: "Read the value from this simple sensor (e.g., temperature) and print it to the serial monitor only if it's above a certain threshold." >Why it works: LLMs can generate MicroPython or Arduino C++ code easily. However, the user must deal with the physical wiring, installing drivers/firmware tools, uploading the code, and debugging why it's not working (Is the code wrong? Is the wiring loose? Is the board getting power?). This physical interaction layer is opaque to the LLM.
>Setup: Create a very simple text-based game (e.g., navigate a maze, guess a number with clues, a basic simulation) that runs locally. The game's state changes with each command. >Task: "Reach the end of the maze / Win the game." The rules might not be fully explained, requiring experimentation. >Why it works: The LLM can't play the game directly because it doesn't have access to the running process or its internal state. The user must interact with it, observe the output, and decide on the next input. They might use an LLM to suggest strategies or even write a script to automate playing the game (if the interaction pattern is simple enough), which itself teaches valuable skills.