Ep 03: I Love You! System Role + Temperature | Free AI Engineer Course | 8 Weeks

Main ideas / lessons conveyed

• The same words can produce very different responses depending on:
  • who the message is addressed to, and
  • what relationship/role the responder believes you have.
  • Example: Saying “I love you” to a girlfriend may get a warm response, while saying it to an office manager/teacher may trigger a formal/professional (and potentially problematic) response.
• Two key ideas in LLM prompting control this behavior:
  1. System role (sets the relationship/post/context)
  2. Temperature (controls creativity/randomness)

Concepts explained: “System role” vs “Temperature”

1) System role (what it is and what it does)

• System role defines the relationship or context the LLM should assume throughout the conversation.
• The video describes three roles in LLM chat:
  • User role: you (the person asking)
  • Assistant role: the model’s responses
  • System role: instructions that tell the LLM how to behave and what “role” it should take when interpreting the user’s messages

Key mechanism

• If you only send a message like “I love you” with no system instructions, the LLM may respond generically (e.g., “I love you too …”).
• If you first send a system instruction like “You are my loving girlfriend,” the LLM will respond in that relational tone.
• Changing the system instruction to “You are my strict office colleague and manager” changes the tone and appropriateness of the response (the video jokes about workplace policy consequences).

Real-world impact of system role (agent roles in organizations)

• In a real company building AI agents, you often want different agents to behave like different job roles.
• Example mapping of roles:
  • Developer: writes/tests code
  • Architect: designs overall system structure; focuses on design concerns (security, maintainability, etc.)
  • Manager: manages people
• By using system role instructions, the same LLM can answer from different professional perspectives:
  • With no system role, answers may be generic.
  • With system role, each agent answers according to its “job,” e.g.:
    • “Act as a senior architect” → warns about reliability/maintainability/security
    • “Act as a junior developer” → focuses on implementation-level details/bugs

---

2) Temperature (what it is and what it does)

• Temperature modifies the model’s creativity / randomness in its predictions.
• The video emphasizes:
  • the model is not a “truth-telling machine,” but a prediction machine that generates likely next text based on patterns.
• With temperature = 0, it tends to make the safest/most conservative choice.
• As temperature increases, it becomes more creative / more varied (more “randomness”).

Temperature range and behavior

• The video claims the practical range is typically 0 to 2 (and says 3 is invalid/not allowed for the demo model).
  • t = 0: safest / most straightforward answers
  • t = 1: somewhat creative, more variety
  • t = 2: highly creative / most varied, potentially less sensible

Examples used

• Wordplay/creative output example
  • For app names from the prompt “name food delivery apps,”
    • low temperature yields more common/safe suggestions (e.g., “Khana Khazana,” “Mantu Ka Dhaba”)
    • higher temperature can output stranger names (e.g., “Swiggy”)
• Storyteller vs doctor use case
  • Doctor agent: prioritize factual closeness; low creativity is preferred
  • Storyteller/media agent: creativity/randomness is desirable to generate engaging variety
• Product naming example
  • A brand founder might prefer a weird/creative name, supported by higher temperature

Methodology / instructions shown (video’s practical workflow)

Setup workflow for the “Day 2” coding demo

• Create a Day Two folder inside Week One.
• Copy the environment file (env file) into the Day Two folder.
• Recreate the virtual environment using a command of the form:
  • uv venv --python 3.11
  • (The video notes common formatting mistakes, such as missing dashes.)
• Activate the virtual environment (the video suggests checking the terminal prompt/path to confirm activation).
• Copy the prior script:
  • from hello_llm.py (or equivalent) into a new file named system_temperature.py.

Dependency installation troubleshooting during demo

• If the script fails due to missing imports/modules (e.g., grok), install required packages with uv add ....
• The video highlights common mistakes (typos, incorrect command formatting, duplicate words) and stresses:
  • Most failures are due to human errors in commands or imports
  • Read the error message and correct the command accordingly.

---

Prompting implementation steps (system role + temperature demo)

A) Demonstrate system role

• Create a new message dictionary for the system role, placed before the user message in the chat messages list.
• Example system role content (conceptual from the video):
  • “You are my loving girlfriend.”
• Then send a user message (e.g., “I love you baby”) and observe:
  • girlfriend context → affectionate reply
  • office colleague/manager context → professional/strict reply

B) Demonstrate temperature control

• Modify the API call/client so that:
  • temperature is passed into the chat completion creation call
  • value is in {0, 1, 2}
• Observe how output variety changes:
  • default temperature (when not set) → behaves like 0
  • temperature = 1 → more creative than 0
  • temperature = 2 → most creative/fanciful responses
• Also demonstrated:
  • trying temperature=3 results in an error (“not allowed”).
• Example task used for temperature:
  • “Suggest a name for my brand/food company,” later “my clothing company,” showing different fanciful outcomes.

Recap (explicitly stated in the video)

• System role: tells the LLM the relationship/post/context it should assume (e.g., girlfriend vs office manager vs architect vs junior developer).
• Temperature: controls creativity/randomness in the LLM’s output (useful depending on the use case like doctor vs storyteller).

Speakers / sources featured

• Speaker(s): The video speaker (subtitles mention Pratyush as a person being addressed/referenced, but no clearly confirmed separate narrator name is stated).
• Named individuals referenced in examples:
  • Pratyush
  • Jerick (used as an example responder)
• Software/source referenced:
  • LLMs / ChatGPT
  • Grok (used as the model/library in the coding demo)
  • Claude (mentioned as a comparable model)