Learning Loop ♾️ 05-2025

1. Top-K: Sampling Method (Limiting the Diversity of Choices)

Top-K is a sampling method in LLMs that helps you get consistent results.
Normally, if you ask the same question multiple times, the answer may change.

But when you set K = 2, the model only picks from the top 2 likely options — so the output stays stable.

Example

• I asked: What’s the color next to Orange?

  • Without any Top-K setting → the answer was Yellow.

  • With K = 200 → the answer changed to Red.

  

  Top K - Check Out The Full Chat

• With K = 2 → it went back to Yellow.

  

  Final Test - Top K

So, limiting diversity (smaller K) makes the model stick to the most likely answers.

Top-K vs Temperature

• Temperature controls the randomness of choices (I have explained temprature in detail in my 1st edition).

• Top-K controls the diversity of choices.

2. Perplexity – Set Source

This feature lets you control where your search engine looks.
I found it most useful when I wanted to understand sentiment around a topic.

This is where you can find it:

Why it matters

• Most LLMs don’t let you pick or control their sources.

• With Perplexity, you can direct the model to use a specific source.

• This makes fact-checking and context more reliable.

Use cases

• Understanding community sentiment or discussions.

• Pulling data directly from SEC filings — giving a clean, reliable fact summary.

💡 Note: Refer output report fromy my experiment on ACHR using Social and Finance source.

3. Summarize YouTube Videos with Gemini

This is a no-brainer use case.

We all come across videos with thumbnails that look super interesting — but they’re 3 hours long and you just don’t have the time.

That’s where Gemini comes in.

How to use it

Just drop the video link into Gemini with the right prompt, and it gives you a clean summary of the key points — saving hours of watching.

You can also ask it give you only real insight and avoid generic statement. This feature only work, if you are logged in from your gmail account only.