• AIpeeps
  • Posts
  • Learning Loop ♾️ 06-2025

Learning Loop ♾️ 06-2025

Your bi-weekly dose of learning AI, without the hype.

Hello peeps,

Hello everyone, and a very warm welcome to the 6th edition of AIpeeps!

What we’re covering today:

Top-P Sampling – probability-based sampling that makes text feel natural
ChatGPT GO (India) – new features, speed, and what I tested so far
Formatting in Excel – why it fails, what works, and community call-out
Wrappers in AI – why startups build on top of LLMs instead of building them

💡3 Curious things I learnt

1. Top-P Sampling

Top-P is another sampling method used by LLMs.

Here’s how it works:

  • You let the model choose from words that fall within a cumulative probability threshold.

  • Instead of restricting to a fixed number of top words (like in Top-K), Top-P looks at probabilities until the cutoff is reached.

  • This makes the text more natural, more human-like, and often more creative.

Top - P

👉 Last week: I explained Top-K — where you limit the model’s choices to K words.
👉 This week: With Top-P, you don’t limit the number of words, only the probability range.

Why it matters:

  • Top-K → great for consistency (predictable answers).

  • Top-P → great for fluency and creativity (varied but still coherent text).

Think of it as the difference between confining the model to a small menu (Top-K) vs letting it explore all dishes that meet a certain quality level (Top-P).

I’ve also shared some examples of how outputs shift when switching between these two:

Top P and Top K

2. ChatGPT GO

Yes — ChatGPT Pro is now available, but only in India 🇮🇳.
Makes sense, since India is the second-largest ChatGPT user base worldwide.

👉 Price: ₹399
👉 How to get it: Go to Upgrade → Select the plan → Pay via UPI

CHATGPT GO STEPS

What I’ve tried so far:

  1. Projects

    • Independent memory, documents, and instructions grouped together.

    • Useful if you’re working on:

      • a business idea

      • a newsletter (like this one)

      • or even a long-term side project.

    • It keeps context separate so you don’t mix work with personal prompts.

  2. Image generation

    • Noticeably faster compared to the free tier.

    • Took me under 60 seconds to generate an image.

    • As a free user, the same prompt used to take much longer.

💡 Note: That’s my initial exploration. In the next issue, I’ll cover more on add-on features in detail.

3. Formated Excel Output with ChatGPT

This one was a fail (at first).

When I tried asking ChatGPT to make Excel sheets with formatting + formulas, it only gave me raw data — no colors, no fonts, no neat formatting.

Here’s what I figured out:

  • You must spell out every detail: colors, fonts, column widths, and formulas.

  • If you just say “make it nice,” it won’t.

  • The more specific you are, the closer you get to the actual formatted sheet.

💡 I’m still working on this.
If you’ve managed to crack the perfect prompt for formatted Excel outputs, reply back — I’d love to share community wins here.

My prompt for below output - “create a table of SIP of 10K for 15 years in excel with nice formating having blue color and Arial 10 ; formulae of sum should also be there”

Formatted Output

🕵🏻Decoding the Jargon

Wrappers in AI

A wrapper = the same core product, just packaged differently.
Think about the food industry → same ingredients, different packaging or branding.

In the AI world:

  • There are only 5–7 core LLM providers (OpenAI, Anthropic, Google, etc.).

  • Most new AI startups are not building their own models.

  • Instead, they wrap around these LLMs with:

    • Better UI/UX

    • Domain-specific data (RAG)

    • New workflows or integrations

Example: platforms like Lovable don’t create the LLM — they wrap it with features that make it more usable for a niche.

So when you hear about “new AI tools,” most are wrappers built on top of the same few LLMs.

Try it Out this week:

  • Create a Project in ChatGPT Pro → test independent memory + docs.

  • Brainstorm wrapper ideas you could build around LLMs → think startup potential.

  • Use Top-P in your next prompt and compare it with Top-K → see the difference for yourself.

Hope you learned something new today!

🚀 “Found this useful? Pass it on to your close circle — knowledge hits different when it’s shared.”

Till next time,

Curiously yours,

Gaurav Jain

AIpeeps

Reply

or to participate.