- AIpeeps
- Posts
- Learning Loop ♾️ 08-2025
Learning Loop ♾️ 08-2025
Your bi-weekly dose of learning AI, without the hype.
Gaurav Jain
October 04, 2025 • Estimated Reading Time: 5 minutes
Hello peeps,
Welcome back to your AI decode session—this is #8!
What we’re covering today:
Context Limit → why ChatGPT forgets new chats but remembers old ones
Comet → the browser that feels like an AI assistant
Create Your First GPT → my WealthScore project and how I built it step by step
Latency → time taken to response by LLMs in detail
💡3 Curious things I learnt
1. Context Limit — Why ChatGPT sometimes forgets recent chats but remembers old ones
Have you ever felt that even after weeks of trying, ChatGPT still doesn’t catch the context of your ongoing conversations?
Yet, strangely, it remembers something from a very old post.
That gap might be happening because ChatGPT’s memory is limited — once it’s full, newer context doesn’t stick.
I’ve already shared earlier how you can delete memory, but this time I realized:
👉 improving context smartly via memory for better outputs is a fresh learning.
Try it yourself. See, if ChatGPT now picks up context from your latest chats as well.
2. Comet — More than just a browser
The UI/UX of Comet instantly gives a premium feel — definitely more than just a browser. Here’s how it starts:
Comet
So far, I noticed many features and options feel similar to Google.
But one thing that completely stands out → “Automate the Click” in the AI age.
I even planned my day with this assistant (see video).
The Assistant took 3 minutes to find restaurant, mall and route back to home. I didn’t even touch the mouse.
Obviously, it’s just a start — but the implications can multiply into 100x more use cases.
Here, you can Download the Comet.
3. Create Your First GPT — My WealthScore project
Yes, it’s time. I decided to create my own GPT.
Coming from a finance background, my starting point had to be something in my niche: WealthScore.
What it does:
Gives you a score based on ratios and models.
Uses uploaded RAG docs for more specialized results.
Even compares your income level region-wise.
Here’s how output looks like:
WelathScore Result
How I built it (with screenshot)
Go to Explore GPT → hit Create.
Follow the steps, upload docs, add instructions.
I refined it through 6+ iterations with GPT’s help + my own vision.
Custom GPT
Best part? When I asked it for a region comparison (via RAG not search result), the output was spot on.
This is still evolving and I am working on improving output and overall experience of user — but you can already try it.
👉 Give it a spin and share your feedback.
🕵🏻Decoding the Jargon
Latency
At a simple level, latency means the time it takes from when you send a prompt to when you start seeing the model’s response.
But in LLMs, latency has three layers:
Request latency (end-to-end)
Total time from hitting enter → full response finishing.
First Token Latency (FTL)
Time until the first word appears.
Depends heavily on model size and infrastructure.
Token generation speed (throughput)
After the first word, how many tokens per second the model can generate.
⚡ Rule of Thumb
Small, optimized models (GPT-4o, Claude Haiku, Gemini Flash) → Faster FTL (300–600 ms), high throughput.
Large, reasoning-heavy models (Claude Opus, GPT-4 Turbo, Gemini Pro) → Slower FTL (1–2s), lower throughput.
Try it Out this week:
Create a GPT for your wild idea and share with me link to try
Delete your ChatGPT memory to improve the context
Download Comet
If you learned something today and liked the newsletter — take a screenshot and post it on LinkedIn with your honest feedback. Tag @Gaurav Jain (or drop the link in a DM) — I’ll pay for your 1-month ChatGPT Plus subscription (only for existing subscriber till 04.10.2025).
Share the love, flex what you learned, and let’s grow this convo. 🚀
Please feel free to share any questions or topics you're particularly curious about.
Hope you learned something new today!
Till next time,
Reply