What is GPT really is ?

Building a “Thinking” Model from a “Non-Thinking” Model Using Chain-of-Thought Prompting ?

Aditya Bhojane — Fri, 15 Aug 2025 08:00:21 GMT

Nowdays most Ai Models already have COT, capability of resson internally it show Thinking step by step and after few min it gives as accurate answer

Why “non thiking” models fail on complex tasks? the reason is very simple
they try to jump into conclusion in one shoot i mean on one go and make mistakes but meanwhile in think approach it breck down task into steps and process it and then give us final output

Now, How to add ‘Thikning” (COT) to ‘non thinking model’
1) We have to impliment chain of thoughts prompting where we have to give good and big context around how to solve a perticular prompt step by step for example :
“Think step by step before answering, show each step of your thinking and then provide a final answer “
the provide examples like ( few shot ) :
Q: If a rectangle has length 8 cm and width 5 cm, what is its area?
A:
Step 1 - THINK : Area = Length × Width
Step 2 - THINK : Area = 8 × 5 = 40

Final Answer: 40 cm²

NOTE : more step can be added its just and example

2) There is one more way to add thiking its same like previous one but one extra thing we have hide model thinking like dont show user that model is thinking

EX : “Think step-by-step but do not show reasoning in the final response. Only output the final answer.”

3) Verification thinking :
To make it more better we can add another model in between as a person like role so that it can verify the answer or a perticular step it will helps us more it more accurate and improve consistancy

Importance of system prompts ?

Aditya Bhojane — Fri, 15 Aug 2025 07:37:48 GMT

A system prompt is basically set of instructions, rules its kind of context we provide to the AI model like a 1st message it acts as how we provide context to it for example persona

There are different types of prompting
1) Zero-Shot Prompting: The Direct Approach :
so this is a most common format we simply ask the model to perfrom a task without giving any prior example, i mean just intructions this we use mostly for straightforward task the model already understand

2) Few-Shot Prompting: Learning by Example :
In this type of prompting we give several examples before actual task for example if we want to make an translator then we have to give task like english ‘go’ → Hindi ‘jaao’ and english ‘How are you’ → hindi “aap kaise hai ?’ kind of like that so the it understand the context well and this type of prompting we use when we want high controll over style and structure

3) Chain-of-Thought (CoT) Prompting: Showing Your Work :
So this is very important type COT, which basically ask the AI to break task step-by step before giving final answer its same like DeepThinking now days most of the Ai model have a feature of DeepThink so you can related this with that, here also we have difine context give lots of examples so that it perform well in problem solving and logical task. most benifite of this type is basically it keep cosistancy and much more accurate as compair to others

4) One-Shot Prompting :
Its just an another type where we give one example before actual task same like few short but here only one example for small context around the assistant

5) Role-Prompting :
This type is basically use to add a persona/role to an Ai for exmple your are a hight level coding expert which have strong understanding of javascript like that we can also set the tone, how he should interact and domain specific task it keeps style and response consistant

Why Prompt Design Matters ?
i think there are few reasons
1) irrelevant responses
2) inconsistent tone
3) confusion
4) It improve user satisfaction
5) good results == lower cost

How Ai internally works ?

Aditya Bhojane — Wed, 13 Aug 2025 12:38:04 GMT

Tokenization : Convert any data like text ( or any other type depends on model ) into a number that model ( software ) can understand and there are methods like BPE, Unigram again it depends on model to model how they tokenize Ex : Hello -> [73, 1542] can be look like this ( token IDs)
Vector Embeddings : This turn each token into a high dimensional vector it can be ( ex:768, 1024, 4096 dimensions ) also make sure put some kind of relationship into it for example if give input like cricket, football and hocky then its sports so it make sure this have relationship so put it close into embedding space for better understanding here is a diagram of it
Positional Embedding : Transformer look token all at once so position also matter for us for example "dog jumped on person" and "person jumped on dog" so just by changing position we have different meaning so to prevent that we use positional embedding
Transformer Blocks : Each block process the sequence to understand context better and that's why many block stacked inside block and do self attention like each token create Query, Key and Value vector based on that it calculate attention score lower marks stops tokens from seeing in future, it repeat this process so many times it stabilize values
Output projection ( Decoding ) : vector block have matrix where based on probability it try to pick best output as per matrix
Detokenization : Convert final list of token into IDs into human readable format

What is GPT really is?

Aditya Bhojane — Tue, 12 Aug 2025 10:16:51 GMT

magine you’re talking to someone and you say:
“Tell me a scary story!”
They think for a moment and start telling you a spooky tale.

But then you change your mind and say:
“Wait! I want a poem instead!”
And they stop the story and start saying a fun little poem.

Now imagine this person can do anything you ask:

Answer questions like “Why is the sky blue?” ☁️
Tell jokes, sing songs, or help with homework
Always listen and give the right answer

That’s what GPT is like!
GPT Is a Magic Toy
It’s like a toy with a brain that:

Talks to you
Listens to what you say
Helps you with stories, poems, answers, and more!

But remember—GPT is not a person.
It doesn’t eat, sleep, or feel things.
It’s a smart helper that knows lots of stuff