AI - What is AI Agent ? Aii Eye - 1
Intent + Action = Functionality --> Intent = Agent --> Action = Tools LLM Model What is LLM ? Trained Data & Code ? Eg., Voice Model & vLLM Model What is Model ? Compressed form of numbers. Number is lik
Intent + Action = Functionality
--> Intent = Agent
--> Action = Tools
LLM Model
- What is LLM ? Trained Data & Code ?
- Eg., Voice Model & vLLM Model
- What is Model ? Compressed form of numbers.
- Number is like dataset , which is huge in number in Petabyte --> its A petabyte (PB) is roughly 1,000 terabytes (or 1 million gigabytes).
- Data --> Train --> weights
- What is weights ? Its like some equation.
- y = mx + c --> Line
- y = ax2 + bx3 + cx + 2 [ a , b , c are weights ]
- Then what is weight ?
- 2 & 0 are the deciding entity.
- For each and every entity or variable or weights , it gets changed.
- Like that for the Model --> we are going to get WEIGHTS and its getting trained with good amount of data and its having parameters.
- Y = mx + c → இது வந்து சின்ன ஈகுவேஷன், இதுல இருக்குற parameters கம்மி. So it's easy to achieve. y and X கு , value கொடுத்தாச்சு , so m & c are the parameters. அதனால Y value ஈஸியா அச்சீவ் பண்ணிடலாம். With many possibilities.
- If the equation is like y = ax2 + bx3 + cx + 2 [ a , b , c are weights ]
- I will add millions of parameters , it will make the y value easy to be achieved. நிறைய parameters இருந்தா , its ஈசி டு achieve.
- These parameters are given weights , that is the one the model is intent to give.
- So eg., Gemini , ChatGPT are nothing but the weights.
- So Model is nothing but the numbers. Weights are numbers.
- Model is going to give only EQUATION.
Numbers
- I told , Models are nothing but NUMBERS.
- Whether its a decimal number or whole numbers.
- Most of the model will have 32 decimal points. After point , there will be 32 points.
- All these 32 points will run in the RAM.
- Storage allocation will be based on this.
- GPT 120b --> means 120 billion parameters are nothing but the a, b , c in the equation.
Conclusion
- Model --> Compressed form of numbers going to have the weights , billion parameters are like a,b,c in the equation.
What does a Model do ?
- Can we train a model ? No , its inspired from a Architecture called TRANSFORMER.
- Model --> Its going to have a output and a input.
- Again --> whats does a model do ? Probability/Possibility of next word depends of the input. ஒரு output வருது, but அத நீங்க expect பண்ணல. So can you control the output ? Eg., Helloனு type பண்றீங்க so output is coming as “Helloworld” which is more probability but you don’t want this.
- 0.60 , 0.20 , 0.20 ... are the probabilities.
- You can control the output by below parameters:
1. Temperature
2. Top K
3. Top P
Temperature
- Value is bounded.
- 0 to 1.
- Using this we can control the thought process or choosing the next NICE word.
Top K
- How many number of words I need to see.
- உள்ள Model இருந்து நிறையா output வருது, இதுல TOP K குடுத்தா , for example Top K = 3 குடுத்தா , first மூணு அவுட்புட் மட்டும் தான் எடுத்துட்டு வரும். இதுக்கு மேல Temperature act ஆகும்.
Top P
- Probability Value
- Previously we have mentioned --> 0.60 , 0.20 , 0.20 ... are the probabilities. These are called Probability Values.
Now , can we control a MODEL ?
- Yes , to a certain extent. With the below:
1. Temperature
2. Top K
3. Top P
- But all these parameters will go along with the prompt.
- So prompt is also important.
Notes
- Quantisation --> I need this model but I can't run 32 points , then Quantisation comes into picture.
- AI ( If you asking any question in Gemini ) --> You will not directly interact with Model --> It will go to the process of ( parse , store , quadrilles then model ) --> Its like Engineering Pieces , you are not directly touching the model.
TBD
- How you know that this value should be given to Top P ?
- What is the difference between 32 decimal points and parameters ? Whether both are same ?
📰 Read the original article on Dev.to AI
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.







