How modern Large Language Models ( LLMs ) do Magic

Intro

Okay so how do Large Language Models get the current stock price of a company or something similar? Why it's not as intelligent as it seems.

( This is running on my old 8GB RAM laptop and it's not that slow. Innit? )

So what just happened there?

The Language model clearly displays the stock data after I ask it so how does it do it?

Explanation

• Well there is an LLM trained and / or prompted to display text in JSON format with certain parameters based on certain keyword used. The JSON then ( could* we will get into this when we make an agent system ) gets used with software ( That YOU designed ) as configuration files basically. Also in this specific case the LLM does some autocorrect on whatever you did to match it with some actual stock ticker symbol and to give you some information.

• The software sends a output based on the configuration file the LLM sent, which will again depends on what you sent to the LLM and how you configured the software to utilise the parameters in the first place.

• And finally you get an output that looks like an LLM response but it's a combination of how you setup and configured everything to interact with each other.

So basically when you ask an LLM a question regarding the Stock price it sends some of the keywords you mentioned as parameters to the software / system / service it is linked to ( THAT YOU'VE CREATED ) and the software displays the information in a clever way. ( IN THE WAY YOU DESIGN THEM TO INTERACT, no magic guys sorry... Overengineered? Yes ).

TL:DR

Basically the LLM responds with a configuration file and some additional text and the software in the background displays and formats everything for you in a clever way.

You can see how this is overengineered and maybe a bit unreliable and yes there are better ways of getting and interacting with information.

Basically the Magic of LLMs is in how good of a service you have built to be able to take LLM output as input and do something with it.

I do think there is actual utility for this especially as LLMs get smaller and better. But... For now...

You could just :

# Import necessary libraries
import yfinance as yf  # Import yfinance library for fetching stock data

# Function to fetch stock price based on company name
def fetch_stock_price(company_name):
    try:
        # Fetch stock data using yfinance ( Read the docs )
        stock_data = yf.Ticker(company_name)  # Create a Ticker object for the specified company_name
        current_price = stock_data.history(period='1d')['Close'].iloc[-1]  # Get the closing price of the stock for the last day
        return current_price  # Return the fetched stock price
    except Exception as e:
        print(f"Error: {str(e)}")  # Print an error message if fetching data fails
        return None  # Return None if fetching stock price fails

# Main program
if __name__ == "__main__":
    print("Welcome! This program fetches the current stock price for a company from Yahoo Finance.")
    
    while True:  # Start an infinite loop
        # Prompt the user to enter a company name
        company_name = input("Enter the company name or 'exit' to quit: ")
        
        if company_name.lower() == 'exit':  # Check if user wants to exit
            print("Exiting the program.")
            break  # Exit the loop and terminate the program
        
        # Fetch the stock price based on the company name entered
        stock_price = fetch_stock_price(company_name)
        
        if stock_price is not None:  # Check if stock price was fetched successfully
            print(f"The current stock price of {company_name} is ${stock_price:.2f}")  # Print the fetched stock price
        else:
            print(f"Failed to fetch stock price for '{company_name}'. Please check the company name and try again.")
            # Print an error message if fetching stock price failed

To run this paste the code into a file and make sure the filename ends with .py and run :

python filename_orwhaeveryouwannacallit.py # Or follow the steps on how to execute Python code whichever way you usually do it

And then you can do something like autocorrect yourself or make a dictionary file to match Company names with ticker symbols because this will use the company / ticker symbol not name (This is how the yfinance library works). Now I know it's not an exact replacement and you'd probably have to know a little bit about it before you just ask for the stock price but I think it's much more efficient than training an LLM paying around millions of $ for GPUs and then having to program something to take LLM output as input.

Conclusion

The best way to utilise the magic would be to have some agent system with a Big desicision making LLM that talks to other LLM " magicians " as tools when he needs to use them, that's a whole topic as well... And it's far from reliable. Even the best LLM ( OpenAI's GPT-4 ) is laughably unreliable let alone the smaller models.

But we will be making this type of system soon.

This was a quick little tease video and post! If you guys are interested in this send your questions and in the next posts we will go over everything, properly and I will make a little tutorial on how to setup everything and I'll give away the code. It's a bit too much for one engadging blog post and I want to leave you wanting for more!