Large Language Models (LLMs)

Author

Andrei-Ovidiu Ersen
June 02, 2024

Large Language Models (LLMs)

LLMs is what people think about now when thinking about AI. It has become very popular because of ChatGPT. Also, talking with a LLM can seem like talking with a human, so this gives us the impression of AI being really intelligent.

But how does an LLM work?

Again, without going in technical details, it works the same as any other type of AI: it ”learns” from examples. Based on some example texts, it “learns” what word should come next.

But that can seem kind of weird, right?

But if you think about it, this is how we learned how to speak when we were children, by just listening to other people.

Still, LLMs have no logic behind them, they just complete with what makes sense. They fail in more complex tasks. For example, given a complex math problem, they will most likely output a random result.

How would a human solve such a problem?

By decomposing into smaller problems. That is where AI agents come into play.

AI agents

AI agents use LLMs, but instead of giving the answer directly, they have intermediate thoughts. They can also have access to tools. These can be databases, web searching, creating files and so on.

Check out my AI agents to understand more: https://atomicautomators.com/our-chatbots/

AI agents can thus become more efficient than humans in certain tasks. What we also need is to integrate AI agents in our workflows.

Check the AI chatbot on my website: https://atomicautomators.com/

These AI agents can be used in any task. What we need to know is how a human would divide the bigger problem into smaller ones.

For example, a complicated math problem can be solved the following way:

  • LLM writes a code for the computation;

  • it uses a tool to run it;

Thus, there are higher chances that the answer is correct.

Conclusion

To enable AI to solve problems in any domain, we need to have people specialized in that domain and people specialized in AI.

This is why I am creating a community, where people can learn about AI and discuss ideas for implementing it in their domain. We need as many people as possible to create the best products for every domain.

See you in the next newsletter,

Andrei