Gist
Large Language Models have transformed how we interact with AI, but different models offer different capabilities, costs and levels of control.OpenAI's ChatGPT became popular as a powerful conversational AI service capable of generating content, writing code and maintaining conversations.
The phrase Large Language Models (LLMs) has become one of the most common terms in the AI world. We hear about new models almost every day, but I was curious about something else.
What actually makes one LLM different from another?
LLMs are trained on huge amounts of data and learn billions of parameters during the training process. At the core of these models are Transformer-based neural networks, which allow them to understand the context of a question and generate a response based on what they have learned.
The interesting part is that these models are not limited to answering questions anymore. They can write code, create content, summarize information, have conversations and even correct their responses based on the conversation.
But there is also a problem.
The same question can sometimes produce different answers depending on how we ask it. The model can also provide incorrect information or show bias in its responses.
That made me look at two of the models that were getting a lot of attention at the time: OpenAI's ChatGPT and Meta's Llama 2.
Let's start with ChatGPT
OpenAI introduced ChatGPT in November 2022, and it didn't take very long for it to become popular.
The idea was simple. Instead of interacting with a traditional application, we could simply ask a question using a prompt and have a conversation with the AI.
What made ChatGPT interesting was the range of things it could do. It could generate computer code, write essays, create poems and even come up with jokes that were not too bad.
The model could also remember parts of a conversation, respond to follow-up questions and make corrections when the user pointed out a mistake.
But it wasn't perfect.
Sometimes it gave incorrect information. Sometimes the response could be biased. And interestingly, asking the same question in a slightly different way could produce a completely different answer.
OpenAI continued improving the model, and the introduction of GPT-4 in March 2023 was another significant step.
But there was one thing that developers and businesses couldn't ignore.
Cost.
At the time, OpenAI's API pricing was based on tokens, with 1,000 tokens roughly representing 750 words. GPT-4 was significantly more expensive than the earlier models, which made the cost an important consideration for applications that needed to make a large number of requests.
And then Meta entered the conversation.
Meta Llama 2 changed the conversation
When Meta announced Llama 2 in July 2023, what caught my attention was not just the model itself.
It was the fact that Meta was making an open-source model available for both research and commercial use, subject to its licensing terms.
Llama 2 came in three variants with 7 billion, 13 billion and 70 billion parameters.
That gave developers another option.
Instead of depending entirely on a hosted AI service, organizations could take the model, customise it and build it into their own applications.
That flexibility is probably the biggest difference I see between the two approaches.
ChatGPT is a powerful, ready-to-use service.
Llama 2 gives developers more control over how they want to use and customize the model.
But that flexibility also comes with its own challenges.
Fine-tuning a model for a specific requirement can require significant computing resources, technical expertise and cost.
So open source doesn't automatically mean free.
There is still an engineering investment involved.
Security makes the comparison more interesting
When we start talking about using AI in commercial applications, performance isn't the only thing we need to consider.
Security and privacy matter just as much.
According to Meta's evaluation at the time, Llama 2 showed a lower proportion of information leakage compared with ChatGPT in the tests they conducted.
The study reported around 4% for Llama 2 Chat compared with 7% for ChatGPT.
These numbers are useful, but I wouldn't use them alone to decide which model is more secure.
Security depends on much more than the model itself. The application architecture, data being provided to the model, access controls and how the model is deployed all matter.
Still, it was encouraging to see security becoming part of the model comparison rather than being treated as an afterthought.
So, which one should we choose?
This is where I don't think there is a simple answer.
If I want a powerful model that I can access through an API and start building with, ChatGPT is an attractive option.
If I want more control over the model, the ability to customize it and the flexibility that comes with an open-source approach, Llama 2 becomes interesting.
The decision eventually comes down to what we are trying to build.
For some applications, the quality of the response will be the most important factor.
For others, the ability to customize the model or control where the data is processed may matter more.
And then there is always the cost.
In the business world, a technically impressive model still needs to make commercial sense.
My take
When I look at ChatGPT and Llama 2, I don't see one model simply replacing the other.
I see two different approaches to how we can use AI.
OpenAI showed how powerful a conversational AI service could become when the model, infrastructure and user experience are packaged together.
Meta's Llama 2 showed that there is another path where developers get more control over the model and can customize it based on their needs.
At that point, the interesting question is no longer which AI model is the best?
It is:
Which AI model makes the most sense for the problem we are trying to solve?
That is probably the question we should be asking every time a new AI model enters the market.
