How Does Generative AI Work at a High Level

Transcript

How does generative AI work at a high level? I want to demystify how AI works from a ten-thousand-foot perspective.

Note that this is a guide that is intended to help lawyers spot legal issues relating to AI, so we are trying to abstract certain concepts about AI rather than focusing on pure technical details.

So whether you’re just curious or already incorporating AI into your daily work, understanding the basics can really help you make the most of it. So let’s jump in.

At a high level, generative AI tools like ChatGPT or Claude function on the basis of something called inputs that go into an AI algorithm and produce outputs.

This concept is key to how the AI delivers the responses we interact with every day. So let me explain a little bit more. Think of inputs as the information we provide to the AI.

Different AI platforms may have slightly different names for these inputs.

For example, for OpenAI’s ChatGPT, an input is data that you submit, like a question, a statement, or even just a keyword. For example, if you ask ChatGPT, “What’s the weather like today?” That’s an input.

With Claude, this input is actually called a prompt. So when you ask Claude something, you’re providing it with a prompt, whether it’s a question, a task, or just a thought to expand upon.

So whether it’s called an input or a prompt or something else, the basic concept is that whatever the user enters into the AI algorithm for it to start processing goes in. And then once that AI receives this prompt, it processes it using powerful algorithms trained on vast amounts of data, working to understand the context and generate a relevant response.

The response that the AI provides is called the output. In OpenAI’s terms, an output is the result of the AI process, such as the answer it gives to your question.

So when you ask ChatChippy Tea to draw a picture of a dog, the output is the image that it drew.

Think of it as a simple cycle. Input or prompt goes in, the AI processes it, and the output comes out.

So what makes this process special is how the AI learns patterns and understands the nuances of language, aiming to generate meaningful and helpful responses even without specific directions from you.

The AI can also learn and evolve based on the training data it is fed. This is a big departure from traditional rules-based programming, which we described in a different video.

This is where developers would need to code in specific rules for the program to follow in order to generate specific intended output. So I like to analogize the AI algorithm to a sausage maker. Once the input goes in, the output may not resemble the input at all. It is actually highly unlikely that you could reverse-engineer the output back to the original input.

This is because the output is generated based on, among other things, the data that the AI algorithm is trained on, the context of the question you asked, and the patterns that the AI algorithm detects. Newer versions of generative AI using voice prompts can even sense and detect humor and sarcasm, which could also affect how it creates an output based on your specific prompt.

The same AI algorithm can even produce different outputs for the same inputs if you enter the same input multiple times. This is due to the nondeterminism and probabilistic nature of AI.

And while both platforms strive to give helpful, insightful outputs, remember that these outputs are generated based on patterns the AI learned from existing training data.

This means that not everything the AI says is verified information or fact checked, and the responsibility for how we use those outputs rests with us, especially if we are utilizing AI in our legal work. So, to wrap up, the process of input or prompt leading to output is at the core of how these AI models work at a high level.

And if you’re interested in learning more about AI technology or ways to leverage generative AI tools in your legal work, make sure to subscribe and hit the notification bell. Thanks for watching.