So far, we’ve learned that LLMs like Claude are “taught” through the process of machine learning - the technological process that underpins the input. But what about the output? How does Claude generate responses?

It does that thanks to technology known as generative AI.

What’s generative AI?

While machine learning and deep learning are generally focused on understanding or interpreting data, generative AI goes a step further to create new data that didn't previously exist, mimicking the original data in a novel way. It's like an artist who creates new paintings inspired by all the art they've seen before.

Generative AI is used for a variety of creative and practical tasks, such as creating realistic images and videos, generating text for chatbots and virtual assistants, developing new video game environments, and more.

So what happens behind the scenes when I chat with Claude?

When you type a query into Claude and you get a response, there are a few things happening “under the hood” to make it all work seamlessly:

1. You ask a question, e.g. “What was Steve Jobs's philosophy on making innovative products?”.
2. Claude uses natural language processing (NLP) to understand the meaning, context and intent behind the question.
3. Claude then uses knowledge retrieval to find the most relevant information from its training data to help answer the question.
4. Once Claude has gathered the necessary information, it uses a process called "language generation" or "text generation" (this is where generative AI comes into play) to put together a coherent and meaningful response.

Does it ever go wrong?

LLMs have been trained so extensively that they do a pretty impressive job of providing accurate responses.

But this is relatively new technology, and it’s certainly not foolproof.

Sometimes AI will give an inaccurate or misleading answer, known as a hallucination, due to:

• Ambiguity in the input query, leading the AI to make inaccurate assumptions.
• Gaps in the AI's knowledge, especially if the required information is missing from its training data.
• The AI's design limitations, such as its inability to access real-time information or understand context beyond its training cut-off date.

What do I get if I use Claude over other AI assistants?

While Claude shares many capabilities with other AI assistants like ChatGPT, there are often differences:

• Quality of model: Newer, more capable models are released at different times, so when ChatGPT’s GPT-4 came out it was the best-performing, but since then Claude Opus 3 came out and seems to be the current best-performing.
• Features: While both ChatGPT and Claude can analyse images, produce code and possess other similar features, ChatGPT has a richer feature set as of now. But new features are being added to all assistants all the time, so the differences are often minimal.
• Claude utilises the Constitutional AI guidelines to ensure its outputs adhere to ethical standards and user expectations, which might differ from the principles guiding other AI models.

With the background knowledge out of the way, let’s get into use Claude for everyday tasks!

Next up in this course: ‍

Getting started - Using Claude for everyday personal tasks