The Recipe for Creating a Free AI that's Always at Your Service

Ingredients:

• Several billions of billions of phrases and texts from different online sources (newspapers, books, official documents, online conversations, websites, etc.).

In passing, thank in thought the journalists, bloggers, and other content creators who have so well structured their content, to be properly referenced on Google... a breeze for an AI to "borrow" (without crediting the authors).

• Algorithms or systems allowing the creation of large neural network architectures.

These systems have become quite common, and are indeed very widely used by the Open Source community. You can find this product in any AI accessory store.

• Powerful computer servers for performing large amounts of calculations.

Microsoft (one of OpenAI's investors, the company that develops ChatGPT) provides strong computing power with its Azure cloud. ChatGPT must be heating up several Olympic pools at the moment.

• NLP (Natural Language Processing) algorithms for generating answers to questions or texts.

Preparation:

1. Start by gathering large amounts of data from the lush jungle of the Web. For the best results, make sure to explore a variety of sources.
2. Sort through this data, keeping as much quality content as possible to train the AI (think about Tay, the AI that became racist and sexist on Twitter in just one day!).
3. In another container, initialize a neural network tailored to learn and integrate all previously gathered knowledge.
4. Place this mixture on powerful computer servers so the AI can grow without lacking resources. Remember that aggregating and understanding and manipulating knowledge requires a lot of energy and mathematical operations.
5. Sprinkle the preparation with natural language processing algorithms. These human-machine interfaces allow the machine to understand the meaning of words, translate them into English, and carry out the requested action.
6. Integrate your data incrementally to train your model. This could take a few days to a few months depending on the extent of data to be ingested by the AI.
7. Chef's tip: In Kenya, it is possible to train your AI for $2 per hour to become more ethical. This is a small trick used by OpenAI, as the Times reports.
8. Finally, test and improve your model by using additional data and indicating good and bad results. The AI learns over time... bon appétit!

CASE STUDY:

ChatGPT has integrated 45 Terabytes of compressed text data. By removing duplicates, its teams obtained a text mass of 570 GB (according to this article by BBC Science), corresponding to 300 billion words.

When we talk about super powerful servers, we mean computers capable of managing 175 billion different parameters for ChatGPT. This means that this AI is capable of understanding a very broad text and performing operations on it.

An example to illustrate this concept of "parameter": if we ask ChatGPT to write a summary on responsible digitalization, it retrieves from its knowledge base all the information on this subject. These data are parameters that it manipulates to produce a concise result.

According to OpenAI, version 4 of ChatGPT will be able to handle one trillion parameters. This should allow it to be faster and more accurate in its responses. As for the resources needed, we dare not think about it.

[Cover photo: DeepMind - recovered from Unsplash]