How AI works in 3 sentences

OpenAI's GPT 4 is a highly complex AI model that contains about 100 trillion parameters (at the time of writing). Nonetheless, how AI models are created and how they work can be captured in merely 3 sentences!

1. An AI model is as a structure of computable logic and/or math.

2. Therefore, it contains values (or parameters) for computation.

3. For a modern AI, a learning algorithm is responsible for sets these values appropriately, with an ultimate goal of making the AI great at performing a specific task.

Let me illustrate in detail with two toy examples.

The 3 sentences that summarize how AI works. (Thank flaticon.com for illustrative icons.)

---

Example A: Predict water mass via its volume

Assume that we have the following experimental data

​Volume [liter]	Mass [kg]
1	1
2	2
3	3
4	4

We wish to create an AI model that can predict mass of a body of water from its volume.

We may propose an AI model whose structure is mathematical, say

m = x v

where v is the volume and m is the mass. The proposed model has one value as represented by x.

This value x needs to be set. In fact, this kind of problem is called regression in the AI world. A learning algorithm suitable for regression problems will likely set x to 1 by learning from the given data as shown in the above table. The resultant AI in this case could be

m = 1 v = v

---

Example B: Predict if a loan applicant is likely to default

The task is to determine risk of granting a loan to applicants. Assume we have the following data collected from the previous lending program:

Customer ID	Salary เงินเดือน	Dependent Child จำนวนบุตร	Default ผิดชำระหนี้
C001	50,000	0	No
C002	60,000	2	Yes
C003	49,000	0	Yes
C004	30,000	1	Yes

Rather than a mathematical structure, I could propose a logic one. A decision tree is a logic structure. At each node, we ask whether a certain criterion is satisfied. If yes, we go on one path to another node, or else we take the other path to a different node. The decision process stops when we can make a conclusion, having accomplished the task.

Thus, in this type of structure, the values to be set are the criteria of all decision steps. This can be accomplished by several learning algorithms for decision trees. As example of AI for this task is shown below. The key concept is that the learning algorithm results in a set of criteria that is best suited to the input dataset.

Logic-based AI mode called a decision tree.

---

Bonus--What does this mean to you?

If you are an AI researcher, you might start tinkering with a new AI structure or devising a new learning algorithm. If you do this, you are likely to test your invention with a benchmark dataset.

If you are looking to apply AI to your career, your first focus is preparing a good data set to train an AI as you will likely employ existing AI structures and their respective learning algorithm. The data is key as garbage data will lead to garbage AI.

---

In summary, 3 sentences that wrap up how AI work and how it is created are:

1. An AI model is as a structure of computable logic and/or math.

2. Therefore, it contains values for computation.

3. For a modern AI, a learning algorithm sets these values with an ultimate goal of making the AI great at performing a specific task.