Evolution of AI: Evolutionary levels of AI

Just as people go through a process in which their cognitive and emotional development evolves, artificial intelligence (AI) is also undergoing several phases of evolution over time. Below is a list of the five main phases of AI evolution:

Symbolic AI (Symbolic AI).

This first phase focused on the representation of knowledge using symbols and logical rules. It was popular during the 1950s and 1960s. In this stage, AI systems relied on humanly programmed rules to solve specific problems.

Machine Learning

From the 1980s onwards, AI entered a machine learning phase. At this stage, AI systems could “learn” data and patterns through the use of algorithms and mathematical models. This enabled the application of AI in a wide range of tasks, such as image classification, natural language processing and recommender systems.

Neural Networks and Deep Learning

With the advent of artificial neural networks and deep learning, this phase took off in the 2010s. This technology enabled the creation of much more complex AI models capable of handling large volumes of data. Deep neural networks were critical to the success of AI in tasks such as image recognition and natural language processing.

Within Deep Learning, we can distinguish three stages:

DL 1.0 → Supervised Learning (2012-2018)

This stage was characterized by being very effective in specific tasks, such as identifying whether a leaf is healthy or not, distinguishing between a dog or a cat, or recognizing gender. However, it was not efficient for less specific tasks, and the primary bottleneck was the availability of labeled datasets.

DL 2.0 → Self-Supervised Learning (2018 – present)

In the beginning of deep learning 2.0, models could learn from datasets without the need for labels. Some examples include Chat GPT, models for generating images, or sound. Furthermore, to enhance model performance, the capability for models to generate content themselves was introduced, allowing users to determine what is good and what is not (feedback provided to Chat GPT).

In this phase, the exploration doesn’t stop there; multimodality is emerging. Multimodal models can learn from data of various types (audio, video, text), resulting in much more intriguing outcomes. The next version of Chat GPT is likely to be multimodal.

DL 3.0 → Reinforcement Learning (RL)

Models in the deep learning 3.0 phase are characterized by being trained without the need for a database through “reinforcement learning” (RL). In this scheme, AI learns from scratch. Some of the most famous examples include AlphaZero, where AI was capable of defeating the world’s best Go players. The significant advantage of this technology lies in its “trial and error” approach, whether in the real world or in simulators, with the potential to acquire knowledge beyond what humans have today.

General AI (AGI)

This phase is still in evolution and refers to the creation of an AI system that can match or surpass human intelligence in a wide range of cognitive tasks. It involves abilities such as reasoning, machine learning, natural language understanding, and problem-solving across different domains. AGI is capable of adapting and comprehending new contexts without specific programming. Despite being an aspirational goal, it is still under development.

Superintelligent AI (Singularity)

This is a hypothetical phase in which AI would significantly surpass human intelligence and could make autonomous decisions without human supervision. This phase raises crucial ethical and security concerns, since it could have profound implications for humanity.

It’s essential to note that these phases simplify the complex history of artificial intelligence, and the field is rapidly evolving. Many experts continue to work toward achieving a more advanced level of AI capable of solving more complex problems.