Hadelin de Ponteves and Kirill Eremenko made their names with Machine Learning A-Z and Deep Learning A-Z, both of which we review here. Artificial Intelligence A-Z was the third in the set and for years it was, in plain terms, a reinforcement learning course. The current version is titled AI A-Z 2026: Agentic AI, Gen AI, Prompt Engineering and RL, and it is taught by de Ponteves, Eremenko, Luka Anicin and the SuperDataScience team with Ligency. The page showed 361,670 learners, a 4.4 average from 51,129 ratings, 17 hours 22 minutes of video and an update in June 2026. The course now has ten parts.
The first three are prompt engineering, generative AI and agentic AI. Parts four to eight are the original core: Q-learning, deep Q-learning, deep convolutional Q-learning, A3C, then PPO and SAC. Part nine covers LLMs and fine-tuning a Llama model on medical terms, and part ten is responsible AI. Across these you build twelve small projects. I will start with what works.
The RL sections follow the A-Z formula of an intuition lecture followed by an implementation. Eremenko's intuition lectures are very good at making an idea like the Bellman equation or experience replay feel obvious, with diagrams and no heavy notation. Then you build an agent that lands a lunar module, one that plays Pac-Man from pixels, one that fights in Kung Fu, and a self-driving car trained with PPO and SAC. These run in Colab and they are satisfying to watch train.
For someone who has heard of reinforcement learning and wants to see it working inside a weekend, I do not know a friendlier route. The cost of that friendliness is rigour. The course says openly that it will not throw complex mathematics at you. That keeps beginners moving, but RL is a field where the maths is the understanding. After this course you will be able to run PPO.
You will not be able to say why it is more stable than a vanilla policy gradient, or debug it when the reward curve goes flat. If you want that, David Silver's lectures or the Hugging Face Deep RL course are the next step, and both are free. The newer front half is weaker. Prompt engineering and generative AI are taught largely through AWS: a chatbot in PartyRock, a custom model and a RAG app with Bedrock and Knowledge Bases, an agent for business assistance in the cloud. It reads like material from the instructors' AWS certification courses moved across, and the responsible AI part at the end, with SageMaker Clarify and Model Cards, confirms the impression.
None of it is wrong. It is just tied to one vendor's console, and consoles change. I think the title is the real issue. Agentic AI comes first in the name and is the thinnest part of the syllabus. Someone buying this in 2026 to learn to build agents will get one section and then find themselves training a Pac-Man player.
Someone who wanted RL has to get through three parts of prompt templates first. The course would be more honest, and I suspect better reviewed, as two separate listings. There is also the usual A-Z caveat. Implementation lectures are typed out line by line while you follow. It feels productive, and for many people it is not.
Close the video and try to rebuild the Q-learning agent from an empty notebook. If you cannot, go back. At Udemy sale prices the downside is small. I rate it lower than its siblings because it is less focused than either. Buy it for the RL, treat the rest as a bonus, and do not expect it to make you an agent builder.