Becoming a self taught AI Researcher

2026-02-01T00:00:00+00:00

🔗My Background

I have worked as a backend software engineer for a good decade and a half. In 2025 I decided to pivot to AI. I want to stay between engineering and research. If you have similar goals, you might find the following tips useful.

🔗Indications and Contraindications

Firstly, and most importantly, I am no expert, so my opinions are subject to evolve as I learn more about the field. I will keep updating this blog post.

The recommendations are specifically targetted for people with decent software engineering and maths background who aim to understand state of the art research and be able to implement the papers.

I have listed only those courses for which you can find video lectures on YouTube.

🔗What to prioritize

The field of AI/ML is certainly quite broad, but I am keeping my interests aligned with what seems to be currently hot (circa 2026).

My subjective opinion is that there are 3 hot areas currently:

Large Language Models
Diffusion Models
Reinforcement Learning

I did do a course on ML, but you can ignore it if you want to focus on more in trend topics. I don’t think you would be missing out on much.

🔗Start with a good introductory course

For all the 3 areas I listed above, you would need to learn about Deep Learning. Even though Andrew Ng’s CS230 Deep Learning course is very famous, my recommendation would be Bhiksha Raj’s 11-785 Introduction to Deep Learning. The latter course has more breadth and depth.

🔗For Large Language Models

For a high level overview of the LLMs and all the associated topics, I can recommend Stanford CME 295 - Transformers & Large Language Models.

But if you really want to get your hands dirty, nothing beats Stanford CS336 | Language Modeling from Scratch. This would make you code a modern LLM architecture, implement a training pipeline, do GPU optimizations, scaling analysis etc. THE course to do if you want to master LLMs.

🔗For Diffusion Models

Diffusion Models require a bit more mathematical maturity. My maths was alright at university, but that was 15 years ago :(. So be patient if you don’t grasp the maths immediately.

There are many perspectives on Diffusion, and I would recommend to look at them from all the perspectives. Some would feel a bit more intuitive than others. I have the following recommendations:

🔗For Reinforcement Learning

My first recommendation would be the DeepMind’s courses: DeepMind x UCL | Deep Learning Lecture Series 2021.

Stanford has two courses, but I haven’t tried them myself, they are on my to watch list:

🔗For Understanding GPUs

Watch lectures 2, 3, 7 and 10 from this course Stanford CS149 I Parallel Computing. This should give you a good mental model of modern GPU architecture. This should help you think about how neural architectures run on a GPU and how you can optimise your code.

🔗Code the models by hand

Nothing teaches you better than implementing the models by hand. We really do live in a golden era of self learning, YouTube is a gold mine for that. Here are some of my favourite YouTubers:

These guys have multi-hour long videos going through excruciating details of how to implement various models.

Rohit's Blog