We need to adapt.

https://en.wikipedia.org/wiki/Overfitting#/media/File:Overfitting.svg

I am following a 12-month learning program, aiming for around 1,700 hours in total. This article is a snapshot from the 400-hour milestone.

I typed this with my own hands because I want it to keep it raw. In this article, I’ll share the resources I picked, why I picked them, and how you might cherry-pick some of them for your own learning path.

When GitHub Copilot became generally available in mid-2022, I was one of the early adopters. It was great, and it was making me 5x to 10x more productive when it comes to code generation. Especially superb on test generation. It was surprising for me that not many people were aware of it, because in my mind, it is something that you have to use if you were coding by hand in 2022. Why not be more productive?

Then ChatGPT launched in late 2022, powered by GPT-3.5. It was viral, and it was impressive. But from my point of view, it was not shocking in the same way, because I had already experienced Copilot inside my IDE. Copilot felt closer to the codebase and more directly integrated into my workflow.

For the next few years, many more great tools emerged, Cursor being one of them by being able to position itself nearer to the developer code base. But for me it wasn’t clicking right, maybe because capability of the models weren’t enough to achieve what it was promising.

Then came the late-2025 release of Claude/Codex-era models, proving that there’s still more room to grow. As a consumer of these products, it was this moment for me to accept its power by experience.

And from then until 2026 autumn, harnesses positioned themselves finely, and I think for a good long time, very close to the OS. Not even the codebase, but closer to the machine. Then I became 100x to 1000x more productive.

The applications that I’m starting from scratch are now maturing in a full month, compared with before — it was taking years.

I, personally, cannot live without knowing the internals of such a technology. And curiosity as being my #1 driving factor, I am in a full time, self-taught, mostly open source, controlled, and dedicated learning program. Here are the main branches that I’m learning at the moment:

  1. Deep Learning
  2. Generative AI
  3. Agentic Coding

For the rest of this article, I want to share the resources I’m learning from.

Deep Learning

Main resources:

As mentioned in many different places online, fast.ai courses provide top-down approach when teaching. This is great for first time learners like me when they are entering a new field. This course helped me to ease-in to the field. Especially first 5 chapters, along with videos was very well thought and conducted.

There’s one thing I found particularly challenging in the fast.ai library coding standards and formats. Most probably because of my background, as many of the application developers, I am obsessed with code quality, variable naming, clean code, and modularity, as well as many other clean code principles. This is what working life taught us as application developers. Every piece of code we wrote has been scrutinized by our peers for the good, and we did the same. Fast.ai coding style did not match the kind of modular, explicit application code I was used to. Mostly optimized for personal, and subjective conventions. Which is good if you want to lock-in for a fixed point of view, but not going to be sustainable, and scalable for fast changing environments with fast changing people.

Does it matter? Because I’m no longer writing code in the same way.
It matters because I’m learning the principles and the library APIs bridging that gap. And with this uncomfortable feeling when I was going on with this course, I stopped reading coding examples — which is a significant part of the book and the course, and focused on the semantics instead. And started to look for a better library. If you want to start learning with fast.ai, you should know that there are a lot of dead-ends in example notebooks. I still recommend the course for its profound semantics.

Coded with precision, and with clean principles in mind, I found what I needed in Lightning AI library, and actually much more than this. I get to know Sebastian Raschka, and his amazingly neat deep learning material. The course is conducted by YouTube videos, and exercise notebooks in Lightning AI website. Sebastian’s well-planned course and Lightning AI’s well-organized code is my choice of today, when I’m developing my first wave regression model.

With Lightning AI course, I made the second pass of learning on Deep Learning topics, as a deliberate act mainly to retain knowledge by repetition, and also get to know different angles to the same topics. What wasn’t clicking for me on the first pass, started to hit like a hammer with Lightning’s course.

Generative AI

Nearing the end of deep learning courses, I still want to keep learning. That’s why I researched a good follow-up course and found:

Relies on HF library, which is a higher level compared to both fast.ai and lightning.ai libraries, and focuses on how large language models work. Updated with the transformer architecture, and till 2025, which is enough for me to catch up.

HF learning material became the second pass for me for the transformer architecture, which was also covered well in Lightning AI course.

HF enabled me to use a vast range of models in my local machine, as we do with Ollama, but also play with their internals with the APIs that it provides. It is also beneficial to get to learn how to navigate around model lists in HF. And this prompted me to deploy my own model to HF at ease.

Agentic Coding

When there’s a new tool for something that I was doing for years, I want to experience it. I want to have my own way on that tool, but also I want to peek into others, to see what is their interpretation of the new way of doing things? These are the resources that are giving us that peek-ability:

It is a learning club/platform/community led by Alexey Grigorev. They have many different courses and recurrent cohorts on relevant topics of 2026: AI DevTools, Deep Learning, Machine Learning, LLMs, etc…

One thing I like about the club is that it can be interactive. If you can catch the cohort time, you can join live classes, and interact with Alexey himself and with other students through Slack.

Alexey is kind enough to provide most of the material free, and has an organized teaching style. Supported with homework, projects, and peer reviews, it keeps your hands-on experience alive and involved.

This is a nice entry point for coding harnesses, if you want to explore, for example how Codex, or Claude works on client level, I strongly recommend that you give a chance to the AI DevTools course.

I watched Matt’s conference talk on recent tools, and what today’s workflow looks like, and found him explaining concepts intuitively, and catching the soul of the time. He is also sharing coding examples and agent skills with community. The video that he is explaining AI workflow could be a good entry point.

While I’m learning around these 3 main topics, for me it is also important to surround myself with people and other resources that are still relevant in 2026, have proven background, and can see the future.

A great architect, but not only that, he is a great teacher. His iconic Monday videos are happy pills that we can take every now and then. I recommend his book Fundamentals of Software Architecture (with Neal Ford) to have an in-depth-and-breadth knowledge about software architecture, and his YouTube channel; for the happy pills.

A great YouTube channel, if my memory serves well, recommended by Mark Richards, and that’s how I inherited it. They are conducting frequent live sessions on YT and LI, covering various topics; architecture, different domains, live coding sessions, experts from different tech areas. I think it is a diamond channel.

David Heinemeier Hansson is a super famous personality especially in the Ruby world, creator of Rails, and a beneficial source to guess the future. Following his writing and talks is enough to get a strong sense of his current thinking.

I recommend his book “The Singularity Is Near” or the 2nd one “.. Nearer”. He knew what was going on back in 2010s, and guessed the exponential growth we are experiencing today. A good source to find a general meaning of what is happening and what might happen soon.

A great engineer shares logs and news from his point of view. I recommend having a look at his blog posts, and following him on X.

He is not only good at teaching DL, but also a great author. I recommend his recent posts around relevant tech topics.

I can recommend more, maybe on a 2nd part of this article.

There are many reasons why you might still want to learn coding, and if you want to learn Java, I strongly recommend this YT channel. Specifically because it is up-to-date, and I like that the teacher is going into depths of programming languages, and doesn’t hesitate to explain language-machine relationships. One big problem for me was the language barrier, because the content was not in English, and with YouTube auto dub, I was able to overcome it up to a point.

Matt Pocock, in his conference:

You should head back to Amazon and buy a ton of those old books.

This is more or less the resources I’m using for the last 2 months, I hope it can provide value for you too, as it did to me.

Since I’m learning topics around ML, DL, Agentic Coding, it is important for me to meet like-minded people around these topics. If you are interested, feel free to reach out to share your experience.