How I stay up to date with AI

Introduction

If you’ve been working actively in AI like me, either in industry or research, you’ve certainly noticed how hard it’s become to keep pace with discoveries and advancements, especially in the last two years. From DeepSeek-R1 ¹ to the Hierarchical Reasoning Model ², from AGI promises to language limits ³, and from simple chatbots to complex agentic workflows, being up-to-date has become overwhelming.

At least this has been my idea so far, until I realized one simple thing: it is just impossible 🤒.

That’s why I’m writing my own recipe here, despite the thousands out there. Not to be another exhaustive to-do list no one will completely fulfill, but:

• 🧑🏻‍💻 for me to remind myself what I’ve read this year and recap the most important concepts;
• 👉🏻 for you in case you missed something;
• 💬 for us to exchange other resources (and that’s sure to be the case)! Leave it in the comments!

Let’s dive in!

First principles

Before just giving out a fancy list of links, I wanted to share my three “simple” rules that I generally follow to keep up-to-date 🎯:

• I subscribed to AI News: you don’t really need anything else. Sections get right to the point, and the “not much happened today” email saves me useless digging. There are tons of other newsletters I could mention, but I started unsubscribing progressively once I noticed I was piling up unread emails rather than learning something new ⁴.

• I follow DevRels or tech profiles from Mistral, HuggingFace, vLLM, Docling, etc. (on LinkedIn and X) to find meetups and hackathons around me. I’m living in Paris and Luma is amazing for that! This is, for example, how I found the hackathon we won last May ⁵. Getting to know people in the field and participating in live events remains one of the best sources of information, and it’s fun 🎉!

• I kept the PhD habit of skimming titles from the best AI conferences currently happening. If you’re not aware of which labs or conferences to look at for your topic of interest, the trending papers section on the HuggingFace website is a great filtered source. It has often brought interesting papers to my attention that I would have never found otherwise.

That’s it, really. Too much more information would result in no information at all. But a list-like thing was promised in the premise, so let’s give it a look!

My 2025 AI tech-list

Below is a curated, non-exhaustive list of resources I’ve explored and found valuable. These span the most important topics of the past two years, from foundational concepts to advanced insights and practical skills.

Books 📚

Theoretical foundations and practical implementation: these two books create the perfect combo for any ML Engineer.

Sebastian Raschka - Build a Large Language Model (from scratch): Covers all theoretical and technical topics from data to architectures and training. Good for everyone, but easier if you’re already using Python daily and you’ve already trained a model at least once. Sebastian’s blog is also amazing: magazine.sebastianraschka.com.

HuggingFace - The Ultra-Scale Playbook: Training LLMs on GPU clusters. Born from a HuggingFace blog post, this addresses the real world complexity, going deep on hardware and infrastructure details for engineers scaling their solutions.

GitHub repos 💻

Hands-on learning through code: the best way to understand theory while building.

nanochat: There’s no better teacher than Karpathy. Nanochat is the perfect combination of theory and coding, and it’s most likely the best free instructional content out there. Well explained, deep, and unlike toy examples, this lets you build your own ChatGPT-like end-to-end with frontend and infrastructure! Perfect exercise after the previous two books 💻. Also check his video on YouTube, pure gold material.

HuggingFace smolagents: This has been the year of Agents. Whether you’re using them at work or not, you need to know why an agent differs from an LLM (Thought, Action, Observation), how to write tools, and what the novel MCP protocol is. The library lets you build simple yet powerful agents, integrate with the most used third-party frameworks (also for RAG), and comes with a course explaining the theory (docs).

Blog posts 📝

Deep dives into current challenges and cutting-edge research: from interpretability to data.

Anthropic Blogs: LLM interpretability is one of the most promising yet difficult challenges we’re facing. These powerful yet mysterious architectures are hard to interpret. Tracing why certain inputs generate certain outputs is crucial for medical or environmental applications. Anthropic “opened” Claude in a beautiful blog post: Tracing Thoughts in Language Models, followed by a deeper article stressing transformer architectures on tasks beyond language.

Thinking Machines (thinkingmachines.ai): Mira Murati’s recent project ⁶ started with super interesting blog posts solving complex LLM problems. The first one, Defeating Nondeterminism in LLM Inference, tackles LLM non-determinism, a problem engineers have tried to solve with temperature 0 and a lot of praying 🙏.

FineWeb dataset (HuggingFace blog post): We always focus on architectures and optimization, but rarely give importance to data, especially open access for reproducibility. The HF blog post transparently presents the complexity of collecting, filtering, and using huge datasets, and how data mix impacts model behavior. See how different data mixes trained SmolLM3 in stages (blog post) to enhance grammar, math, coding, reasoning, and coherency.

Bonus: Two recent HF guidebooks (which I still need to read) seem promising: Evaluation Guidebook and Smol Training Playbook.

YouTube videos 🎥

Visual explanations and deep discussions on AI’s future and current limitations.

Yannick Kilcher: Yannick is a smart researcher and engineer who, on top of technical paper reviews and clear explanations of complex topics, has strong and valid thoughts on our world that more than once made me think critically. Check out his video on why AGI will never happen and the “An Image is Worth 16x16 Pixels” paper explanation where he brilliantly explains the core difference between Transformers and older CNNs/RNNs.

Pre-Training GPT-4.5 - OpenAI video: Super interesting about problems solved with PyTorch, the role of data, and limits of generative-based technology.

Conclusions

So, here we are. After all this, what’s my take? Keeping up with AI is indeed impossible, and that’s okay!

I’ve learned to be selective: focus on quality over quantity, follow a few trusted sources, and most importantly, build things 🛠️. The resources I’ve shared here have helped me not just stay informed, but actually understand and apply what I’m learning.

If you have game-changing resources, share them in the comments! Let’s keep this conversation going 🤗

References