Pablo Fernandez 🧉

Reading the book AI Engineering, how is RAG supposed to be context construction before calling the model API? #ai

2025 was the year of the year in review

"You do not rise to the level of your goals, you fall to the level of your systems" -- James Clear

"You should be far more concerned with your current trajectory than with your current results." -- James Clear

I've been coding almost every single minute of this weekend so far and dreading it finishing. I just want to continue. I'm making progress.
3/3

But even there, I can learn ML concepts quickly using an LLM. My understanding is superficial, so I'll learn them properly later on, but it's enough to trigger experiments to see if I can improve the metrics of the system I'm working on.

This is fascinating and exciting.
2/3

I'm starting to see how the amount of code merging that used to come from a team, can now be produced by a single person with AI. And I'm not even maximizing for throughput since I'm doing ML work and the limiting factor is my ML knowledge.
1/3

I saw this at a hotel.

A year ago this would have made me think of secondary school.

Today it's making me think of AI.

I just double cloned my machine learning project repo to have two independent AI agents working in parallel. So it begins...

How important do you think is having fun as the culture of a company?

I woke up at 4 and couldn't go back to sleep. I was too eager to go back to coding on my machine learning project. It's noon now and I don't know where the time went, it just flew by. So much fun and so productive. But sleepiness is catching up with me.

Finetuning or RAG?

Finetuning to fix behavior, RAG to fix information.

Also RAG is a lot easier (at least at the beginning).

#AI

If you wait long enough, a large enough amount of hydrogen, starts thinking and multiplying vectors.

And if you continue waiting... it looks like the vectors might start thinking too!!!!

</joke>

That means if you are a CEO, your ambition needs to be extremely high, or people will leave.

It's not possible (or very hard) to achieve more than your boss at any company, so the ambition of the boss sets the ceiling for the career progression (and growth) of everyone reporting to them.

Also, I learned from my mistake. This time it's not running in the Cursor terminal, but in tmux inside a proper terminal.

And so it begins (again). Another round of training nanochat. This time with my MLA modifications to compare against my previous non-modified version.

#deeplearning

There's a lot of mistakes to make in business once you are not making all the mistakes hundreds or thousands of people documented already in books, podcasts, curses.

I just saw in a survey that one of the things people want to trust AI more is for employers to them people in how to use it. The reality is nobody really knows how to use it this is to new there are no expert we all beginners.

I see a lot of "learning from failure" in software engineering. I think it teaches you techniques, how an API works, but it's not good at teaching you fundamentals, like the concept of time complexity.

Someone recently told me that learning by trial and error is how they got better at business and I thought "Oh geez, that is a slow expensive way to get those lessons".

"You can go as fast as the trust you have built with the stakeholders around you."
-- Anne Morriss

My social media bubble: I will hang up on AI interviewers.

A study run on recruitment: 78% of candidates choose the AI agent.

Those interviewed by AI are half as likely to report being gender discriminated, women were most likely than men to choose AI.

#ai

My first local training of an LLM, nanochat, finished. Well... it crashed around the testing part, so I restarted it from the last step to make sure it finished correctly and passed the test.

Now I'm running the fine tunning.

#deeplearning

I just understood what agentic AI is... It turns out I've been doing agentic AI for a year and a half. Go figure.

#agenticai #ai #llm

I recently started using Coder at work, to get GPUs for machine learning. I wanted to know how much my use was costing and also, whether I could use it in my spare time for personal learning. Well this pricing page kinda lacks... you know... the price:

Yesterday the LLM fine tunning I was running on my machine threw an error. I saw it, went to bed "I'll tackle it tomorrow".

Today, Windows restarted to apply upgrades. The error output is gone.

I'm so close to switching back to Linux (used it as my desktop for about a decade).

"I replaced my developers with Ai... I'm still waiting for it to clock in... Any day now"

#llm #ai

A better monetization model for LLM applications Explore the challenges of LLM-powered apps with subscription models and discover a win-win solution for users and providers.

Did I hear OpenAI will let you share your tokens with other apps? Sounds like the idea I wrote about a few months ago:

pablofernandez.tech/...

#llm

I'm starting to use voice to talk to Cursor and I feel like freaking Iron Man. Will engineers finally get private offices so we can talk to our computers in peace?

#llm #ai