Pablo

Machine Learning Research Intern at CSIC

[Lorenzo Pacchiardi @LPacchiardi]

🚨 New paper: AI evaluation is structurally unsuitable for continual learning (CL). To address this, evaluation should be centred on the "behavioural trajectories" that CL systems develop, with the goals of characterising possible behaviours and forecasting their evolution. 🧵

[Pablo @ThePablo22]

@josephwdr @elonmusk @AnthropicAI Right = Do what the user wants under the chatbot provider's restrictions.
Wrong = Not doing that. E.g., cheating to pass the tests but not actually solving the problem.

That's alignment.

[Avi Chawla @_avichawla]

ML researchers just built a new ensemble technique.

It even outperforms XGBoost, CatBoost, and LightGBM.

Here's a complete breakdown (explained visually):

[Becas Fundación "la Caixa" @BecariosFLC]

📢 Exciting news!

We’re thrilled to introduce the 100 new #laCaixaFoundFellows selected in the 2025 postgraduate fellowships abroad call 2025. 👏

Meet them! 👇
https://t.co/32YwbJ5wIM

[1/3] 🧵

[Pablo @ThePablo22]

How to End a Talk?
--------------------
- The last slide must list your contributions
- Don't end a presentation with a "Questions?" or "Thanks" slide. It wastes this opportunity
- Ending with "thanks for listening" implies they were listening out of politeness

[Pablo @ThePablo22]

I just finished watching Patrick Winston's 1-hour talk at MIT titled "How to Speak"

These are my notes👇

[Pablo @ThePablo22]

Vision: A problem people care about, and it must be something new in your approach. How do you express that you've done something? List the steps that need to be taken to achieve the solution, and what you have already accomplished

[Daniel Han @danielhanchen]

Yep I'm back for AI Engineer World's Fair!!

I'll talk about RL, GRPO, dynamic 1.58bit quants for DeepSeek R1 and cool tips and tricks + more! It'll be 2 hours (or 3 if demand is more :)) split into multiple topics like last time! Ask all the questions you want!

[Tom Dörr @tom_doerr]

Convert a Docker image into a standalone executable you can run or share

[Max Zhdanov @maxxxzdn]

people say GNNs are turning obsolete, but one thing I noticed in my experiments:

attention is not so good at picking up local fine-grained details and you need GNNs to do the job!

This is because GNNs are a lot more expressive when it comes to pairwise interactions.

1/n

[Radek Osmulski @radekosmulski]

A big misconception I unconsciously held about the @fastdotai way is that it only applied to “sciency” fields.

But that is not the case.

@fastdotai courses worked so well for me precisely because they took a humanising approach, because they incorporated at a deep level an understanding of how we learn and how the world operates.

You can find echoes of the successful ideas in other works:

• the ultimate personal brand builder and learning accelerator in @austinkleon’s “Share your work!”
• project based learning in “Why and How of Woodworking: A Simple Approach to Making Meaningful Work”
• blueprint for building effective, complex systems by starting with a baseline as reflected in Gall’s Law and more broadly Systems Theory
• fast improvement by increasing the speed of meaningful iteration and often starting with a batch of one as espoused by the Toyota Way and the OODA loop
• error analysis as a basis for informing subsequent designs which is central to Structural Engineering
• learning theory and creating a social environment conducive to learning as outlined in “Making Learning Whole” by David Perkins
• and many, many more highly effective practices…

The novelty of @fastdotai is applying all these ideas to the cutting edge field of machine learning, but this enormous feat, packing all this goodness into courses and the community, could only be accomplished by a mind that is extremely well versed in science, but also in humanities.

Ergo, if you only spend your time studying technical fields, you impoverish yourself from the overwhelming majority of human knowledge.

You limit how successful you can be.

[Jack Morris @jxmnop]

Deep Learning is Not So Mysterious or Different (2025)

[Carlos Alberto Haro @h1sort]

this is one of the best papers I’ve read about MLOps