Duarte O.Carmo

A few weeks ago there was a bit of a scandal back home.

We've gone from suffering from the heatwave in the south of Europe to complaining about rain in Copenhagen.

Today, I'd like to talk about Pi.

Source composition and averages.

July was busy — like always — but we made sure to take some time and touch some grass.

For the past year, I've been very curious about the intersection of Large Language Models (LLMs) and the European Portuguese language.

This month's Retrospectiva comes a little later than usual.

In January this year, OpenClaw went viral.

If there's one thing Retrospectiva does for me - other than giving me an excuse to write for you of course - it's making me realise how fast time moves.

About 6 months ago I wrote a small rant on how open source TTS models still sucked. 6 months later, I'm happy to report that isn't the case anymore.

Educational score over time vs. document count for Bagaco v2 A couple of months ago I released Bagaço - a pretraining dataset for European Portuguese.

My Hermes agent – Saramago – giving me coaching feedback and looking for the ideal flights for an upcoming trip.

Update July 2026: The AMÁLIA models have now been released!

Well, now would you look at that.

As you get older, life becomes complicated.

Given some text in Portuguese, how easy is it to determine if it's from Brazil or Portugal1?

Happy February!

Let's say your goal is to train a Large Language Model only on European Portuguese.

Well—that was quick.

Let's face it.

I get up feeling sleepy.

And just like that, it's the end of the year again.

Benchmarking calorie prediction for Taralli There's something very funny about the current Machine Learning and AI landscape.

Big news in the state of Denmark.

Back in May this year, my longtime friend Alex reached out and asked me if I wanted to collaborate with him on a book about DeepSeek.

or at least the open versions of it.

Welcome to Retrospectiva.

Why do this?

The world moves fast, faster every day.

It's that time of year again.

A couple of weeks ago in Lisbon, I went to a friend's birthday dinner.

..but they can also be a lot of fun.

Last month, I built an app called Taralli.

For years, I never really cared about what I ate, how much I ate, or when I ate it.

Even though it never really replaced physical books for me, I've been a big Kindle user for many years.

Last year I made an effort to speak less and learn more.

Last week over lunch, Pedro was telling me how he uses Perplexity to improve the way he searches around the web.

In the period of the generative AI boom, everything is happening all at once.

Notebook download - nbsanity preview Some context We had a database of 50M strings and I couldn't wait to embed them all.

I'm flying over the Atlantic just between France and Spain for my yearly Christmas trip back to Lisbon.

For the past year, this website's podcast companion has been running on a text-to-speech model called XTTS-v2.

For the past years Vitto and I have used Tricount to track our shared expenses.

For a short summer break, Vitto challenged me to visit Iceland with her.

Tanda Runner is a web app that shows me the things I care most about when preparing my next marathon.

I'm lucky enough to work with some pretty talented folks.

I'm a sucker for side projects.

There I was.