Duarte

Duarte

@duarteocarmo.com

Born in 🇵🇹, based in 🇩🇰. Machine Learning Engineer (or whatever it's called now) duarteocarmo.com duarteocarmo.com/photos github.com/duarteocarmo/

998 followers232 following438 posts

Trying to understand what are good open Jevs. WTF is going on with benchmarks this days??? LESS IS MORE GUYS!! LESS IS MOREEEEEEE

There is just something about the Cursor (now X AI or whatever) team and how they design interfaces. Cursor Agent Pane and Grok Bot are just incredibly designed. Just the right amount of knobs! In the age of slopped interfaces, its refreshing to see a team do LESS!

You find a bug in a famous benchmark. Do you fix it and run it in the fixed version? Do you leave it as it for comparable results? I hate benchmarks

Until today my connectors were a mess. Hermes uses this cli in my VM, pi uses the mcp for that there, codex has the integration here, gws cli for some email accounts there. A mess. Now I just use @executorsh. 100% recommend.

People of pi - what do you do when you want it to run every 2 hours or check something 30 mins from now?

I'm so happy I bought that refurbished Hetzner server with 64GB of RAM for 30 EUR/month. It runs pretty much everything at this point.

Useful loop tip #23 Have hermes go through a busy repo from last week, make a summary, and drop the markdown in your NextCloud server so you can go through it at your own pace

Used codex app for a day. Can solve Navier-Stokes but oh lord is this app buggy. Back to herdr 🐑.

The more you press "I don't like" Russ songs on Spotify - they always come back to haunt you in your Discover playlist.

Wonderful two days for Denmark. CPR? Hacked. DTU? Hacked as well? Is this a different attack? JESUS.

Had to use cursor today (sshing into hpc - and wanted to get an idea of files, directories, etc) One thing Cursor still nails is the review pane. Lazygit and hunk are good - but this is better.

I saw a post on LinkedIn saying we should use AMALIA to grade Portuguese high-school exams (the gov. was taking too long) Horrible idea.. But! I benchmarked a bunch of smaller models vs. AMALIA on a bunch of EU-PT benchmarks (including PT-exams) link below

The best thing about LLMs gotten this good is that I can focus on the writing and it will handle all the other bs for me Converting images, generating thumbnails, adding links, animations, visuals ALL THAT BS!!!! (yes it gets the link wrong sometimes, but I catch the mf)

If I can run 3 deepseek flash sessions in the same time astra is thinking about a single prompt - does this mean I cancel codex? jk

Building a slide deck today. AGI is here, but I still prefer powerpoint. -> Move logo here -> gray it out -> wtf is this font -> add this image I think it's faster if I just do it myself 😂

I have to compare two models on a benchmark. Which temp? Which max_tokens? Which top_p setting? Everywhere I look I see "we use the lab recommendation". Ok, so this is the wild west I see.

2024: Use parcel tracking apps 2026: "Check on <tracking_code> and let me know when something changes" Hermes I ❤️ you.

"Ping!" Agent ready. Running 2-3 agents in parallel doesn't need to be mentally taxing. Pi-jumper allows you to jump between @pidotdev sessions running on tmux with a couple clicks. It's part of pi-tools, some extensions I've built around Pi! Blog post below

Presenting Bagaço v3 the next generation of Bagaço, the largest pre-training dataset for European Portuguese! ~30 Billion tokens from across the Web, it now also includes Wikipedia and PDF data thanks to the great work from the folks at Hugging Face! Full write up in the comments

For the curious: - herdr is very nice, but I noticed it made me much more tired than using tmux - I needed something calmer/slower - above is pi extension + SwiftBar - Just tells me if I have a pi session done or waiting for input

.@spotify - explain to me why russ is always on my discover playlists? Even though I've clicked hide this multiple times?

This year I've built evals, classifiers, and even Bagaço, a pretraining dataset for European Portuguese. This raises an obvious question: Can we pretrain LLMs fully in European Portuguese? How do we measure if they're any good? Meet project Ginjinha 🇵🇹 (link in comments)