pedromadruga.com - Profile | ThreadSky | a Reddit-style client for Bluesky

I've been trying Cogito 32B locally and it has by far the highest inference speed for all the 32B models I've tried. It has a nice context of 128k and supports 30 different languages. However, I stopped paying too much attention to benchmark results when new models come out.

submitted 14 hours ago • 1 comment

Tilføjer til din læseliste 👇 news.ycombinator.com/item?id=4369... #dkai #dkdev

submitted 3 days ago • 0 comments

Google hiring for post-agi research 🤔 uk.linkedin.com/jobs/view/re...

submitted 13 days ago • 0 comments

It was always a matter of time 😎 thank you!

submitted 20 days ago • 0 comments

“Clever engineers write clever code. Exceptional engineers write simple code.” And so much more to unpack here. Worth reading.

submitted 20 days ago • 0 comments

Context size is also wild. Welcome LLama4 👋 ai.meta.com/blog/llama-4...

submitted 23 days ago • 0 comments

I have sort of given up on proactively updating the Danish Machine Learning Peoples starter pack. If anyone thinks that they should be added or know someone who should, please don't hesitate to comment here or sent a message 🤗. bsky.app/starter-pack...

submitted 23 days ago • 1 comment

LLM compression, by Apple 🔥 “Our experiments with Llama3 70B, […] show zero-shot accuracy retention at 4- and 3-bit compression to be on par with or better than state-of-the-art methods, while maintaining performance comparable to FP16 baselines.” machinelearning.apple.com/research/see...

submitted 24 days ago • 1 comment

Opensource all the way! github.com/sentient-agi...

submitted 24 days ago • 0 comments

I have been trying Mistral lately - both via Vertex API as well as via LeChat. I find it particularly good at translating (to and from) Portuguese from Portugal, especially when compared with Gemini for example (where it tends to write in Brazilian Portuguese instead).

submitted 24 days ago • 0 comments

The inability to pivot is very detrimental to the success of a team and a product, in the new age of GenAI-based product development.

submitted 25 days ago • 0 comments

Local LLMs achieving a 99.4% accuracy when performing sensitive data anonymization. ai.nejm.org/doi/full/10....

submitted 27 days ago • 1 comment

This would be great! Awesome initiative @serge.belongie.com et al.

submitted 29 days ago • 0 comments

Looking at the 2024 index, one can see how things have change. Just from one year to another.

submitted 32 days ago • 0 comments

IRRJ, the Information Retrieval Research Journal, is the new player in the information retrieval publication landscape. The first issue has recently been published. Have a look at irrj.org/index, and don't forget to follow @IRRJ.sigmoid.social.ap.brid.gy!

submitted 32 days ago • 0 comments

In my latest column for Science magazine, I discuss recent AI "reasoning" models -- how it works, to what extent it captures "genuine" reasoning processes, and what's needed to answer such questions. www.science.org/doi/10.1126/...

submitted 39 days ago • 8 comments

There are books that are such a pleasure to read, where I find myself not reading them in order to try to perpetuate the joy. (It’s a paradox indeed) This book was one of them: app.thestorygraph.com/books/d3bbb9...

submitted 39 days ago • 0 comments

In an agentic world, where models tend to be more specific, I’d say the playfield is leveled by now

submitted 41 days ago • 0 comments

One of my favorite newsletters out there

submitted 43 days ago • 0 comments

Publications from 2025 are shared more on Bluesky than on X/Twitter https://bsky.app/profile/altmetric.com/post/3lkbh7lvglc2i

submitted 43 days ago • 0 comments

An assembly of 18 European companies, labs, and universities have banded together to launch 🇪🇺 EuroBERT! It's a state-of-the-art multilingual encoder for 15 European languages, designed to be finetuned for retrieval, classification, etc. Details in 🧵

submitted 49 days ago • 5 comments

If bluesky made it, so will lemmy.

submitted 49 days ago • 0 comments

A bookmark for all the ones who run local models

submitted 54 days ago • 0 comments

Migrating from icloud drive & gmail to proton. Despite reddit, the integration of proton drive in linux (PopOS) was a breeze (via rclone). There’s also a native .deb for it. And also works.

submitted 56 days ago • 0 comments

Their open source week went a bit under the radar but it was a fantastic initiative.

submitted 59 days ago • 0 comments

'What's the biggest problem we're solving?' I love asking this. This question will give you (and everyone in the room) insights on wether people understand the bigger picture, on what's the biggest problem for them, to understand everyone's motivation, perspectives and ultimately reach consensus.

submitted 61 days ago • 0 comments

Patiently waiting for benchmarks but one can only imagine how many local models could be run

submitted 63 days ago • 0 comments

There's low air quality going on right now (in Denmark and Sweden, at least). This is for PM2.5 particles. But it also seems it's going to stay like this for a couple of days. Be safe out there. Sources: - storage.googleapis.com/gmp-maps-dem... - IQAir app - Google Maps app (EAQI icon)

submitted 64 days ago • 0 comments

There's probably an EU alternative to what you're looking for, no? european-alternatives.eu

submitted 64 days ago • 0 comments

Apple MCP tools A collection of apple-native tools for the MCP protocol. One simple command to give LLMs access to a bunch of apple-native tools like: - contacts - notes - iMessages and more (soon) github.com/Dhravya/appl...

submitted 66 days ago • 0 comments

As my 3rd book of 2025, I've read Show Your Work - and I'm recommending it. It's (rightfully) self-proclaimed as "a book for people who hate the very idea of self-promotion". (It's also a catalyst for resuming my blog) www.goodreads.com/book/show/18...

submitted 79 days ago • 0 comments

The strongest benchmark I'm using these days comes from which model are people - from all sorts of backgrounds and use-cases - experimenting with. And lately, DeepSeek is the one. By far.

submitted 95 days ago • 0 comments

Most of the talk around AI and energy use refer to an older 2020 estimate of GPT-3 energy consumption, but a more recent paper directly measures energy use of Llama 65B as 3-4 joules per decoded token. So an hour of streaming Netflix is equivalent to 70-90,000 65B tokens. arxiv.org/pdf/2310.03003

submitted 105 days ago • 9 comments

It’s algo a great tool for simulating a future discussion because we can make it simulate the other side’s point of view. And the more we “prepare” the llm with the other side’s pov, the better the discussion simulation, and the better we see their reasoning. A positive feedback loop.

submitted 104 days ago • 0 comments

The most influential book I've read in 2024 was Creativity Inc. I've come across so many parallels between AI development, team leading and this book. Some things I've incorporated in daily work and, mostly, is about allowing people's creativity to thrive by embracing the fear of the unknown.

submitted 106 days ago • 2 comments

Tried the github.com/yetone/avant... plugin and I'm impressed to say the least. Important note: I have never tried Cursor IDE before, but tried all sorts of different AI plugins before and this one is impressive. Also easy to configure. I (also) need to write about this...

submitted 133 days ago • 0 comments

@kennethenevoldsen.bsky.social opdaterer sin liste over åbne sprogmodeller som er state-of-the-art på dansk 🇩🇰 huggingface.co/collections/... Som han skriver på DDSC Slack, så skyldes den fremgang der har været nye danske datasæt af høj kvalitet + internationale model releases 💪 #dkai

submitted 134 days ago • 1 comment