Notes

🚨 Announcing Pushin 🚨 Now officially: I'm announcing Pushin.eu, a git hosting platform built and run entirely in Europe. Think GitHub/GitLab but European, better, and faster. Built with #ElixirLang and #RustLang. peterullrich.com/announcing-p...

There has never been a better time to learn software engineering.

Very much worth a read

As a search person, I get a front row seat to AI team problems. Honestly they look like search team mistakes from 10 years ago * Ignoring evals * Thinking retrieval is 'one thing' (in this case RAG) * Thinking in chunks, not metadata softwaredoug.com/blog/2026/08...

I've had the Fairphone 6 for a year now and I'm in general quite satisfied. It took some time to go from iOS to Android, and I still want to try e/OS. www.theguardian.com/technology/2...

When the price per token of proprietary models stops being underpriced, then the edge is sitting on the side of the ones who are running local models.

💪

pedromadruga.com/blog/domain-... Domain Driven AI

pedromadruga.com/blog/domain-... Domain Driven AI

More on the github topic

🚨I've just released Sentence Transformers v6.0! MultiVectorEncoder joins the family: ColBERT-style late interaction models are now a first-class model type, for training, inference & interpretation, alongside dense, sparse & reranker models. Big thread 🧵

Check out our latest open model data -- which models are mentioned in every arXiv ML paper since ChatGPT. A very fun weekend project with Codex! From idea is full dataset in <72hours. ~500K papers processed 😎 Plot: dashboard.interconnects.ai?axrange=all#...

Just found out about this and it's such an interesting concept to read it together. Following the group closely now and sharing for wider reach. Awesome stuff! 👍

A better way of explaining GitHub's outages. Source: @benjdicken on twitter

AI development should follow domain knowledge. I'm calling it Domain Driven AI and I wrote about it. (No AI was used to write this blogpost which basically took me 4 weeks to write it 🥲) pedromadruga.com/blog/domain-...

Its going to be an era of contradictions. Polls will show everyone hates AI overall but also everyone will secretly use AI all the time for lots of stuff. AI companies will be underwater in public opinion but also many people will feel strongly attached to their own favorite model & care about it.

Great use of Claude

Opencode VS PI VS claude code VS ... reminds me of the coices between Sublime VS vscode VS neovim VS ...

The twitter bubble of silicon valley is saying that Sol has surpassed Fable but I'd really like a better source of information about it.

Opencode looks good so far, although I have only been using it for a couple of days. I chose it over PI because I don't need a barebone option. Most importantly, I can easily switch from different providers and also between cloud and local models easily. Anthropic models seem to be lagging...

Luddites are at it against AI coding, but it feels they're less by the day. I wonder what's going to be their next victim.

For the past year my setup includes a combination of github worktrees, tmux, kitty tabs, with a configuration that syncs all of these together so I can plan, build and review in parallel. Similar to what herdr is, which I have tried now. Recommended.

If we.all agree that a lot of our learnings come from side projects and experiments, why not have teams increase doing this research? This would make people more knowledgeable and a benefit to any organization. Research and experiments are essential in teams work on AI-based products.

Something that I've stayed completely away from when using LLMs is analysing data from a dataset.

Stifling innovation will very rarely be a positive thing.

After SIGIR2025, I selected the best talks (imo) and summarized them using the Feynman Technique. The output was an overview explained in an incredibly straightforward way. I shared them all here github.com/pmadruga/sig... This year, I'm skipping SIGIR2026 but will probably do the same.

"AI is a tool, just like other tools we use. And it's clearly a useful one." Linus 🙏

you 👏🏼 cannot 👏🏼 automatically 👏🏼 detect 👏🏼 plagiarism > the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text link.springer.com/article/10.1... 1/

Software engineering without innovation is just technical complacency.

MTP is wild. Just a few days after the llamacpp support.

Wait, what? source: www.anthropic.com/engineering/...

So much looking forward to put this on a Raspberry Pi or similar.

Mistral, do your thing.

I need to know the inference speed on a raspberry pi because these sizes look great.

Rust implementation for Speech-to-Text based on Qwen3 models by Michael Yuan * Self-contained binary build — no external dependencies * Uses libtorch on Linux with optional Nvidia GPU support * Uses MLX on MacOS with Apple GPU/NPU support github.com/second-state...

LibreOffice it is. "Danish government agency to ditch Microsoft software in push for digital independence"

Could be interesting, macOS only though but with a small footprint. GitHub - ggml-org/LlamaBarn: A cosy home for your LLMs.

LLM cloud inference dominates usage, but should it? Local models and accelerators have improved massively over recent years. Perfect routing to best local model "reduce energy consumption by 80.4%, compute by 77.3%, and cost by 73.8% versus cloud-only deployment" arxiv.org/pdf/2511.07885

I went from Obsidian after a few years, to Notesnook, to Standard Notes. I am back using Obsidian. I can configure it enough to still keep it minimal while addressing all my use cases. I do notice that the Android version is slower than the iOS one. But manageable.

These are the engineers I like to work with. Boris is (again) on point.

A company's "AI Winter" starts when its AI-based product only has focus on engineering and zero on AI innovation. However, trying to avoid engineering debt creates AI debt. And AI debt has deeper (& longer) consequences due to rapid changes happening in the field. Aim for a balance, always.

The claude code agent teams feature seems great in theory but there's an upper limit on the cognitive load of reviewing every change each agent makes. I'm still reviewing every single change, yes.

Today's reading item

💯

On Github Copilot's memory system and what their engineering wrote fosstodon.org/@pamelafox/1...

The best compliment i can give OpenAI's Codex 5.3 is that it feels way more like Claude Code

Quilicura, Chile, one of the communities I wrote about in EMPIRE OF AI, has launched a brilliant initiative to inspire more responsible AI prompting. Today, don't use AI; ask the townspeople instead: quili.ai. So heartened to see this creative act of resistance.

The future of software engineering isn't less human, but more focused on higher-level problem-solving. Engineers will evolve to master AI tools, leveraging them to build more robust, thoughtful solutions. #FutureOfWork 6/6

This can't be good for Europe.

For all the hate that LangGraph gets, there is a hundred times more people using it in battle-tested grounds (such as gitlab, elastic, klarna, cisco and so on). Also, it's silly to hate a framework.

With the amount of AI-assisted tools for development coming out - thus impacting teams' development practices from the ground up - the concept of cross-functional teams seems ancient. A developer that understands these tools will hardly fit in the same team with one that rejects them vehemently.

Domain driven AI - where domain knowledge drives AI implementations - is the best approach for AI-based products.

Software engineering is, fundamentally, a solved problem. AI Science isn't. The ones who treat AI challenges as software engineering ones are doomed to fail.

This year, I've tried Helium and Waterfox browsers. I found the former to be buggy and there's something about Chromium-based browsers that doesn't bode well with me. And I've always been a Firefox user, since its Firebird days. Waterfox was perfect since the get-go. Both on Mac, Linux and Android.

The Going Dark initiative or ProtectEU is a Chat Control 3.0 attempt View Article | Join the HN Conversation Summary of HN discussion 🧵👇

This was also a year where I have been migrating to European-based services. Went from iCloud to Proton Drive, Gmail to Proton Mail and so on. Migrated from iPhone to Android, since I bought a FairPhone and am very happy with it. FP is much better than I anticipated!

Doing a "year in review" on the technologies I've tried/read about. One thing that comes to mind is the hostility towards MCP, especially by people who didn't try it. And I used to do Javascript, which got a lot of heat back in the day. MCP is just a Protocol, people.

I just read "AI Agents in Action". It's well-supported by code and drawings. Being quite heavy on OpenAI and Microsoft tooling is a drawback. Online reviews are pretty mixed as well. It's not a beginner-level book but needs more depth. Decent read, all in all. www.manning.com/books/ai-age...

📢 Launching EuroLLM-22B - a fully open, EU-made LLM supporting all 24 EU languages (+11 more)! 🤝 Built by a Europe-wide consortium with #HorizonEurope & @eurohpc-ju.bsky.social support. 🔓 Available now on Hugging Face: lnkd.in/e-xgQXTX 🔗 Learn more: lnkd.in/e6FdpYrN

Dolphin-v2 🐬 new document parsing model released by Bytedance huggingface.co/ByteDance/Do... ✨ 3B - MIT license ✨ Works on any document: PDFs, scans, photos ✨ Understands 21 types of content: text, tables, code, formulas, figures & more ✨ Pixel-level precision via absolute coordinate prediction

The influence of domain knowledge is huge when it comes to applied AI. Domain expertise can prevent overengineering and/or unnecessary costs. Working in the legal industry as an AI scientist it became obvious early on how crucial it is to build the bridge towards domain experts. Every day.

A chatbot that interacts with NeurIPS 2025 papers: neurips.zeroentropy.dev Went in to find Information Retrieval related papers and there's a few. Happy reading!

We tested one of the most common prompting techniques: giving the AI a persona to make it more accurate We found that telling the AI "you are a great physicist" doesn't make it significantly more accurate at answering physics questions, nor does "you are a lawyer" make it worse.

We have come full circle.

TITANS & MIRAS: real continual learning MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs) TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module let’s dive in! research.google/blog/titans-...

I am a fan of SDD but it often gets confused with Vibe Coding, specially in companies where innovation is met with resistance. This articles shares some criticism, yet by someone who actually tried it. I disagree with the article and it seems that HN people in thread do too. Give SDD a try!

I have been trying lefthook and damn this thing is fast. Also, the ability to run scripts is handy if you like to keep the config tidy.

💯 bsky.app/profile/veri...

“Before being technical, science is visionary” Carlo Rovelli, Reality Is Not What It Seems

Business-applied AI development requires a much higher level of experimentation than traditional software development. This is because AI development requires a very tight connection to domain knowledge. And each domain requires its own approach, for the most part.

A short (way too short) yet fantastic reading. Rovelli has the gift of explaining complex things simply. I have quickly moved to reading The Order of Time, another of his books. The way he explains time is just so accessible - thus mind blowing.

It’s reasonably amusing to see the resistance from some traditional software developers to some AI tooling. MCP for example, gets a lot of focus from tech doomers because it’s new. It’s Ok to not use those don’t fit a use case but until they try it it’s just preemptive criticism.

TIL!

The term Agentic Coding is something that resonates quite a lot. But I’ve heard Spec Driven Development and it is spot on. Both these are almost polar opposite to what Vibe Coding is (imo), even though all are AI powered. Also, for prototyping: vibe coding all the way.

An important component of the success of AI-based products is the ability to intertwine data science and (software) engineering. Enforcing a separation of those is a recipe for failure. When it works though, it feels like a dance: there might be toe-stepping but the song is the same.

When people ask what am I doing with so many Pi’s

this seems like a very good idea actually

Phi4 didn’t even try

So Anthropic is now the enterprise AI market leader. After having been playing with Opus 4 for the last month, this is rather unsurprising. Open-source LLMs adoption has stalled. Src: menlovc.com/perspective/...

Context engineering (CE) makes sense. It exposes one of the limitations of RAG: the retrieved chunks lack a structure, thus turning the prompt into a spaghetti of information (inc w/ rerankers). CE, by providing formatting guidelines, improves the final output - while making testing easier.

I've finally clean the dust of my blog and have a few nice blogposts coming in. I'll be writing mostly on Applied AI (scalability, techniques, technical leadership of a successful AI product, etc) from experience in the industry. Sharing my learnings here: pedromadruga.com/newsletter/

“In science if you know what you are doing you should not be doing it. In engineering if you do not know what you are doing you should not be doing it.” - Richard Hamming, The Art of Doing Science and Engineering

Lumo. Mistral it is. 👍🏼

Seems to be using only self-hosted models, according to their Privacy. Proton (@proton.me)

What are your favorite recent papers on using LMs for annotation (especially in a loop with human annotators), synthetic data for task-specific prediction, active learning, and similar? Looking for practical methods for settings where human annotations are costly. A few examples in thread ↴

One of the best parts of SIGIR2025 were the talks I had with the authors during breaks. So much information to unpack that can only be obtained by being there. Another great thing was all the amazing work presented.

Dust is off from my blog! pedromadruga.com/blog/art-of-...

🚨 New paper from us: Given they are trained on human data, can you use psychological techniques that work on humans to persuade AI? Yes! Applying Cialdini's principles for human influence more than doubles the chance of GPT-4o-mini agreeing to objectionable requests compared to controls.

On my way to #SIGIR 2025 - see you in Padova 🇮🇹 . 👋

Particularly relevant for speed and cost-saving purposes: post-training quantization seems to yield really good results. It seems that between binary, trinary and 4- and 8-bit quantization, seems 4-bit is the sweet spot since it it's as good as 8-bit. jina.ai/news/quantiz...

“Careless people” is next up on my reading list. 👇

"Developers and User Experience folks often step outside “their lanes” to contribute widely in getting work done." For AI development, this is also spot on since much of the sucess of AI applications comes from domain knowledge of the industry where the developer is part of. Recommended 👇

One of the tests I do to local LLMs is how they perform in portuguese from Portugal. Mind you there’s not a lot of content in this, especially when compared to portuguese from Brazil. Mistral performs quite well so far. Looking forward to test the new 24B model (also, Apache 2 license!).

“The best team is not made out of the best people but the right people.”

Bifrost seems interesting at least. An LLM proxy that claims to be 40x faster than LiteLLM, production oriented (ie handling high workloads) and using 68% less memory. Written in Go. www.getmaxim.ai/blog/bifrost... github.com/maximhq/bifr...

Incredibly excited to join #SIGIR 2025 in July. Hope to see you all in Italy and talk about information retrieval.

Kinda cool to see some (very) basics of DeepSearch by google, being opensourced. github.com/google-gemin...

Finally, an overthinker.

Europe’s choice is clear. To put science at the heart of its economy. To become the home of scientific freedom and collaboration. And to welcome talent from all over the world. I’m glad to present the first elements of our Choose Europe Initiative. → europa.eu/!TTbWbJ