Sofie Van Landeghem

Genuine question: does the newsletter also cover developments of open-weights LLMs?

How can we better understand how models make predictions and which components of a training dataset are shaping their behaviors? In April we introduced OLMoTrace, a feature that lets you trace the outputs of language models back to their full training data in real time. 🧵

ChatGPT telling me it's giving me a **curated** list is giving me the creeps.

GitHub - fastapi/fastapi: FastAPI framework, high performance, easy to learn, fast to code, ready for production FastAPI framework, high performance, easy to learn, fast to code, ready for production - fastapi/fastapi

Here's FastAPI 0.115.14 🚀

@patrick.wtf fixed support for unions of @pydantic.dev models in Form data. 🐛🎉

(This will make it easier to implement single sign-on 😉)

github.com/fastapi/fast...

GitHub - fastapi/fastapi: FastAPI framework, high performance, easy to learn, fast to code, ready for production FastAPI framework, high performance, easy to learn, fast to code, ready for production - fastapi/fastapi

Here's FastAPI 0.115.13 with a bunch of bug fixes and tweaks 🎁

Special thanks to Yurii for all the work handling discussions and PRs! 🙌

github.com/fastapi/fast...

We’ve just made it even easier to build with our open models and datasets ✨

What’s in our docs platform?
⚙️ Setup instructions
🚀 Hosting + deployment
⭐ Integration examples
⚡ Optimization tips

2025-05-31-pycon_italia Keynote given at PyCon Italia 2025: https://2025.pycon.it/en/keynotes/data-doesnt-lie-but-it-can-mislead-how-to-ensure-integrity-of-your-machine-learnin…

Data doesn't lie, but it can mislead!

At PyCon Italia 2025, I gave a keynote about how to ensure integrity of your ML applications by inspecting your data, developing good baselines and interpreting your performance numbers.

Slides: speakerdeck.com/sofievl/2025...

Video: youtu.be/qXx-4m-Z0q4?...

Looking forward to presenting my keynote tomorrow at PyCon Italia. If you're into data and/or Formula One, come join at 5PM ! @pyconit.bsky.social

I’ll give a talk about @pydantic.dev AI and Logfire today 12:25 at @pyconit.bsky.social 🙏

Every now again it’s useful to repeat advice about accessing papers that are behind a paywall that excludes you. Email the author. My estimate is that 90% of academics are so thrilled that a living, breathing, possibly even reading, person shows interest that they will swiftly send you a copy.

GitHub - fastapi/typer: Typer, build great CLIs. Easy to code. Based on Python type hints. Typer, build great CLIs. Easy to code. Based on Python type hints. - fastapi/typer

Here's a Typer release for you: 0.15.3 🥟

This has a couple of bug fixes by @oxykodit.bsky.social 🙌

github.com/fastapi/typer

move slowly and build things

At PyCon Italy, I'll be talking about a topic close to my heart: data quality and how your data modeling decisions impact your downstream AI/ML/NLP applications 🕵️

Hoping to see many of you Italy!

It is *so critically important* that unit tests be *completely understood* by the person writing them. That is your safety net! That is the foundation upon which the rest of your code is built! Let copilot generate all of your code as long as you write the unit tests! 😭

Hey @jbaert.bsky.social, kunnen jullie met Nerdland geen Belgische nationale overstapdag organiseren (WA -> Signal), zoals ze in Nederland hadden?

Tijdens het Nerdland festival ofzo? 😎

This blog post will hit its 25th birthday in about a week. 25 years, folks. That's enough time for pretty much everybody to have read it

www.joelonsoftware.com/2000/04/06/t...

Gemma 3, OLMo 2 32B, and the growing potential of open-source AI Leading open-weight models and the first open-source model to clearly surpass GPT 3.5 (the very last version).

Up to 5/15 top models in ChatBotArena being open weights. Yes, there are caveats as always, but this is THE BEST opportunity people have ever had to build things on open models.

R1, Gemma 3, DeepSeek V3, QwQ, Command A 🫡
www.interconnects.ai/p/gemma-3-ol...

I get your point, but I don't think a single datapoint constitutes a "benchmark" 😜

You thought I was joking? I'm not, come and work with me, for real.

It's potentially fun and loaded with emojis. 😅

If you like FastAPI and know a lot about Kubernetes, cloud, observability, all that stuff, send an email.

And send this to your badass SRE/DevOps friend. 🚀

Now I'm looking for someone with 7+ years of FastAPI experience 🤭

Jokes aside, I'm doing interesting new things and want some extra help full-time 🚀

I'm looking for tons of skills and knowledge in SRE / DevOps / Platform / Cloud / name-it ...and of course, FastAPI

Email jobs@fastapilabs.com

Some tips on prompting the new Gemma 3 models, instruction tuned. At the Google Gemma Day research release.

Training recipe for Gemma 3

Super excited to start cooking with the newly released open-weights Gemma 3 models and their long context window 🤩

So exciting to be in Paris at the launch today! #LLM #gemma #ai #developers

I'll be in Paris for the next two days, attending Google's Gemma Developer Day. Hoping to catch some interesting talks and people! #google #gemma #llm #ai #opensource

GitHub - fastapi/sqlmodel: SQL databases in Python, designed for simplicity, compatibility, and robustness. SQL databases in Python, designed for simplicity, compatibility, and robustness. - fastapi/sqlmodel

Here's SQLModel 0.0.24 with support for Python 3.13 🐍🚀

This is the last release to support Python 3.7 and Python 3.8 🧟

...should have dropped support for them a while ago 😅

The backwards compat took a LOT of effort from @oxykodit.bsky.social, thank you! 🙌

github.com/fastapi/sqlm...

I have a couple of tricks I use with my open source projects, starting with @fastapi.tiangolo.com 🤓

Someone once called it "documentation driven development" 📝

If you're curious, here's the short tale of a bug that shows that 🐛👇️

WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, includin...

Huge shoutout to colleagues at Google & Unbabel for extending our WMT24 testset to 55 languages in four domains, this is game changer! 🚀

I really hope it puts the final nail in the coffin of FLORES or WMT14. The field is evolving, legacy testsets can't show your progress

arxiv.org/abs/2502.124...

GitHub - fastapi/sqlmodel: SQL databases in Python, designed for simplicity, compatibility, and robustness. SQL databases in Python, designed for simplicity, compatibility, and robustness. - fastapi/sqlmodel

New SQLModel 0.0.23 with a couple of bug fixes 🐛

Thank you @oxykodit.bsky.social for all the help! 🙌

github.com/fastapi/sqlm...

A great TED talk by Thomas Wolf about how open-source and open-weight models contribute to resiliency of applications and the AI eco-system in general 🚀