Davide Eynard

New any-agent release (v0.24.0) is out.

▪️ AgentJudge returns full eval traces
▪️ Serve agents via MCP
▪️ Use agents as tools
▪️ Cost + token tracking
▪️ A2A protocol compatibility

We're making agents more composable + observable.

⬇️ Links in next post

GitHub - mozilla-ai/wasm-agents-blueprint: Testing WASM-powered AI agents Testing WASM-powered AI agents. Contribute to mozilla-ai/wasm-agents-blueprint development by creating an account on GitHub.

🌐 New blueprint: AI agents that run in your browser via Webassembly & Pyodide. Just open an HTML file and go.

They work with OpenAI API compatible models or local models via Ollama/vLLM.

Take a look : github.com/mozilla-ai/wasm-agents-blueprint

A couple comments about what I’ve been learning about algo engineering, but first off: if you’re interested in helping us improve Discover and all of our algorithmic systems, we are hiring algo engineers jobs.gem.com/bluesky/am9i...

8 Hard Truths About Our Attention Crisis The path to reclaiming control is closer than you think

We know platforms like Facebook, Instagram, and X are designed to keep us scrolling, but have we fully understood what that costs us collectively?

I had the opportunity to chat with @andypiper.me about this. Here's my post about our discussion

open.substack.com/pub/kathleen...

Ready to build AI agents? Here’s the first Any-Agent Cookbooks!
Cookbook 1: Build your first AI agent
Cookbook 2: Create an interactive booking assistant with MCP

Explore the cookbooks:
- mozilla-ai.github.io/any-agent/co...
- mozilla-ai.github.io/any-agent/co...

YouTube video by Progress Reimagined with Kathleen Chu Breaking Free From Social Media Silos With The Fediverse

"We've spent 15 years sleepwalking into this digital landscape," says @andypiper.me

If you've found refuge in the Fediverse, you already understand why that matters.

Watch and share
www.youtube.com/watch?v=GZ4R...

🎥 Watch the Mozilla.ai x @eleutherai.bsky.social webinar recording! See experts demo the Audio Transcription & Document Conversion Blueprints, including setup, best practices, and tips.

Recording: mozilla.hosted.panopto.com/Panopto/Page...

Explore more: blueprints.mozilla.ai/all-blueprints

YouTube video by Progress Reimagined with Kathleen Chu AI Myths DEBUNKED: You DON'T Need Huge Models & Massive Data

AI Myths Debunked!

Challenging popular AI assumptions! 🤖 @aittalam.bsky.social of @mozilla.ai breaks down why:

1. You DON'T need massive computational power and data sharing for effective AI solutions
2. Generic "leader board champion" models AREN'T always best for YOUR specific needs

YouTube video by Progress Reimagined with Kathleen Chu Here's what your kids are missing when using AI, warns AI engineer and researcher

I finally got the excuse to ask @aittalam.bsky.social from @mozilla.ai the burning question that’s probably on the minds of every parent:

How does having access to AI affect our kids' learning abilities? 😁

youtube.com/shorts/YJMA8...
#ai #leanring

YouTube video by Progress Reimagined with Kathleen Chu Why Think Small Is The Future of AI: Mozilla.ai Engineer on Building Your Own Timeline

🎙️ Our ML engineer @aittalam.bsky.social explains "Why Thinking Small Is The Future of AI" on @kathleenchu.bsky.social 's Progress Reimagined podcast.

www.youtube.com/watch?v=HDM8...

... small AI and some of its principles including:

- if you don't need it, don't use it 🙂
- own your algorithms
- run them on your hardware
- use them to serve *your* needs

@kathleenchu.bsky.social has been a great host and I ❤️ her work, which I suggest you to check out regardless of my AI rants 😉

Progress Reimagined Progress Reimagined explores innovative ideas, economic models, and technologies that enhance social equity, environmental sustainability, and overall well-being. Click to read Progress Reimagined, by...

"Your Digital Life, Your Rules": kathleeninweb3.substack.com/p/34654196-8...

What if we could apply a "think small" approach to AI?

@kathleenchu.bsky.social was so kind to invite me to her podcast "Progress Reimagined" and we had a great chat about small AI and some of its principles including:
🧵

Breaking: Pleias releases a new generation of small reasoning models for RAG and source synthesis. Pleias-RAG-350M and Pleias-RAG-1B come with built-in support for source citation, SOTA performance and an accuracy comparable to models ten times their size. huggingface.co/PleIAs/Pleia...

This #NationalLibraryWeek, watch a preview of an upcoming documentary from Structure Films about the #InternetArchive's fight to defend online freedom & public access to knowledge.

Full clip ➡️ vimeo.com/1056104575/d...
Support ➡️ fiscal.thegotham.org/project.cfm/...

Meet Any-Agent, a Python library that lets you build and switch across agent frameworks.

Try it: github.com/mozilla-ai/a...

We built it because we were struggling. Now we’re sharing.

🔹 Unified interface
🔹 Logging with OpenInference
🔹 Built-in eval

Blog: blog.mozilla.ai/introducing-...

Mozilla.ai is going to university!

This Thursday, Mario Cariñana will speak at Universitat Politècnica de València (UPV) on open-source AI, our work at Mozilla.ai, and building trustworthy AI systems.

¡Nos vemos el jueves!

Oh yes, then custom bayes classifiers (I think there's a reference in Vicki's link) are definitely a fun thing to play with!
Another thing I'd find interesting are efficient built-in binary set operators (e.g. re-ranking via dot product, set similarity... Even mapping might be seen as a set op? 🤔)

… if you want you can train a small MLP on top of them. Yeah I know this boils back to the learned embedding case you were trying to move from… But if you learn them yourself it’s more fun 🙂

If you are not constrained to NLP there’s a lot you can play with in CV (eg descriptors such as SIFT, HoG etc). Handcrafted (especially visual!) descriptors are easily interpretable + relatively cheap to calculate, still quite limited for ad-hoc tasks but…

The demo is not local (which is what BYOTA is all about!), but I hope the message is still clear: we *can* run useful TL algorithms locally, without sharing our own data. We don’t need a lot of data and compute, if our goal is not to convince people to spend more time on a given platform.

I already talked about why we should own our timeline algorithms (fosstodon.org/@mala/113967...) and promised a blog post soon… Here’s the long awaited post AND a whole new @mozilla.ai Blueprint 🔥

Live demo: mozilla-ai-byota.hf.space

#BYOTA

Benchmarking DeepSeek R1 Models using Lumigator: A Practical Evaluation for Zero-Shot Clinical Summarization New state-of-the-art models emerge every few weeks, making it hard to keep up, especially when testing and integrating them. In reality, many available models may already meet our needs. The key quest...

What’s the smallest model that gets the job done? 🐤

We benchmarked DeepSeek R1 for zero-shot clinical summarization using Lumigator 🐊 to test efficiency, accuracy, and cost trade-offs.

Read more: blog.mozilla.ai/benchmarking...

At Mozilla Builders Demo Day in SF today 🛠️

Catch @jularase.bsky.social and Stefan French presenting 🐊 Lumigator and our 🔵 @mozilla.ai Blueprints!

Stop by, check out the demos, and grab some Mozilla.ai swag!

#MozillaAI #MozillaBuilders #OpenSourceAI