Gilles Louppe

Today is international Frauenkampftag ("women-fight-day") how we call it in German to highlight that women have been fighting for equal rights since the 1910s. More than 100 years later, women still have to fight for opportunities and this post is for all of you:

Simulation Based Inference Blueprint The 1st Simulation-Based Inference Blueprint workshop will take place on the 26th and 27th February, 2026 at CERN. The workshop aims to summarize the current status of statistical techniques and softw...

Happy to see a Simulation-Based Inference Blueprint workshop being held at CERN. First papers were in 2015, and it took several years to get the first physics papers out. Hope to see it become more mainstream.
indico.cern.ch/event/160067...

I have added a new tutorial on discrete diffusion models:
github.com/gpeyre/ot4ml

I have been using Openclaw, in its own user space, no sudo, no access to any of my data, most things disabled (incl. skills). Risks are still high.

Quite disagree on this. Writing is helpful for thinking, but those are personal notes. A paper is something else. A paper is a message to the readers, not to yourself.

All that said, I'm genuinely bullish. The core experience of directing work from your phone while an agent executes on a real machine is something new. We just need to teach them social norms, and for them to know when they don't know. You know, the easy stuff.

Performance is wildly inconsistent. One moment it's brilliant, the next it's making mistakes a beginner wouldn't make. It's like working with someone who fluctuates between senior engineer and "first day on the job" energy at random.

They also love taking initiative. Sounds great until you realize "initiative" means doing things you never asked for and definitely didn't want. A human assistant would check first. The agent just assumes it knows best. It does not.

Agents sometimes make Big Decisions (running long scripts, deleting files, reorganizing everything) without understanding consequences. A human intern would at least hesitate before `rm -rf`-ing a project folder. Agents just go for it with full confidence.

You'd think an AI agent would be well-organized. Nope. The workspace turns into a dumpster fire shockingly fast. Eventually the agent can't even find its own files and just... starts over. No learning, no memory. Just vibes.

The context window is still a real bottleneck. Compressing conversation history = losing information. You end up repeating yourself constantly, like talking to someone with amnesia who insists they're fine.

I've been running AI agents 24/7 on a spare machine for a month now. Haven't had this much fun tinkering in years: prototyping random ideas from my phone while on the couch, in the car, wherever. This new way of using computers is genuinely exciting. But I've learned some things. 🧵

In an effort to create a more accurate climate model, scientist Tapio Schneider teamed up with researchers at Google, who generated a library of 8,000 ultra-detailed cloud simulations.
www.quantamagazine.org/climate-phys...

Pluton avec son atmosphère

Quelle image incroyable de Pluton côté nuit par la sonde New Horizons.
Elle révèle l'atmosphère de la planète naine et quelques reliefs à sa surface !

Always appreciated when brainstorming research with Claude and it suggests reading one of my own papers 😅

University of Liège Scientists Reveal Similarities Between Auroras on Ganymede and Earth Des observations menées sur la plus grande lune du système soaire révèlent que les processus physiques à l'origine des aurores sont universels et ne se limitent pas aux planètes.

An international team of astrophysicists, led by researchers from @universitedeliege.bsky.social, has published new observations of Ganymede revealing a striking similarity between its auroras and those on Earth. A 🧵 about 🔭. #planetsci #jupiter #juno #aurora www.star.uliege.be/cms/c_136561...

Maths and biology conspired to make something beautiful: the world’s largest floating leaf (genus Victoria), spanning up to 3 m across. We decoded this enigma with mathematical modelling to show the lattice is an economy of material– the most structurally efficient way to cover a large surface area.

My professorship promotion application has reached ERC-comparable word count. Someone please intervene. [Also, let's ban Word once and for all 😱]

Next time you want to use a vibe coded compile think about benchmarking the resulting binary. ccc on SQLite is up to 150k slower on some SQL queries and the benchmark took 2 hours vs 10sec for GCC compiled harshanu.space/en/tech/ccc-...

A small polymerase ribozyme that can synthesize itself and its complementary strand The emergence of a chemical system capable of self-replication and evolution is a critical event in the origin of life. RNA polymerase ribozymes can replicate RNA, but their large size and structural ...

How could a simple self-replicating system emerge at the origins of life? RNA polymerase ribozymes can replicate RNA, but existing ones are so large that their self-replication seems impossible. Could they be smaller?

Excited to share our latest work in @science.org on a new small polymerase.
1/n

So how does it work? Meet Miniature (mini + IA, get it?), a complete AI agent in ~225 lines of Python. No frameworks, no abstractions. Code: github.com/glouppe/miniature

🎉 Announcing TabICLv2: State-of-the art Table Foundation Model, fast and open source

A breakthrough for tabular ML: better prediction and faster runtime than alternatives, work by Jingang Qu, David Holzmüller @dholzmueller.bsky.social , Marine Le Morvan, and myself 👇

GitHub - glouppe/miniature Contribute to glouppe/miniature development by creating an account on GitHub.

Voilà github.com/glouppe/mini...

But you are right, i'll add a disclaimer that running this piece of code does not come without risks.

See it as a teaching opportunity! I will extensively discuss over this code :-)

That's it. Tool use is structured text generation. Agency is a loop. Memory is context. Every AI agent framework is just a dressed-up version of this.

Memory is the simplest trick. A plain text file (memory.txt) loaded into the system prompt. The model can write to it via a "remember" tool. No vector database, no embeddings. Just text injected into context. Stateless model, persistent memory.

The agent loop (runtime.py:104) is simple: take user input, prepend system prompt + conversation history, send to LLM. If the output contains a JSON tool call, parse it, execute the function, feed the result back, and let the LLM respond again. If no tool call, the turn is over.

To make the model useful, you steer it. That's what the system prompt does (runtime.py:47). It tells the model who it is, what tools it has, and how to format tool calls as JSON. The model doesn't "know" it has tools, it just follows instructions in its context window.

It starts with the generation loop (model.py:26). A transformer autoregressively predicts the next token given all previous tokens. Sample one, append it, repeat. Add KV caching to avoid recomputing everything each step.