Max Puelma Touzel

2 years ago I asked myself what work I could do that might have impact as the climate crisis deepens. Social/economic phenomena seems central to solutions. So, I started a move from NeuroAI/RL to AI-powered social science. I want to share a bit of my trajectory since. 🚀🧵

📰FIIR Workshop Alert! Future of Information Integrity Research 🎯Submit abstracts/papers (Tiny and Regular) in ACM PDF format → ⚠️Important: 1⃣Indicate your chosen track at the top of your PDF 2⃣Demo papers must include a link to the demo 📅Deadline: January 6th, 2026

@ryankatzrosene wildfires at low latitudes are much more locked to daily cycles? is the heat of the day such a driver? (more so than say the cold of northern Canada

From: svs.gsfc.nasa.gov/5572/

🤓 Man, this is some very cool stuff from NASA! It visualizes one month of aerosol particles moving throughout Earth's system (from August 2024). Green=Sulfate Red=Carbon Purple=Dust Cyan=Sea Salt [Keen readers will note the pun above😉]

evaluation is a killer app of collective intelligence within this AI revolution. Also maybe a gateway to alternative development models/revolutions?

we're in our evals arc. some thoughts from @divyasiddarth on our reasoning and approach. blog.cip.org/p/notes-on-bui…

@ttkeskinturk what's the standard reference for what made people move on? Is it Sewell's "A Theory of Structure: Duality, Agency, and Transformation" (1992)? I'm trying to understand the field as an outsider.

fun

Speculative drawing

@ryankatzrosene @JAtanackov Man, looks pretty good to me. Teach me your ways! Was looking up on Sunday at 9pm and saw nothing.

And was interesting to note that this came on the 166th Anniversary of Carrington Flare! (Thanks @JAtanackov for noting that!)

This ended up being the best we could do with this CME in the Ottawa-Gatineau region. After this substorm, Bz went (and held) North for hours. And then clouds got even heavier. Craving a clear-sky, sustained, powerful auroral experience!

Thought-provoking, not only for the engineers, but to the physicists who admit identifiability of microscopic dynamics for most complex systems is practically impossible.

*very* excited to share a new *efficient* method for learning *marginally stable* and NONLINEAR dynamical systems, w. brilliant students Evan Dogariu and Anand Brahmbhatt @AnandBrahm15501: arxiv.org/abs/2508.11990 more info in thread

🥈"A Guide to Misinformation Detection Data and Evaluation" received Best Paper Runner Up at KDD Datasets and Benchmarks Track! If you're at KDD, check out our presentation and poster! Monday 18:30 - 21:00, Poster 170 Wednesday 15:30 - 17:30, Room 604

Today (w/ @UniofOxford @Stanford @MIT @LSEnews) we’re sharing the results of the largest AI persuasion experiments to date: 76k participants, 19  LLMs, 707 political issues. We examine “levers” of AI persuasion: model scale, post-training, prompting, personalization, & more 🧵

@andrew_leach Not in the field/an expert, but also have this apparently inaccurate assumption about the industry's precarity from news coverage over the years. You indicate the industry itself propagates this narrative. Can you say more about that? It seems counterproductive to their interests

These calcs account for taxes, royalties, etc. If you decide to leave those bbls in the ground, you're foregoing the returns plus the taxes and royalties that would have come along with them. The industry often pretends it's very precarious, and so easy to ignore this value. /fin

My sense is that Canada's oil industry has done too good a job at convincing Canadians, including many Albertans, that the industry is a risky, marginal, barely profitable adventure that could all collapse in a heap if someone sneezes nearby. That's not the case. 1/

In case folks are interested, here's a video of a talk I gave at MIT a couple weeks ago: youtu.be/FmN6fRyfcsY?si…

@xuanalogue You are likely already familiar with the information bottleneck and more broadly rate distortion theory. This seems like it no?

I'm also curious what the keyword is for something like "utility-weighted lossy compression", where you want to make sure some bits of information are better preserved over others. Might be useful for fair aggregative summarization of multiple viewpoints.

Is there work that uses the lens of lossy compression & information theory to analyze and improve text summarization? Given that LLMs are probability models it seems quite possible to turn that theory into practice.

@michael_nielsen What does such a pledge mean in practice? Aren't automated gun technologies already quite advanced? Presumably, the main current hurdle to weaponized robots is not the weapon control, its the robot. Once the robot is there, it may be relatively straightforward to put a gun on it?

The history of Boston Dynamics is fascinating. Note the no-weaponization pledge at the end. I'd always assumed it was basically an extension of the US DoD, but it's actually owned by a South Korean company(!) Of course, pledges can be retracted, but the later company history surprised me

@corefpark Interesting results. Since you mention phase transitions, I am thinking of how the width limit works here, e.g. in light of the theoretical results of that models with finite width can nevertheless encode an infinite context window. arxiv.org/pdf/2408.01367

We further investigate how this critical context size for an in-context transition scales with graph size. We find a power law relationship between the graph size and the critical context size. Maybe Percolation? :) 12/n

New paper! “In-Context Learning of Representations” What happens to an LLM’s internal representations in the large context limit? We find that LLMs form “in-context representations” to match the structure of the task given in context! 1/n

@VoiceOfFranky But isn't it possible that people are identifying in other ways that are meaningful to them?

@mptouzel I believe it is evidence of an age of nihilism, or something very close to that . We also find that values are declining as is attachment to symbols of national identity . Strongest decline in under 35 Canada .

@DanielaGabor As in under the boot of the EU as oppressor?