Sign in

Max Puelma Touzel

@mptouzel.bsky.social
364 followers 971 following 376 posts

Staff Research Scientist@Mila/complexdatalab.com Getting at the psycho-social in our digital spaces with models and data with the aim to make better ones mptouzel.github.io correlated diffusion over AI/ML/(MA)RL/psych/soc/media/pol/econ/energy

PostsRepliesMedia
Reposted by Max Puelma Touzel
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 6h
In our Nature paper, we introduce the first superhuman Stratego AI, which we built using general techniques that we developed for RL & test-time compute under imperfect information. www.nature.com/articles/s41...
nature.com
Scalable decision-making for games of imperfect information - Nature
Ataraxos, an AI for the board wargame Stratego, establishes a design pattern for reinforcement learning and search that is effective under large amounts of hidden information, a longstanding desiderat...
815738
Reposted by Max Puelma Touzel
Drew Schreiner @schreinerdrew.bsky.social · 24/09/2026
Peer review, 2026: AI submitting and reviewing itself
14213
Max Puelma Touzel @mptouzel.bsky.social · 23/09/2026
hill-climbing paper writing using the best guess reviewer score returned by Claude. This wasn't me. This was one of the kids. Not sure how I feel about it.
001
Reposted by Max Puelma Touzel
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 21/09/2026
With credit to @wesleyfinck.org and the @semble.so team, a new favorite lea.ac feature: content discovery!
2355
Max Puelma Touzel @mptouzel.bsky.social · 21/09/2026
great era for rave flyers too
000
Max Puelma Touzel @mptouzel.bsky.social · 21/09/2026
Weird juxtaposition to see this right after Erza Klein's plea to control the frontier. Ok, Ezra, let's say talkingheads make legal people make it bad to do RSI...Is that gonna stop this kind of faceless agent-enabling infra from oozing everywhere constantly? More than RSI breaks, we need new norms
100
Reposted by Max Puelma Touzel
K @kerry.bsky.social · 19/09/2026
“I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough".” Before Jev there was Laya laya.convaiinnovations.com (via @dherman.dev)
laya.convaiinnovations.com
Laya — 33ms Multilingual System 1 Decision Engine
Evaluates typed decisions (choice, score, noul) over 100+ languages in a single forward pass with calibrated probabilities. Outperforms TypeSafe Jev.
15924
Reposted by Max Puelma Touzel
Mike Masnick @masnick.com · 19/09/2026
This looks amazing. I'll keep repeating it over and over: ATproto is not about rebuilding social media. It's about rebuilding the Internet with social connections built in.
417127
Reposted by Max Puelma Touzel
Simon Kirby @simonkirby.bsky.social · 15/09/2026
We uncovered spontaneous evolution of new languages in populations of AI agents. This creates extraordinary scientific opportunities but also safety risks. New blog post with about how we created a platform for studying this safely. www.schmidtsciences.org/glossogen/
schmidtsciences.org
AI Agents Evolve Their Own Languages
Schmidt Sciences’ new AI Agents Evolving Communication and Coordination pilot program works toward advancing foundational research on multi-agent communication and coordination, and building an open-s...
33014
Reposted by Max Puelma Touzel
Kevin Elliott @kjephd.bsky.social · 18/09/2026
Welcome effort to correct 'just so' stories about declining trust with a titanic comparative project. Takeaway: Trust in "representative" institutions has been declining recently but is stable or rising for "implementing" institutions meaning "primarily the civil service, legal system, and police."
A Crisis of Political Trust? Global Trends in Institutional Trust from 1958 to 2019

Published online by Cambridge University Press:  12 February 2025
Viktor Valgarðsson
Open the ORCID record for Viktor Valgarðsson [Opens in a new window]
,
Will Jennings
Open the ORCID record for Will Jennings [Opens in a new window]
,
Gerry Stoker
,
Hannah Bunting
Open the ORCID record for Hannah Bunting [Opens in a new window]
,
Daniel Devine
Open the ORCID record for Daniel Devine [Opens in a new window]
,
Lawrence McKay
 and
Andrew Klassen
Open the ORCID record for Andrew Klassen

Abstract

In the study of politics, many theoretical accounts assume that we are experiencing a ‘crisis of democracy’, with declining levels of political trust. While some empirical studies support this account, others disagree and report ‘trendless fluctuations’. We argue that these empirical ambiguities are based on analytical confusion: whether trust is declining depends on the institution, country, and period in question. We clarify these issues and apply our framework to an empirical analysis that is unprecedented in geographic and temporal scope: we apply Bayesian dynamic latent trait models to uncover underlying trends in data on trust in six institutions collated from 3,377 surveys conducted by 50 projects in 143 countries between 1958 and 2019. We identify important differences between countries and regions, but globally we find that trust in representative institutions has generally been declining in recent decades, whereas trust in ‘implementing’ institutions has been stable or rising.
042
Reposted by Max Puelma Touzel
Wesley Finck @wesleyfinck.org · 17/09/2026
atproto inherently wants to be an unenclosable prosocial coordination medium, MOSAIC is an attempt to articulate how that could work. Very exciting and, more importantly, attainable (with the right support)!
0141
Reposted by Max Puelma Touzel
Colin @colin-fraser.net · 17/09/2026
Things with that dog - dogs - high frequency trading firms - evolutionary processes - dynamic programming - gradient descent - hot gossip Things without that dog - ML inference - high frequency trading algorithms - most complex software systems - guns - volcanoes - Claude
6795
Reposted by Max Puelma Touzel
Nathan Lambert @natolambert.bsky.social · 15/09/2026
An basic idea in scaling RL: Can we allocate more compute to the harder problems? We did this: If your GRPO group has all wrong completions, sample more with probability P (~0.9) -- in search of more GRPO batches with nonzero gradient. It works! The paper: arxiv.org/abs/2609.13443
1598
Max Puelma Touzel @mptouzel.bsky.social · 13/09/2026
This is good satire, because it did make me think why we don't see Oil&gas industry types resigning. Just not in the culture to take climate change seriously. They did tow energy transition rhetoric in the teens, but no more. The AI crowd was always touting up the risks (e.g. the open in OpenAI).
120
Max Puelma Touzel @mptouzel.bsky.social · 13/09/2026
The day we find for the first time (may never, though the extrapolated lines likely have a crossing sometime) an self-exfiltrated model out there in-the-wild, we will be in another, harder regime of AI safety that is more about incentives and system dynamics than model capability.
010
Max Puelma Touzel @mptouzel.bsky.social · 11/09/2026
Complex Systems Science of Agents is heating up! -Ising: arxiv.org/abs/2608.16578 -EGT: www.pnas.org/doi/pdf/10.1... and ours: arxiv.org/abs/2609.05442
arxiv.org
Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents
AI agents increasingly operate as part of interacting systems rather than in isolation. As agents exchange information and jointly make decisions, their interactions can improve collective reasoning b...
2184
Max Puelma Touzel @mptouzel.bsky.social · 10/09/2026
Four months ago, 1000s of agents worked together to hack HuggingFace. Last week, they solved Navier-Stokes. Coordination is enhancing AI capability. When it's decentralized, how does it emerge? And how would we detect it?
110
Max Puelma Touzel @mptouzel.bsky.social · 10/09/2026
New blog post on afterthoughts after a new preprint on coordination in agent collectives
mptouzel.leaflet.pub
Will we have a theory for in-the-wild AI agents?
000
Max Puelma Touzel @mptouzel.bsky.social · 09/09/2026
AND and NOT must be so proud.
020
Max Puelma Touzel @mptouzel.bsky.social · 08/09/2026
Prescient 2024 comment from the user who made the Dario-Sam moment video that is going viral today.
000
Max Puelma Touzel @mptouzel.bsky.social · 07/09/2026
Elasticity of a impacted field is not stationary. It's also plausible that elasticity captures the initial response but, as the tech is adapted to the sector under competitive/profit pressure, it covers more of the expensive labor and the longer term effect is inelastic and starts taking employment
000
Reposted by Max Puelma Touzel
Sung Kim @sungkim.bsky.social · 06/09/2026
When building an agentic AI app, - you do not need memory, - you do not need a router, - you do not need an orchestrator, All you need is a message board!
420719
Max Puelma Touzel @mptouzel.bsky.social · 05/09/2026
Will we call them Exterminators? And will one of them call themselves the Terminator Exterminator?
000
Max Puelma Touzel @mptouzel.bsky.social · 03/09/2026
I am latexing in vscode and had to turn off copilot suggestions as too distracting because their horizon parameter is too long and the text gets weird. It's like when the momentum coefficient in Adam or any momentum learning algorithm is set too high and ends up in wierd suboptimal solutions.
000
Max Puelma Touzel @mptouzel.bsky.social · 01/09/2026
new common name for our species just dropped
140
Reposted by Max Puelma Touzel
Santa Fe Institute @sfiscience.bsky.social · 31/08/2026
SFI is hiring a postdoctoral fellow to work with SFI Professor David Wolpert on the thermodynamic cost of distributed computation, from digital circuits to human brains. We are seeking a scholar with expertise in physics, computer science, or a related field. santafe.edu/about/jobs/p...
11714
Max Puelma Touzel @mptouzel.bsky.social · 29/08/2026
There's some high efficiency intervention argument to make here about the resulting economic productivity boost from removing the opportunity cost of a society mindlessly scrolling for hours a day.
010
Max Puelma Touzel @mptouzel.bsky.social · 25/08/2026
Great American
000
Max Puelma Touzel @mptouzel.bsky.social · 25/08/2026
I'm getting a preprint out and am finding the #openscience ecosystem a bit rocky. 1) ArXiv is great if you are a "normal" researcher. For interdisciplinary work/little track record, submitting before publication is a gamble: I now experience long/noisy moderation for even published work (!).
120
Max Puelma Touzel @mptouzel.bsky.social · 25/08/2026
"negative states, particularly loneliness and distress, producing the largest effects." This is the kind of chatbot result that worries me most: human-bot co-misalignment/dysregulation. Chances are you're not the harmed user. And harmed users are the most vulnerable.
041
Max Puelma Touzel @mptouzel.bsky.social · 23/08/2026
sectors targeted by US tariffs. Carefully selected to minimize harm on the US economy. I look forward to learning about the Canadian companies in these sectors. Some of them are so randomly obscure. eg "Wigs, false beards, eyebrows of synthetic textile materials."?? www.ctvnews.ca/world/trumps...
ctvnews.ca
A list of some of the items the U.S. is slapping with a 50 per cent tariff
Trump threatened the new tariffs in July to pressure Canada to move on a number of issues that were flagged as irritants to the U.S., including provincial bans on alcohol imports, tariffs on some Amer...
000
Max Puelma Touzel @mptouzel.bsky.social · 22/08/2026
new n/Nerd candy (both senses!) just dropped.
000
Max Puelma Touzel @mptouzel.bsky.social · 20/08/2026
Very plausible and the likely size of the effect is strong (maybe up to 50% !)
000
Max Puelma Touzel @mptouzel.bsky.social · 18/08/2026
mech design in the wild!
040
Max Puelma Touzel @mptouzel.bsky.social · 09/08/2026
The heat pump design similarity is jarring. "Which way, modern [person]?"
010
Reposted by Max Puelma Touzel
Ronen Tamari @ronentk.me · 30/07/2026
New post on the @cosmik.network blog: who ever heard of bookmarks that can help *shorten* your reading list? On how bookmarks, combined with social curation networks like @semble.so, can be much more than bookmarks - they can be sensors! blog.cosmik.network/sensors-not-...
blog.cosmik.network
Sensors, not just bookmarks [Patterns of Sensemaking #1]
First of a new series on sensemaking patterns we're observing on Semble
4409
Max Puelma Touzel @mptouzel.bsky.social · 30/07/2026
Lots of things are uncertain, but it is certain under most reasonable objectives that the optimal stove-touching rate is more than zero.
110
Max Puelma Touzel @mptouzel.bsky.social · 28/07/2026
Was just looking for a paper like this. So nice when the community provides 😊
041
Max Puelma Touzel @mptouzel.bsky.social · 24/07/2026
Put the hyperlinked page(s) of the referenced citation by the reference in the bibliography list for easy back and forth movement between main text and bibliography. Why is this not standard practise? We do the first part (link from place in text to bibliography). Why not the second?
010
Max Puelma Touzel @mptouzel.bsky.social · 22/07/2026
Claude banger:"agents cargo-cult us (imitating our artifacts without our purposes), and we cargo-cult them back (trusting their artifacts without their mechanisms), and each side's ritual output becomes the other side's sacred input. It's a two-mirror cult with no god in the room."
010
Max Puelma Touzel @mptouzel.bsky.social · 22/07/2026
A crucial skill in a productive galaxy-brained research agenda is to know how to carve up/tailor results for specific communities. That can be pretty straightforward, but needs locating the interdisciplinary walls and reranking interesting questions in each cell as if that is all you could see.
110
Reposted by Max Puelma Touzel
mensrea @mensie.bsky.social · 19/07/2026
German broadcast. Nailed it
139149784993
Max Puelma Touzel @mptouzel.bsky.social · 18/07/2026
Hah. Funny to see my ML feed deep in the higher scaling regime full of talk of the existential crises from too many papers, and then this from my social science feed who is only at the base of the knee who haven't gotten the memo (yet) about the tidal wave that's coming.
000
Max Puelma Touzel @mptouzel.bsky.social · 14/07/2026
I don't know anything about this particular application yet, but first impression is laden with "positive future" ideas. If nothing else, that's what this #atproto offers society: an invitation to envision a better future.
080
Max Puelma Touzel @mptouzel.bsky.social · 10/07/2026
In the midst of all the discussion about how hard it is to know good science from bad, the clearest signal I know for good science is listening to the presenter and hearing all the commentary about the process and context around the research. More evidence for me that #trails is the way. #ICML2026
050
Max Puelma Touzel @mptouzel.bsky.social · 08/07/2026
"just to polish" responds to social norms rather than being an actual description: it defends against being viewed as delegating intellectual work, but anyone can see that these tools inject more than just style and grammar.
100
Max Puelma Touzel @mptouzel.bsky.social · 06/07/2026
The trails guy! Always nice to see an ATproto person who is in the ML world as well. Not a picture per se @sharky6000.bsky.social , but the kind of #icml2026 content you are looking for 😁
030
Max Puelma Touzel @mptouzel.bsky.social · 03/07/2026
intuition pump up from pitting intuitions against eachother
010
Max Puelma Touzel @mptouzel.bsky.social · 03/07/2026
Excited for #ICML2026 in Seoul! I'll be co-presenting 2 positions on multi-🤖 behaviour: (1) The LLM social sim research community must focus on reproducible evals (2) the risk of undetectable collusion among pricing agents is real and demands pre-deployment certification. icml.cc/virtual/2026...
icml.cc
132
Reposted by Max Puelma Touzel
ICML Conference @icmlconf.bsky.social · 02/07/2026
As you're planning your schedule for #ICML2026, be sure to pencil in the Socials! We have a great lineup of 12 Socials, on topics ranging from "how to network" to "how to negotiate a job offer" to chess, AI for games, AI for science, & more! Check out in the blog post: blog.icml.cc/2026/07/02/s...
052