Lightnews — Scholar-powered news

Joel Z Leibo @jzleibo.bsky.social · 12d

Concordia was always an entity-compoment pattern. But it was improved in 2.0. Also, we realized that it was an important part of the story to emphasize and explain to anyone who didn't already know about it. So that's what we did in the 2.0 tech report.

Here:
arxiv.org/abs/2507.08892

Multi-Actor Generative Artificial Intelligence as a Game Engine

Generative AI can be used in multi-actor environments with purposes ranging from social science modeling to interactive narrative and AI evaluation. Supporting this diversity of use cases -- which we ...

arxiv.org

1 1

Joel Z Leibo @jzleibo.bsky.social · 19d

Tutorial here: www.youtube.com/watch?v=2FO5...

Concordia Library v2.0 Release - Tutorial with Alexander (Sasha) Vezhnevets [Google DeepMind]

YouTube video by Cooperative AI Foundation

www.youtube.com

2

Joel Z Leibo @jzleibo.bsky.social · 19d

Concordia 2.0: www.cooperativeai.com/post/google-...

Google DeepMind Releases Concordia Library v2.0

Google DeepMind has released Concordia v2.0, a platform providing a robust environment for creating multi-agent simulations, games, and AI evaluations.

www.cooperativeai.com

3 3 9

Joel Z Leibo @jzleibo.bsky.social · Jul 22

One can also explain human behavior this way, of course 😉, we are all role playing.

But I certainly agree with the advice you offer in this thread. Humans harm themselves when they stop playing roles that connect to other humans.

Murray Shanahan @mpshanahan.bsky.social · Jul 22

However, a good explanation for much of the sophisticated behaviour we see in today's chatbots is that they are role-playing. 3/4

4

Joel Z Leibo @jzleibo.bsky.social · Jul 12

Reposting our paper from a few months ago as recent events again underscore its relevance:

"a theory of appropriateness with applications to generative AI"
arxiv.org/abs/2412.19010

A theory of appropriateness with applications to generative artificial intelligence

What is appropriateness? Humans navigate a multi-scale mosaic of interlocking notions of what is appropriate for different situations. We act one way with our friends, another with our family, and yet...

arxiv.org

2 8

Joel Z Leibo @jzleibo.bsky.social · Jun 30

Congratulations!!!! 🎉

1

Reposted by Joel Z Leibo

Wolfram Barfuss @wbarfuss.bsky.social · Jun 17

How to cooperate for a sustainable future? We don't know (yet), but I'm thrilled to share that our new perspective piece has just been published in @pnas.org. Bridging complexity science and multiagent reinforcement learning can lead to a much-needed science of collective, cooperative intelligence.

1 6 11

Reposted by Joel Z Leibo

Ted Underwood @tedunderwood.com · Jun 12

Love the insight that language models have been more widely useful than video models.

For Levine, this is because LLMs kind of sneakily and indirectly model the brain. I would be tempted to say instead that shared languages are more powerful than single brains. But either way, it’s a good read!

Language Models in Plato's Cave

Why language models succeeded where video models failed, and what that teaches us about AI

open.substack.com

5 5 39

Reposted by Joel Z Leibo

Eliot Higgins @eliothiggins.bsky.social · Jun 6

Bellingcat has followed up its 2023 article on testing LLMs on geolocation with a new review of LLM's geolocation skills, with dramatically improved results www.bellingcat.com/resources/ho...

Have LLMs Finally Mastered Geolocation? - bellingcat

We tasked LLMs from OpenAI, Google, Anthropic, Mistral and xAI to geolocate our unpublished holiday snaps. Here's how they did.

www.bellingcat.com

5 55 200

Joel Z Leibo @jzleibo.bsky.social · Jun 3

Confabulate is exactly the right word for it -- used exactly this way in neuroscience for decades. I see some people also pointing out that Frankfurt's "bullshit" also works, and I agree with that too. Confabulate seems better for obvious reasons though

1

Reposted by Joel Z Leibo

David Pfau @davidpfau.com · May 9

The idea of "AI alignment" grew out of a community that thought you could solve morals like it was a CS problem set. Nice to see a more nuanced take.

Joel Z Leibo @jzleibo.bsky.social · May 9

Announcing our latest arxiv paper:

Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt
arxiv.org/abs/2505.05197

We argue for a view of AI safety centered on preventing disagreement from spiraling into conflict.

Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt

Artificial Intelligence (AI) systems are increasingly placed in positions where their decisions have real consequences, e.g., moderating online spaces, conducting research, and advising on policy. Ens...

arxiv.org

3 25

Joel Z Leibo @jzleibo.bsky.social · May 9

Thanks to my coauthors @sebk.bsky.social, @sindero.bsky.social, Sasha Vezhnevets, Wil Cunningham, and @manfreddiaz.bsky.social

6

Joel Z Leibo @jzleibo.bsky.social · May 9

It's a revised version of the post with the same name we previously made on LessWrong. This version adds references and attempts to incorporate responses to most of the comments we had on the original.

1 4

Joel Z Leibo @jzleibo.bsky.social · May 9

Announcing our latest arxiv paper:

Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt
arxiv.org/abs/2505.05197

We argue for a view of AI safety centered on preventing disagreement from spiraling into conflict.

Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt

Artificial Intelligence (AI) systems are increasingly placed in positions where their decisions have real consequences, e.g., moderating online spaces, conducting research, and advising on policy. Ens...

arxiv.org

1 6 23

Joel Z Leibo @jzleibo.bsky.social · Apr 23

Just posted today, here's a podcast interview where I spoke about multi-agent artificial general intelligence.

www.aipolicyperspectives.com/p/robinson-c...

Robinson Crusoe or Lord of the Flies?

The importance of Multi-Agent AI, with Joel Z. Leibo

www.aipolicyperspectives.com

7

Joel Z Leibo @jzleibo.bsky.social · Apr 22

The post also serves as an introduction to some of the ideas in our broader "Theory of Appropriateness" which we published on arxiv a few months ago
arxiv.org/abs/2412.19010

A theory of appropriateness with applications to generative artificial intelligence

What is appropriateness? Humans navigate a multi-scale mosaic of interlocking notions of what is appropriate for different situations. We act one way with our friends, another with our family, and yet...

arxiv.org

3

Joel Z Leibo @jzleibo.bsky.social · Apr 22

First LessWrong post! Inspired by Richard Rorty, we argue for a different view of AI alignment, where the goal is "more like sewing together a very large, elaborate, polychrome quilt", than it is "like getting a clearer vision of something true and deep"
www.lesswrong.com/posts/S8KYwt...

Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt — LessWrong

We can just drop the axiom of rational convergence.

www.lesswrong.com

3 1 6

Joel Z Leibo @jzleibo.bsky.social · Apr 11

That's nothing, it also improves performance if you tell the worker what they ate for breakfast!

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · Apr 10

The multiagent LLM people are going to do me in. What do you mean you told one agent it was a manager and the other a worker and it slightly improved performance

3

Joel Z Leibo @jzleibo.bsky.social · Apr 1

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

A Theory of Appropriateness with Applications to Generative Artificial Intelligence

YouTube video by MITCBMM

youtu.be

3 8

Reposted by Joel Z Leibo

Nathan Lambert @natolambert.bsky.social · Mar 26

A silly example of this being unexpected for AI forecasters

4 3 23

Reposted by Joel Z Leibo

Marc Lanctot @sharky6000.bsky.social · Feb 24

Looking for a principled evaluation method for ranking of *general* agents or models, i.e. that get evaluated across a myriad of different tasks?

I’m delighted to tell you about our new paper, Soft Condorcet Optimization (SCO) for Ranking of General Agents, to be presented at AAMAS 2025! 🧵 1/N

1 17 67

Joel Z Leibo @jzleibo.bsky.social · Feb 21

CAIF's new and massive report on multi-agent AI risks will be really useful resource for the field
www.cooperativeai.com/post/new-rep...

Cooperative AIPlaintext Code Block

www.cooperativeai.com

1 3

Reposted by Joel Z Leibo

Jeff Dean @jeffdean.bsky.social · Feb 15

I can't believe they've just cancelled the Epidemic Intelligence Service program at CDC.

This program trains the best & brightest epidemiologists, who then go on to have distinguished careers in public health, serving at CDC, in state health departments, overseas, ...

27 110 440

Reposted by Joel Z Leibo

turrigiano.bsky.social @turrigiano.bsky.social · Feb 14

Happy to see SfN sign on to this

Dr. Asmeret Asefaw Berhe @aaberhe.com · Feb 13

Professional societies stand together in defense of science & scientists.
“science has led to humanity’s greatest advances, improving people’s lives & the health of our planet … (we’re) committed to supporting, elevating, & fighting for science & those who further it.“
www.unitedsciencealliance.org

1 9

Joel Z Leibo @jzleibo.bsky.social · Feb 12

Thanks!

1