Manfred Diaz

I was following this one during the COVID pandemic, but it has been inactive for quite some time. The original talks' recordings are amazing, though!

Yeah, it's been a period for all of us simultaneously! I have also been pretty busy with thesis/job search. Hopefully, it will be back running in the Fall term!

@aamasconf.bsky.social 2025 was very special for us! We had the opportunity. to present a tutorial on general evaluation of AI agents, and we got a best paper award! Congrats, @sharky6000.bsky.social and the team! 🎉

A Tutorial on General Evaluation of AI Agents Artificial Intelligence (AI) and machine learning (ML), in particular, have emerged as scientific disciplines concerned with understanding and building single and multi-agent systems with the ability ...

In the afternoon we will be giving a tutorial on general evaluation of AI agents.

sites.google.com/view/aamas20... 10/N

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...

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.

Congrats, Seth!

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.

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...

The quality of London's museums is just amazing! Enjoy!

YouTube video by MITCBMM A Theory of Appropriateness with Applications to Generative Artificial Intelligence

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

Our new evaluation method, Soft Condorcet Optimization is now available open-source! 👍

Both the sigmoid (smooth Kendall-tau) and Fenchel-Young (perturbed optimizers) versions.

Also, an optimized C++ implementation that is ~40X faster than the Python one. 🤩⚡

github.com/google-deepm...

Working at the intersection of social choice and learning algorithms?

Check out the 2nd Workshop on Social Choice and Learning Algorithms (SCaLA) at @ijcai.bsky.social this summer.

Submission deadline: May 9th.

I attended last year at AAMAS and loved it! 👍

sites.google.com/corp/view/sc...

If the AAMAS website is a good reference for this, it may not be, but uncertain atm.

Come to understand ML evaluation from first principles! We have put together a great AAMAS tutorial covering statistics, probabilistic models, game theory, and social choice theory.

Bonus: a unifying perspective of the problem leveraging decision-theoretic principles!

Join us on May 19th!

Re #2: The key finding there is that the stationary points of SCO contain the margin matrix but, as I said in the note, there is still more work to do!

Thanks! I have been meaning to update the manuscript to standalone without the main paper but instead I may have change the content to a different format 😉. Coming soon!

Ah, I see the confusion... I never used the "identically distributed assumption," only the independence assumption (from 8 to 9).

I'm not sure if I understood your question correctly, but yes? As the post you shared says, "Voila! We have shown that minimizing the KL divergence amounts to finding the maximum likelihood estimate of θ." Maybe I am missing your point 😬

Elo drives most LLM evaluations, but we often overlook its assumptions, benefits, and limitations. While working on SCO, we wanted to understand the SCO-Elo distinction, so I looked and uncovered some intriguing findings and documented them in these notes. I hope you find them valuable!

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

I had the convexity results for the online pairwise update (Section B.1.1.1) in my notes (manfreddiaz.github.io/assets/pdf/s...), but it is not clear to me if they hold for the other non-online settings. Worth taking a more detailed pass over the paper!

That's a nice finding, @sacha2.bsky.social! @sharky6000.bsky.social I skimmed over it, and it seems neat! There is an important distinction, though. They work with the "online" Elo regime, departing from the traditional gradient/batch gradient descent updates. (e.g., FIDE doesn't use online updates)

lol 😀

Not that Michael Jordan, but this one en.wikipedia.org/wiki/Michael...

I believe this example conveys, as Prof. Jordan hinted, the need for fresh conceptual frameworks that shift our perspective, help us avoid conceptual confusion, and increase our ability to build the future of AI. I believe ML-SoA provides such framework, but I’d love to hear more perspectives!

YouTube video by Jack Clark CoastRunners 7

Through the ML-SoA design-centric view, reward hacking, misspecification and similar issues originate from designers' bounded rationality. The surprising “alignment problems” (www.youtube.com/watch?v=tlOI...) reflect ML designers' inability to comprehend the consequences of the programs they write.

For instance, in the AI safety literature, we hear phrases such as “the agent hacked the reward”, “the robot gamed the reward”, and similar statements. But does an ML model really "hack" or "game" anything? Or is this just a misleading metaphor? Chances are, it’s the latter.

None of these benefits is more critical than removing, to a greater extent, the anthropomorphism that has existed in AI, currently exacerbated by the rise of LLMs, whose principal peril is the conceptual confusion it sometimes causes. Let me clarify what I mean by "conceptual confusion".

Milnor-Myerson Games and The Principles of Artificial... In this paper, we introduce Milnor-Myerson games, a multiplayer interaction structure at the core of machine learning (ML), to shed light on the fundamental principles and implications the...

We recently argued here (openreview.net/forum?id=LNY...) that examining ML through this perspective, which I call Machine Learning Through the Sciences of the Artificial (ML-SoA), offers solid foundations for understanding AI and ML in a broader context and introduces multiple benefits.

Back in the 1960, Herbert Simon defined in "The Sciences of the Artificial" an artificial entity as a product of human creation. In this context, AI shares foundations with other scientific theories that study human-designed structures such as institutions, markets, economies, and elections.

AI, Science and Society Conference - AI ACTION SUMMIT - DAY 1 YouTube video by IP Paris

Last week, Michael I. Jordan's insightful talk at the AI Action Summit (www.youtube.com/live/W0QLq4q...) reminded us of the meaningful connections between AI, ML, economics, game theory, and mechanism design. But I'd argue the relationship goes deeper—it's profound, historical, and foundational. ⬇️