Arash Behboodi (@behboodiarash) header
Arash Behboodi (@behboodiarash) profile picture

Director @Qualcomm AI Research. Information theory & AI, and substrates of computation. Exploring the mathematical and philosophical foundations of intelligence

Joined October 2023
Photo by Arash Behboodi (@behboodiarash)Photo by Arash Behboodi (@behboodiarash)Photo by Arash Behboodi (@behboodiarash)
Arash Behboodi retweeted
congrats to @berggruenInst for awarding the 2026 prize for philosophy and culture to tom nagel. berggruen.org/eu/news/usd1-m…
19
66
14
438
25,872
Arash Behboodi retweeted
Been running Opus 5.5 agents over the notebooks of Gauss, Riemann, Hilbert and Grothendieck. around 12,000 pages so far. Heres the best finding so far: Riemann's 1860 shock wave paper is famous for getting the energy part wrong however his unpublished sheets at Göttingen have it right the Rankine Hugoniot conditions a decade or two before Rankine and Hugoniot plus results usually dated 19061937, all in his own notation. A second sheet in another folder has the same physics. There are probably thousands of potential undiscovered stuff in the research papers of the smartest ever to live imo all of them should be made public and reviewed by ai years of work can probably be transcribed by agents in days
77
354
65
3,621
268,806
Arash Behboodi retweeted
I am saddened to say that Karl Astrom, a towering figue in control theory, and a pioneer of adaptive control, passed away yesterday, Nov 12, at the age of 92. I imagine that in this AI-crazed age, with its constant pursuit of "flavor-of-the-month", many may not be aware of his contributions. However, one can draw a clear arc from his work to many of the ideas currently prevalent in machine learning. As early as 1965 he wrote (what I assume is) the first paper on the control of partially-observed Markov decision processes, which play an important role in modern reinforcement learning: Optimal control of Markov decision processes with incomplete state estimation, J. Math. Anal. Applic., 1965 His landmark paper with Wittenmark lays the foundations for "learning and control": On self tuning regulators, Automatica, 1973 I have been personally indebted to him in multiple ways. When I was a second year graduate student I attended the 1993 Institute of Mathematics and its Applications (IMA) Workshop on Adaptive Control in Minnesota that he had been organizing with Graham Goodwin and PR Kumar. After a brief discussion of some of my research he go so excited that, midway through the workshop, he added me to the program. I got pummeled by the faculty after my talk (even though I was right and they were wrong 😀). I last saw him two years ago at Lund, where he spent almost all of his career. He was still active, visited the department, and attended talks regularly, always asking pointed questions (though he had to sit in a chair next to the speaker in order to hear!). Altogether a remarkably kind and generous human being. A true scholar. RIP Karl.
22
1
101
4,421
Arash Behboodi retweeted
Replying to @ziv_ravid
@ziv_ravid These suggestions are good ones, in my opinion, and should be considered regardless of how many submissions a venue receives. +1 for 3-4 pages consisting only of an abstract and the core idea. Sadly, I doubt they will change the situation; people adapt to them.
1
1
1,132
An important new paper from the lab : in a few years, vision models have progressed to such an extent that they almost match human perception in the domain of geometry (but may still fall short of having a genuine language of shapes). bsky.app/profile/maxencepajo… @Maxence_Pajot
4
15
95
7,099
Arash Behboodi retweeted
«ژینا گیان تۆ نامری، ناوت دەبێته‌ ڕەمز» #زن_زندگی_آزادی
12
90
9
568
11,458
Arash Behboodi retweeted
Just heard that James Munkres passed away. His undergrad textbook "Topology: A First Course" is a classic. I appreciated his kindness when I was an assistant professor at @MIT. douglassfh.com/obituary/jame…
14
87
15
641
36,636
At the end, these are rumors with their own caveats and asterisks; however, reusing compute from the units of the same model as in loop models is a promising direction, removing the need for latent supervision as other latent reasoning models.
130
Today’s AI needs to be able to deal with uncertainty! This has been a central theme of my work, so I was delighted to talk about it at length with @FryRsquared for the latest @GoogleDeepMind Podcast. Intelligence means knowing the limits of one's own knowledge, and when to say you are unsure about something. If we want useful AI systems that we can trust to make important decisions, then they need to be able to do the same. They should recognise blind spots, avoid overconfidence, and reason carefully about their decisions. I think you’ll enjoy the episode, but I can’t say for sure. 🙂 You can watch it here: goo.gle/4y6WNlR
7
17
5
131
27,145
🚨1 week away!🚨 The submission deadline for NeurReps 2026 is 1 week away (Monday, August 24th) Submit your work: neurreps.org/#cfp
5
2
12
2,774
Arash Behboodi retweeted
Doubling down on Milton Friedman (not defending the whole Encyclopedia piece). The comments to my tweet (along the lines of "there is so much more to Friedman than the caricature in the piece") raise a larger issue, that of “the role of the economist in the city,” to use Jean Tirole’s expression. Friedman was a remarkable and sophisticated thinker. But when I talk about “TV Friedman,” I do not talk about TV simplifying a complex Friedman message, but about Friedman himself expressing his views on TV (in “Free to choose” in particular). And the message was unambiguous: Markets work, the state typically stands in the way. No ifs, no buts. The influence, if any, we (academics) have on perceptions and on politics does not derive from our publications in Econometrica or the American Economic Review, but on how we sell their conclusions to policy makers and to the public. And, in the case of Friedman, the message was simple and unambiguous. This is relevant to assessing his contribution.
33
69
19
613
206,745
Arash Behboodi retweeted
Text-to-image models have become remarkably good at producing realistic images. But realism isn’t the same as correctness. Ask for several distinct people, a specific composition, or a high-resolution image generated locally, and today’s models still struggle in surprising ways. In this episode, Fatih Porikli, Vice President of Technology at @QCOMResearch, joins me to discuss what remains unsolved in image generation and several approaches his team presented at CVPR to address those challenges. We explore why better training objectives can improve controllability, how separating scene planning from rendering may lead to more reliable image generation, techniques for generating 16-megapixel images efficiently on edge devices, and new methods for eliminating the visible artifacts that often appear in AI-powered image editing. Along the way, we discuss reinforcement learning for image generation, agentic image generation pipelines, on-device AI, and what the next phase of progress in generative vision systems is likely to look like. 🗒️ Full show notes: twimlai.com/go/773. 📖 CHAPTERS =============================== 00:00 - Introduction 02:04 - What’s Still Unsolved in Image Generation 05:18 - Disco: Improving Identity Diversity 09:09 - Better Training Objectives and Reinforcement Learning 13:09 - Are We Asking One Model to Do Too Much? 16:06 - Agentic Image Generation Pipelines 18:42 - Evaluating Diversity with the DiverseHumans Benchmark 22:14 - Ar2Can: Separating Scene Planning from Rendering 33:07 - PixelRush: Efficient High-Resolution Image Generation 42:50 - PixelRush: Generating 16-Megapixel Images with Limited Memory 48:30 - InvertFill: Artifact-Free Image Editing 53:31 - PyramidalWan, ReHyAtt, and Attention Surgery for Efficient Video Generation 55:18 - Closing Remarks
1
2
3
593
Highly recommended. Check this out for our works on AI inference on Edge devices.
Realistic images are easy. Correct images are hard. In the latest @TWIMLAI podcast, @Qualcomm AI Research VP of Technology Fatih Porikli discusses the next frontier of image generation: controllability, agentic AI pipelines, high-resolution on-device generation, and eliminating editing artifacts. Listen now: bit.ly/4zgFzUc #AI #GenerativeAI #ComputerVision #OnDeviceAI #CVPR
1
129
Arash Behboodi retweeted
We asked an unreleased research version of Claude to take a stab at the Riemann hypothesis. It didn’t solve it, but it did make strides on a related problem: it increased the lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis from 41.6% to 67.2%. anthropic.com/research/riema…
923
1,984
1,339
18,192
10,346,672
Arash Behboodi retweeted
Fascinating read. Fascinating result, Dimitris. I still remember the day you and Alex came to my office to describe the work you had been doing on this problem. Not sure I would characterize you as an "anxious" first-year grad student 😀. And, yes, this result would have led to a lot of acclaim back then. But I would not characterize it as a problem no one cares about anymore. It is a fundamental result. Especially, since the algorithm is so simple. I can think of many possible generalizations. Does the result hold in the non-square case, or for higher-order constellations? Most importantly, perhaps, is it possible to match the ML solution with a polynomial-time algorithm at SNRs below 2logN? What is the threshold? And I remain interested in the MCMC angle. I had hoped that the "spectral independence" technique of Anari, Liu, and Oveis-Gharan, would crack this, but I've not been able to... Anyway, kudos for your work. AI has changed the face of research forever...
4
10
2
95
21,442
Took 5+ days to simplify and verify, extremely exhausting, and no i don't trust lean to do it for me.
The proof is long, but basic, here is a roadmap to it
4
7
3
148
17,486
Also very nice analysis from @KrzakalaF . AMP is a fascinating method. The first eye-opening moment for me was the AMP optimality results for compressed sensing (see refs. in the post) in connection with the results of Wu and Verdu on analog lossless compression.
Yesterday, an interesting new exact-recovery theorem, and a remarkable AI-assisted proof appeared here: xitter.gq/DimitrisPapail/status/… While the result was interesting, the AMP algorithm, rigorous state evolution/replica formula, and the 2 log N prediction have a substantial prior history. I tried to separate what was known, what is new, and what AI may really have changed :
3
221
Arash Behboodi retweeted
Congratulations on solving a beautiful problem. This is bringing back old memories ; I still remember first discussing this problem with Sergio Verdu when he was on sabbatical at Berkeley in the late 1990's . With a wink of an eye it's now 25 years later. Some perspective on this problem from someone who has been around. This problem belongs to the class of problems called "multiuser or MIMO detection", pioneered by Sergio Verdu and received a lot of attention in the late 80's and early 90's, In the late 90's and early 2000's, there was a resurgence of interest in this problem via connections with random matrix theory. I myself have worked on a bunch of these problems, evaluating performance of different MIMO detectors. The problem that AI and @DimitrisPapai have solved is the ultimate of these problems because the optimal MIMO detector is the joint ML detector. I am very happy to see it solved. That being said, I have to say that MIMO detection as a whole has made little impact on how actual wireless communication systems are built. Because wireless communication systems, like all communication systems, use coding, while MIMO detection only talks about uncoded systems; the N information bits here in this problem are uncoded. It turns out that much lower SNR can be achieved with coding. Take this problem as an example, with coding, a constant SNR* = 1.32 is sufficient to support an information rate of 1 bit per antenna, in contrast to the growing SNR = 2 log N result just proved. That's one reason why although this problem is beautiful as a mathematical problem, it has not received the full attention of the information theory community even in those days of peak activity. Claude Shannon has already told us in his 1948 information theory paper: to achieve the ultimate limit of communication, one needs to do coding. Uncoded systems are sub-optimal. So SNR = 2 log N is only optimal among uncoded communication systems, but SNR* = 1.32 is optimal among all communication systems. The problem that Claude Shannon solved in his 1948 paper was the capacity of point to point communication, The problem here is one example (so called point to point MIMO channel), and hence the optimal SNR* = 1.32 can already be computed by Shannon's 1948 theory (no AI needed). Most of the information theorists' attention in the hey days of wireless communications research has been to extend Shannon's theory to network information theory problems, such as the broadcast channel, the relay channel and the interference channels. In those problems , one ask what is the best information rate (or equivalently, the minimum SNR needed) that a network of users can communicate with each other. These problems have been proposed since the 60's and the 70's and to this day none have been solved in their full generality. I'd love to see AI take a crack at some of those open problems.
7
14
2
151
54,960
Network information theory has been on my mind for a while, after AI breakthroughs in Math. I am quite curious to see if problems like broadcast channels can be solved by techniques to close the gap between the Marton inner bound and the upper bounds /1
Congratulations on solving a beautiful problem. This is bringing back old memories ; I still remember first discussing this problem with Sergio Verdu when he was on sabbatical at Berkeley in the late 1990's . With a wink of an eye it's now 25 years later. Some perspective on this problem from someone who has been around. This problem belongs to the class of problems called "multiuser or MIMO detection", pioneered by Sergio Verdu and received a lot of attention in the late 80's and early 90's, In the late 90's and early 2000's, there was a resurgence of interest in this problem via connections with random matrix theory. I myself have worked on a bunch of these problems, evaluating performance of different MIMO detectors. The problem that AI and @DimitrisPapai have solved is the ultimate of these problems because the optimal MIMO detector is the joint ML detector. I am very happy to see it solved. That being said, I have to say that MIMO detection as a whole has made little impact on how actual wireless communication systems are built. Because wireless communication systems, like all communication systems, use coding, while MIMO detection only talks about uncoded systems; the N information bits here in this problem are uncoded. It turns out that much lower SNR can be achieved with coding. Take this problem as an example, with coding, a constant SNR* = 1.32 is sufficient to support an information rate of 1 bit per antenna, in contrast to the growing SNR = 2 log N result just proved. That's one reason why although this problem is beautiful as a mathematical problem, it has not received the full attention of the information theory community even in those days of peak activity. Claude Shannon has already told us in his 1948 information theory paper: to achieve the ultimate limit of communication, one needs to do coding. Uncoded systems are sub-optimal. So SNR = 2 log N is only optimal among uncoded communication systems, but SNR* = 1.32 is optimal among all communication systems. The problem that Claude Shannon solved in his 1948 paper was the capacity of point to point communication, The problem here is one example (so called point to point MIMO channel), and hence the optimal SNR* = 1.32 can already be computed by Shannon's 1948 theory (no AI needed). Most of the information theorists' attention in the hey days of wireless communications research has been to extend Shannon's theory to network information theory problems, such as the broadcast channel, the relay channel and the interference channels. In those problems , one ask what is the best information rate (or equivalently, the minimum SNR needed) that a network of users can communicate with each other. These problems have been proposed since the 60's and the 70's and to this day none have been solved in their full generality. I'd love to see AI take a crack at some of those open problems.
1
8
1,538
Along the lines of the work of Gohari, Chandra Nair, etc. (see chandra.ie.cuhk.edu.hk/pub/p…), do we need to go beyond the classical methods we have: Marton coding, superposition codes, cut-set upper bounds, etc. Same for relay, interference, two-way channels.
1
73