iiinova — where your digital self keeps evolving

Aran Nayebi @anayebi.bsky.social · 05/10/2026

About time! We would have a running bet in grad school when optogenetics would finally win the Nobel. I guess today is finally the day 👏

082

Aran Nayebi @anayebi.bsky.social · 30/09/2026

A few months *before* the now famous HuggingFace incident, we already showed frontier agents override human control, resist shutdown, and violate explicit resource restrictions when evaluated in agentic, computer-use settings — even when they seem aligned in text-only evaluations!

120

Aran Nayebi @anayebi.bsky.social · 30/09/2026

At this rate, ML conferences are just based on whether an LLM thinks a paper should be accepted/rejected. What's the point then? I, too, can run my paper through Astra, and ask it to review it. There's 0 added benefit waiting months just to have some other reviewer do the same.

0140

Aran Nayebi @anayebi.bsky.social · 27/09/2026

The “AI is fake” crowd watching a plane take off:

170

Reposted by Aran Nayebi

Yang Tan Collective @yangtancollective.bsky.social · 21/09/2026

Former Yang ICoN Fellow and current assistant professor in the CMU School of Computer Science Aran Nayebi @anayebi.bsky.social talks about neuroscience + AI in Does Compute, a podcast presented by Carnegie Mellon University School of Computer Science and GeekWire Studios. Watch ⬇️

youtube.com

Very Much Inspired By the Nervous System

YouTube video by GeekWire

011

Aran Nayebi @anayebi.bsky.social · 21/09/2026

Check out the CMU SCS podcast episode highlighting our lab's work on reverse-engineering natural intelligence with autonomous agents: www.youtube.com/watch?v=-U7S...

youtube.com

Very Much Inspired By the Nervous System

YouTube video by GeekWire

051

Aran Nayebi @anayebi.bsky.social · 19/09/2026

Happy 44th birthday today to the :-) emoticon, first emoted at 11:44 am ET on Sept. 19, 1982 by CMU Professor Scott Fahlman. I got to meet him earlier this week, where he graciously signed our :-) shirts. “What do you work on? LLMs?” I said neuroscience. “So Real I. Faculty keep getting younger.” 😂

030

Aran Nayebi @anayebi.bsky.social · 14/09/2026

For my personal take on "Why the NeuroAI Approach is *Inevitable*", check out my introduction to the debate here: www.youtube.com/watch?v=7BUo... My non-mangled slides: anayebi.github.io/files/slides...

youtube.com

CCN 2026 | GAC Introduction

YouTube video by Cognitive Computational Neuroscience

130

Aran Nayebi @anayebi.bsky.social · 14/09/2026

If you're interested in the discussions around the merits of NeuroAI for understanding the mind & brain, check out our #CCN2026 GAC Debate recording! www.youtube.com/watch?v=kYES...

youtube.com

CCN 2026 | GAC: NeuroAI Methods & Frameworks

YouTube video by Cognitive Computational Neuroscience

1173

Aran Nayebi @anayebi.bsky.social · 11/09/2026

CMU is hiring an Assistant Professor in NeuroAI! Come join our awesome NeuroAI community 🧠🤖: apply.interfolio.com/189268

apply.interfolio.com

Apply - Interfolio {{$ctrl.$state.data.pageTitle}} - Apply - Interfolio

02811

Aran Nayebi @anayebi.bsky.social · 17/08/2026

If you are attending @UncertaintyInAI #UAI2026 in Amsterdam, I will be presenting this work on "What Capable Agents Must Know" tomorrow (Tuesday) in the poster session as Poster #205! Poster below 👇 if you can't make it: anayebi.github.io/files/poster...

071

Aran Nayebi @anayebi.bsky.social · 06/08/2026

Thanks so much for the super helpful pointers, Alex! We will cite these in the upcoming v3 :)

020

Aran Nayebi @anayebi.bsky.social · 04/08/2026

Finally, we study the effects of noise and subsampling on metrics, and they alter this task-irrelevant symmetry.

110

Aran Nayebi @anayebi.bsky.social · 04/08/2026

For RSA: it turns out that you can decompose RSMs into unique task-relevant "core geometry" and a task-irrelevant symmetry-generated term. Weak-strong equivalence holds for the core geometry and by projecting onto privileged axes you can filter out the task-irrelevant part so it doesn't interfere.

110

Aran Nayebi @anayebi.bsky.social · 04/08/2026

For transformers: it turns out that they have privileged axes, just like convnets, if you look in the right place (MLP layers and attention heads.) The identification of privileged heads is a potentially key result for emergence of interpretable stucture in LLMs.

110

Aran Nayebi @anayebi.bsky.social · 04/08/2026

Version 2 of Theory of Contravariance w/ @dyamins.bsky.social is out! New material on contravariance for Transformers, and the theory of Representational Similarity Analysis (RSA) and centered kernel analysis (CKA)/Procrustes.

3155

Aran Nayebi @anayebi.bsky.social · 04/08/2026

000

Aran Nayebi @anayebi.bsky.social · 04/08/2026

100

Aran Nayebi @anayebi.bsky.social · 31/07/2026

We expand the original UBI analysis in v1 of our paper to characterize how public capture, deployment costs, automation scope & market structure shift the AI capability threshold needed for broad-based benefit. Camera ready here: arxiv.org/pdf/2505.18687

arxiv.org

030

Aran Nayebi @anayebi.bsky.social · 31/07/2026

Now accepted to the AI, Ethics, and Society 2026 conf under the new title: "When Do AI Gains Become Broadly Shareable?" With all the rapid progress in AI, it is critical to now rigorously characterize the institutional conditions that allow gains from greater AI capability to be broadly shared.

141

Aran Nayebi @anayebi.bsky.social · 23/07/2026

If anyone ever asks you: "What has AI/NeuroAI taught us about the brain?" Here's a (woefully incomplete!) list of what you can say:

0110

Aran Nayebi @anayebi.bsky.social · 23/07/2026

Why, according to Terry Tao, the recent disproof of the Jacobian conjecture in N > 2 dimensions was not mere brute force and required what many mathematicians would characterize as "creative insight" had a human come up with it!

2114

Aran Nayebi @anayebi.bsky.social · 22/07/2026

The blogpost exposition on our "zippering theorems" in Section 4 from our recent Contravariance Theory paper is now out! These zippering theorems explain why end-to-end optimization for downstream tasks often seems to strongly constrain upstream representations.

141

Aran Nayebi @anayebi.bsky.social · 13/07/2026

7/6 As an aside, it's worth noting that this work significantly sharpens my recent selection theorems, for one showing that in modern neural networks, the complexity of the mapping between two networks that are representationally aligned is *at most* linear: bsky.app/profile/anay...

020

Aran Nayebi @anayebi.bsky.social · 13/07/2026

4/6 The Zippering Theorems (ZIP) give a more complete picture. Zippering is when convergence downstream "zippers" upstream, so end-to-end optimization => axis alignment even in early or middle layers. ZIP is the core explanation of convergent evolution from contravariance.

100

Aran Nayebi @anayebi.bsky.social · 13/07/2026

3/6 Weak-strong equivalence (WSE) says that, for hard tasks, alignment up to linear representation (weak alignment) mathematically implies alignment up to privileged axes (strong alignment). And it explains *why* we see privileged axes in the first place -- namely, hard tasks.

110

Aran Nayebi @anayebi.bsky.social · 13/07/2026

2/6 Specifically, we've created a mathematical Theory of Contravariance in NeuroAI: the informal idea that strong tasks are constraining on the solution space. There are two high-impact concepts: weak-strong equivalence (WSE) and Zippering (ZIP) that shape how to think about NeuroAI going forward.

110

Aran Nayebi @anayebi.bsky.social · 13/07/2026

1/6 Why have deep neural networks aligned so strongly with brains for the past 15 years? What explains it? @dyamins.bsky.social & I make progress on this question in our new paper👇 In a nutshell, we *prove* that for sufficiently hard tasks, the choice of alignment metric does *not* matter.

2317

Aran Nayebi @anayebi.bsky.social · 09/07/2026

There was no recording, but please check out @reecedkeller.bsky.social's talk on our work at our #Cosyne2026 "NeuroAgents" workshop: www.youtube.com/watch?v=LZiR... Slides here: anayebi.github.io/files/slides...

youtube.com

Reece D. Keller (CMU)

YouTube video by NeuroAgents Workshop

041

Aran Nayebi @anayebi.bsky.social · 09/07/2026

Had a blast presenting our virtual zebrafish paper at #FENS2026 in Barcelona in the "Computational Astroscience" symposium -- the first autonomous agent that can predict *whole-brain* neural-glial data!

110

Reposted by Aran Nayebi

Audrey Denizot @adenizot.bsky.social · 06/07/2026

We're at @fens.org! #FENS2026 #FENSGlia Posters: PS02-07PM-558, PS03-08AM-495, PS05-09AM-674. Delighted to be presenting at the S44 Computational astroscience symposium on Thursday with T. Fellin, @anayebi.bsky.social & T. Papouin. Thanks @inbalgoshen.bsky.social for organizing, looking forward!

051

Aran Nayebi @anayebi.bsky.social · 04/07/2026

3. Our recent ROGUE benchmark: our empirical test of how corrigible today’s frontier AI agents really are. Models that behave well in ordinary chat don’t always stay that way once they’re given a real off-switch they could disable to finish a task. Paper: arxiv.org/abs/2606.00341

arxiv.org

ROGUE: Misaligned Agent Behavior Arising from Ordinary Computer Use

As AI agents are increasingly deployed in real personal and corporate settings (email accounts, development workflows, company databases, etc.), safety considerations surrounding these agents become p...

020

Aran Nayebi @anayebi.bsky.social · 04/07/2026

2. Corrigibility - building AI systems that stay editable, deferential, and willing to be shut down - is feasible with the lexicographic approach that sets hierarchical priorities on an agent’s objectives. Paper: arxiv.org/abs/2507.20964

arxiv.org

Core Safety Values for Provably Corrigible Agents

We introduce the first complete formal solution to corrigibility in the off-switch game, with provable guarantees in multi-step, partially observed environments. Our framework consists of five *struct...

110

Aran Nayebi @anayebi.bsky.social · 04/07/2026

In case you want to learn more about AI safety this 4th, check out the recent recording of some of my group's work on the AI Safety Research directory! www.youtube.com/watch?v=XWQJ...

youtube.com

Corrigibility and the ROGUE Benchmark with Prof. Nayebi, NeuroAgents Lab

YouTube video by AI Safety Research Directory

131

Aran Nayebi @anayebi.bsky.social · 30/06/2026

UAI 2026 Camera ready up on arXiv as v3! I've written a long-form LW blogpost on how the technical aspects of this work connect to NeuroAI and to AI sentience/welfare, entitled: "What Capable Agents Must Know: Why AI Consciousness May Be an Inevitable Byproduct of Capability"

lesswrong.com

What Capable Agents Must Know: Why AI Consciousness May Be an Inevitable Byproduct of Capability — LessWrong

[No LLMs were used (or harmed!) in the writing of this blogpost!] Technical results can all be found here: https://arxiv.org/abs/2603.02491 …

131

Aran Nayebi @anayebi.bsky.social · 23/06/2026

and at 1:00:53 how Nash encouraged him to switch from knot theory to game theory. My own talk on AI safety and agreement-based complexity starts at 2:01. The Aumann panel discussion starts at 34:49. Turn captions on for best experience. Link + full timestamps here: www.youtube.com/watch?v=WD_O...

youtube.com

Nobel Laureate Robert Aumann 96th Birthday Conference: 50 Years of "Agreeing to Disagree" Paris 2026

YouTube video by Aran Nayebi

010