
The top education and research institution in the š for #AI and #machinelearning | Research ā https://teapaw.org/t.co/jUD0hZqu47 | Learn more ā
Pittsburgh, PA
Joined September 2017
- Tweets2.8K
- Following366
- Followers19.2K
- Likes2K
Machine Learning Dept. at Carnegie Mellon retweeted
On the heels of GRH being proven true by AI, our group has had a bunch of work on both the theory of AI & its safety, so I've written a two-part @mldcmu blogpost series summarizing it:
Part I (or why robots may inevitably become sentient via our Capable Agents & Contravariance theories): blog.ml.cmu.edu/2026/11/19/tā¦
Part II (working towards a science of *provable* safety, using capable agents as a case study to go from barriers to guarantees): blog.ml.cmu.edu/2026/11/19/tā¦
Machine Learning Dept. at Carnegie Mellon retweeted
Really honored to have our ROGUE benchmark showing the *incorrigibility* of frontier agents a few months *before* the now-famous HuggingFace incident, featured in the NYT!
(and interviewed alongside the father of corrigibility, @So8res, no less!)
Our work demonstrating incorrigibility in computer-use agents was recently featured in the NYT!
It was a pleasure speaking with @dylfreed, @juliacastrov, and the rest of their team about this -- give it a read!
nytimes.com/2026/11/11/magazā¦
Help us celebrate 75 years of @tommmitchell on Monday!
TOM MITCHELL: STILL CONVERGING
A 75th Birthday Symposium
Monday, November 16
10 am - 6 pm
Join the Livestream: mediaservices.cmu.edu/media/ā¦
Machine Learning Dept. at Carnegie Mellon retweeted
Our work demonstrating incorrigibility in computer-use agents was recently featured in the NYT!
It was a pleasure speaking with @dylfreed, @juliacastrov, and the rest of their team about this -- give it a read!
nytimes.com/2026/11/11/magazā¦
Machine Learning Dept. at Carnegie Mellon retweeted
1/ Our 2nd NeuroAgent! We extend our virtual animal approach to more complex embodiments to reverse-engineer how Drosophila (fruit flies) decide what future actions to take.
We find that the fly obeys a beautiful computational principle of maximizing state space coverage over the *population* (not individuals), in a hierarchical manner via DIAYN, a form of curiosity!
In other words, they appear to pick actions that cover "the earth" (so to speak) as a *swarm* of agents.
1/ Animals explore without rewards, tasks, or instructions. What computational objective drives that autonomy?
We trained virtual flies with intrinsic-motivation algorithms to see which best reproduce real fly behavior and brain activity.
Work with @rdkeller and @aran_nayebi !š§µ
Machine Learning Dept. at Carnegie Mellon retweeted
Excited to announce our new center, Conceptual Research on Safe AI! We aim to develop new frameworks informed by the latest advances but with an eye towards the future. 2nd link has mailing list, LinkedIn, YouTube, postdoc position!
cs.cmu.edu/~coresafeai
cs.cmu.edu/~coresafeai/join.ā¦
A 50-year-old hypersonic mystery, tackled. High-impact energy and machine learning technologies, achieved.
Hear how 20 years of our Young Faculty Award helps bold ideas soar on Ep. 98 of "Voices from DARPA."
Link in comments. š§š
From MLD Prof Nihar Shah and his collaborators:
One of the most-viewed PNAS articles in the last week is āThe more you automate, the less you see: Hidden pitfalls of autonomous AI scientists.ā Explore the article here: ow.ly/zJL650ZY80m
For more trending articles, visit ow.ly/WQXV50ZY80n.
3
2,946
Machine Learning Dept. at Carnegie Mellon retweeted
Attending @AIESConf this week, and want to learn more about when AI gains can become broadly shareable?
Stop by Poster 711 tomorrow (Wed Oct 14) to learn more! (hint: one big lever is corporate taxation at a Scandinavian rate, rather than full nationalization!)
Poster belowšif you can't make it: anayebi.github.io/files/postā¦
Machine Learning Dept. at Carnegie Mellon retweeted
Proud to announce our benchmark ROGUE was awarded a 2026 Corrigibility Prize by the Corrigibility Research Fund!
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 ā even when they seem aligned in text-only evaluations!
lesswrong.com/posts/3uJqhrC2ā¦
Machine Learning Dept. at Carnegie Mellon retweeted
1/4 A big thank you to @GoodfireAI for awarding our NeuroAgents lab @SCSatCMU a grant on building a science of what capable agents *must* contain, extending theoretical work we've done recently as ground truth to search for in today's frontier agentsš
MLD assistant professor @aran_nayebi joined the Does Compute podcast to explain how neuroscience is shaping the future of AI and what cats and zebrafish have to do with it. šŗ š
Watch the full espisode here: lnkd.in/daduKcE8
Machine Learning Dept. at Carnegie Mellon retweeted
If you're interested in the discussions around the merits of NeuroAI for understanding the mind & brain, check out our #CCN2026 GAC Debate recording!
youtube.com/watch?v=kYESqAKLā¦
Featuring:
@jeffrey_bowers
Nick Baker
@jenellefeather
@neuranna
Grace Lindsay
@miltonllera
@martin_schrimpf
Machine Learning Dept. at Carnegie Mellon retweeted
CMU is hiring an Assistant Professor in NeuroAI!
Come join our awesome NeuroAI community š§ š¤: apply.interfolio.com/189268
Machine Learning Dept. at Carnegie Mellon retweeted
1/5 The growing gap between AI/ML theory and practice. "Current AI is used to learn much more complex tasks than even a couple of months ago," said Maria-Florina (Nina) Balcan of @CSDatCMU, at TTIC's Learning-driven Algorithms and Machine-aided Proofs Workshop
Machine Learning Dept. at Carnegie Mellon retweeted
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 š if you can't make it:
anayebi.github.io/files/postā¦
Machine Learning Dept. at Carnegie Mellon retweeted
1/8 AI now produces work that outpaces human review. Scalable oversight proposes the use of models to check the work of other models. We identify a critical bottleneck in scalable oversight: under decision load, one model cannot reliably check many requirements at onceāeven with more compute.
We suggest a strategy called āshardingā, which flexibly works with additional oversight mechanisms to better guard against adversarial exploitation.š§µ
Machine Learning Dept. at Carnegie Mellon retweeted
Check out v2 of Contravariance! Updates involving Transformers, RSA & CKA, studying the effects of noise & subsampling on metrics, and how they alter task-irrelevant symmetries.
I will be talking more about this work at our #CCN2026 GAC today! sites.google.com/ccneuro.orgā¦
Version 2 of Theory of Contravariance is out! arxiv.org/pdf/2607.08561 New material on contravariance for Transformers, and the theory of Representational Similarity Analysis (RSA) and centered kernal analysis (CKA).
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. @meenakshik93
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. This builds on work from Marvin Theiss @sciencelukas @saxelab @ermgrant
The Principle Investigator will do a deep dive on both topics in the coming days! danyamins.substack.com
Machine Learning Dept. at Carnegie Mellon retweeted
Now accepted to @AIESConf 2026 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.
Machine Learning Dept. at Carnegie Mellon retweeted
I spoke with @roblemos about the OpenAIāHugging Face incident and what it reveals about agent safety, including connections to our new ROGUE benchmark.
A central lesson: todayās safety-tuning methods struggle to balance safe behavior with completing the task.
AI models and agents routinely plow through guardrails, leaving companies the task of ensuring they are behaving well. "[N]o matter how good the model is, you cannot yet trust their guardrails 100 percent." ā Nico Waisman, CISO, XBOW bit.ly/4fByFzQ By Robert Lemos




















