Aran Nayebi

Feel free to check out my new LessWrong post for a high-level summary of this work! www.lesswrong.com/posts/dP8J6v...

Matt's slides on Interactive World Models: www.cs.cmu.edu/~mgormley/co...

My slides on the Science of AI Alignment: www.cs.cmu.edu/~mgormley/co...

...and that's a wrap for Fall 2025! In the final lecture of the semester, Matt Gormley & I covered bleeding-edge research topics in Generative AI, namely Interactive World Models + Science of AI Alignment.

Next semester we plan to have our recordings publicly available on YouTube -- stay tuned!

The 2nd paper circumvents the first paper's main "no free lunch" barrier of encoding "all human values", by identifying small value sets that yield the *first* formal guarantees on corrigibility.

In the AAAI Machine Ethics Workshop (W37) Proceedings 👇:
bsky.app/profile/anay...

We have 2 papers accepted to #AAAI2026 this year!

The first paper 👇 on intrinsic barriers to alignment (establishing no free lunch theorems of encoding "all human values" & the inevitability of reward hacking) will appear as an *oral* presentation at the Special Track on AI Alignment.

Slides: www.cs.cmu.edu/~mgormley/co...
Full course info: bsky.app/profile/anay...

In today's Generative AI lecture, we cover code generation & autonomous agents, discussing how Github Co-Pilot works, diving into multimodal agents (like Gemini 3 Pro!), and ending on AI scientists & AI for science. Lots more to explore in this rapidly growing space!

Join us December 5th at University of Toronto (in-person and online) for a special seminar by Dr. Aran Nayebi on reverse-engineering the brain and building neuroscience-inspired artificial intelligence.

#neuroAI #compneuro @anayebi.bsky.social @utoronto.ca @uoftcompsci.bsky.social

Slides: www.cs.cmu.edu/~mgormley/co...
Full course info: bsky.app/profile/anay...

In today's Generative AI lecture, we dive into reasoning models by dissecting how DeepSeek-R1 works (GRPO vs. PPO, which removes the need for a separate value network + training with a simpler rule-based reward), and end on mechanistic interpretability to better understand those reasoning traces.

Finally, we briefly discuss Querying Transformers for text-image alignment, as a hold-over from last lecture on multimodal foundation models!

We also discuss data quality & amount (where you get great performance with a smaller model trained on lots of tokens), how to get good data depending on your application, and Moravec's paradox for robotics foundation models.

In today's Generative AI lecture, we primarily discuss scaling laws and the key factors that go into building large-scale foundation models.

Slides: www.cs.cmu.edu/~mgormley/co...

Full course info: bsky.app/profile/anay...

Autonomous Behavior and Whole-Brain Dynamics Emerge in Embodied Zebrafish Agents with Model-based Intrinsic Motivation Autonomy is a hallmark of animal intelligence, enabling adaptive and intelligent behavior in complex environments without relying on external reward or task structure. Existing reinforcement learning ...

Full paper (to appear in NeurIPS 2025!) here: arxiv.org/abs/2506.00138

Congratulations to my Ph.D. student Reece Keller for winning the best talk award at #CRSy25 on our project building the first task-optimized autonomous agent that predicts whole-brain data! Check out the post below for other cool talks!!

Detailed summary: bsky.app/profile/reec...

Congrats to this year's Nobel Prize winners!
Philippe's seminal work is in fact what our recent closed form UBI AI capability threshold builds on: bsky.app/profile/anay...

Thanks @undo-hubris.bsky.social for the invite & for hosting!

Slides: anayebi.github.io/files/slides...

Paper 1 (alignment barriers): arxiv.org/abs/2502.05934
Paper 1 summary: bsky.app/profile/anay...

Paper 2 (corrigibility): arxiv.org/abs/2507.20964
Paper 2 summary: bsky.app/profile/anay...

YouTube video by ILIAD Conference Day 5 Aran Formal Guarantees of Corrigibility

My ILIAD ’25 talk, “Intrinsic Barriers & Pathways to Alignment”: why “aligning to all human values” provably can’t work, why reward hacking is inevitable in large state spaces, & how small value sets bypass “no free lunch” limits to yield formal corrigibility.

www.youtube.com/watch?v=Oajq...

A nice application of our NeuroAI Turing Test! Check out
@ithobani.bsky.social's thread for more details on comparing brains to machines!

Academic paper: bsky.app/profile/anay...

AI Is taking jobs: Could universal basic income become a reality? Forecasts that AI could erase tens of millions of jobs by the end of the decade appear to be making the notion of a guaranteed income less radical.

Honored to be quoted in this @newsweek.com article discussing how AI could accelerate the need for UBI.

Read more here: www.newsweek.com/ai-taking-jo...

Next time we discuss how to optimize these reward models via DPO/policy gradients!

Slides: www.cs.cmu.edu/~mgormley/co...

Full course info: bsky.app/profile/anay...

Specifically, we cover methods which don't involve parameter-updating, e.g. In-Context Learning / Prompt-Engineering / Chain-of-Thought Prompting, to methods that do, such as Instruction Fine-Tuning & building on IFT to perform full-fledged Reinforcement Learning from Human Feedback (RLHF).

In today's Generative AI lecture, we talk about all the different ways to take a giant auto-complete engine like an LLM and turn it into a useful chat assistant.

Slides: www.cs.cmu.edu/~mgormley/co...

Full course info: bsky.app/profile/anay...

In today's Generative AI lecture, we discuss the 4 primary approaches to Parameter-Efficient Fine-Tuning (PEFT): subset, adapters, Prefix/Prompt Tuning, and Low-Rank Adaptation (LoRA).

We show each of these amounts to finetuning a different aspect of the Transformer.

YouTube video by CMU Robotics Institute RI Seminar: Aran Nayebi : Using Embodied Agents to Reverse-Engineer Natural Intelligence

6/6 I close with reflections on AI safety and alignment, and the Q&A explores open questions: from building physically accurate (not just photorealistic) world models to the role of autoregression and scale.

🎥Watch here: www.youtube.com/watch?v=5deM...

Slides: anayebi.github.io/files/slides...

5/6 I also touch on the Contravariance Principle/Platonic Representation Hypothesis, our proposed NeuroAI Turing Test, and why embodied agents are essential for building not just more capable, but also more reliable, autonomous systems.

4/6 This journey culminates in our first task-optimized “NeuroAgent”, integrating advances in visual and tactile perception (including our NeurIPS ’25 oral), mental simulation, memory, and intrinsic curiosity.

3/6 By grounding agents in perception, prediction, planning, memory, and intrinsic motivation — and validating them against large-scale neural data from rodents, primates, and zebrafish — we show how neuroscience and machine learning can form a unified *science of intelligence*.