Twitxy Saved

Yatin Dandi

@DandiYatin

PhD student (EPFL, Switzerland), interested in the theory of deep learning and statistical physics of computation
yatindandi.github.io Joined May 2015
1.6K Following 365 Followers 179 Posts

Latest posts 5

Yatin Dandi @DandiYatin
Anyone looking for a super-smart and super-kind advisor in AI safety/alignment should apply asap to Maksym!
Maksym Andriushchenko Verified account @maksym_andr Aug 6, 2025
🚨 Incredibly excited to share that I'm starting my research group focusing on AI safety and alignment at the ELLIS Institute Tübingen and Max Planck Institute for Intelligent Systems in September 2025! 🚨

Hiring. I'm looking for multiple PhD students: both those able to start in Fall 2025 (i.e., as soon as possible) and through centralized programs like CLS, IMPRS, and ELLIS (the deadlines are in November) to start in Spring–Fall 2026. I'm also searching for postdocs, master's thesis students, and research interns. Fill the Google form below if you're interested!

Research group. We will focus on developing algorithmic solutions to reduce harms from advanced general-purpose AI models. We're particularly interested in alignment of autonomous LLM agents, which are becoming increasingly capable and pose a variety of emerging risks. We're also interested in rigorous AI evaluations and informing the public about the risks and capabilities of frontier AI models. Additionally, we aim to advance our understanding of how AI models generalize, which is crucial for ensuring their steerability and reducing associated risks. For more information about research topics relevant to our group, please check the following documents:
- International AI Safety Report,
- An Approach to Technical AGI Safety and Security by DeepMind,
- Open Philanthropy’s 2025 RFP for Technical AI Safety Research.

Research style. We are not necessarily interested in getting X papers accepted at NeurIPS/ICML/ICLR. We are interested in making an impact: this can be papers (and NeurIPS/ICML/ICLR are great venues), but also open-source repositories, benchmarks, blog posts, even social media posts—literally anything that can be genuinely useful for other researchers and the general public.

Broader vision. Current machine learning methods are fundamentally different from what they used to be pre-2022. The Bitter Lesson summarized and predicted this shift very well back in 2019: "general methods that leverage computation are ultimately the most effective". Taking this into account, we are only interested in studying methods that are general and scale with intelligence and compute. Everything that helps to advance their safety and alignment with societal values is relevant to us. We believe getting this—some may call it "AGI"—right is one of the most important challenges of our time.

Join us on this journey!
View quoted post on X
0 replies 0 reposts 4 likes 1.1K views View on X
Yatin Dandi @DandiYatin
Excited to share our new paper! We show that re-using batches dramatically lowers the sample complexity to O(d) for functions hard to learn under one-pass SGD.

Physics shines throughout the work through DMFT and symmetries!

arxiv.org/abs/2402.03220
1/3 Did you hear about the hardness of learning caused by information & leap exponents, non-staircase functions, or CSQ algorithms? Yes? Then brace for a surprise: All this goes away if only you reuse your batches at least twice. See: arxiv.org/pdf/2402.03220…
View quoted post on X
2 replies 0 reposts 20 likes 5.4K views View on X
Yatin Dandi @DandiYatin
Excited to share my first work in the field of mathematical physics and random graphs: arxiv.org/abs/2305.03591

We first prove phase transitions related to friendly bisections in sparse graphs and then use a universality argument to relate them to local optima in spin glasses!
The mathematical establishment of the conjectures from spin glass theory continues. This time on counting of metastable states by Bray & Moore'81 and friendly bisections in a pair of works synced to come out the same day 🤝: arxiv.org/abs/2305.03543 & arxiv.org/abs/2305.03591.
View quoted post on X