Kirill

Very cool! Fable found my 6x great-grandfather from the 1700s (we had only traced that line to the 1870s) Stack: familysearch . org data (Russian church books scanned by Mormons in the 90s) + YOLO line segmentation + Finnish National Archives' OCR model for handwritten Cyrillic

What happens when thousands of people submit public comments urging stronger climate action from bank regulators...and most of them don't move the needle? Researchers from @ChicagoBooth used AI to open the black box of climate policy rulemaking—the findings are striking.

Chicago Booth's Rimmy Tomy and predoctoral researcher Kirill Skobelev used large language models to study how public comments shape climate policy at national banks. What they found: Most comments did not move the needle. @CAAI_Booth ms.spr.ly/6017vU0VT

@AnthropicAI #Fable benchmarked on real data from the operating room. It's not what you think. We ran the latest model through the canonical datasets in our field and found unexpectedly poor performance across standard #surgicaldatascience tasks. What does this mean: commercial-off-the-shelf LLM/VLM don't win the day in medicine. Domain expertise and custom small models may remain advantageous, especially in real world workflows that have to be brought to the edge. But counterintuitively, this may remain true even for cloud.

RT @XYHan_: For Surgical AI, Claude Mythos/Fable isn't all that scary. Loses to a YOLO. See updated benchmarks from @kskblv 🏥🤖 Real alpha…

We benchmarked Fable on surgical tool detection so you don’t have to. At almost 2 cents/image, Claude Fable 5 still loses to a compact image classifier. Will frontier VLMs ever catch up in surgical perception? Read our paper to learn more: arxiv.org/abs/2603.27341

AI can schedule appointments and transcribe doctor's notes—but it still can't reliably spot a scalpel in a surgical video. @ChicagoBooth's @XYHan_ @kskblv, @SurgicalDSC, and @ChildrensNatl found more data ≠ better #AI in the OR. chicagobooth.edu/research/cente…

🚀 Next up at the SDSC Data Science Roundtable: "What will get us to Surgical AGI?” with Dr. @XYHan_ and Kirill Skobelev XY HAN, PhD Dr. Han is an Assistant Professor of Operations Management at the University of Chicago Booth School of Business. He holds a PhD in Operations Research and Information Engineering from Cornell University, an M.S. in Statistics from Stanford University, and a B.S.E. in Operations Research and Financial Engineering from Princeton University. His research focuses on uncovering the mathematical and operational foundations of modern artificial intelligence, using large-scale computational experiments and geometric analysis to better understand how AI systems learn and perform. Dr. Han has made several influential contributions to the field, including co-discovering the Neural Collapse phenomenon in deep neural network training and co-inventing the Survey Descent method for nonsmooth optimization. His work has been recognized with the ICLR 2022 Outstanding Paper Award and as a finalist for the ICCOPT 2022 Best Paper Prize for Young Researchers. Through his research, he continues to advance the theoretical understanding of AI while developing tools that improve real-world performance across complex domains. KIRILL SKOBELEV Kirill Skobelev is a predoctoral researcher at the University of Chicago Booth School of Business's Center for Applied AI and an incoming PhD student in Computer Science at Northwestern University. He holds a BA in Economics from the New Economic School. His research interests include AI and decision-making in long-horizon, sparse-reward domains such as medicine, business, policy, and law; interpretability methods that create a two-way supervision channel between humans and AI; and meta-science. 🗓️ Tuesday, May 5, 2026 ⏰ 11:30am–12:30pm Eastern 🔗 Register: us06web.zoom.us/meeting/regist… #SurgicalAGI #AIinHealthcare

🧐 If scaling works so well everywhere else… why doesn’t it work in surgery? That’s the question we set out to explore in research jointly supported by the Center for Applied AI at Chicago Booth and the SDSC team: arxiv.org/abs/2603.27341 🤖 Across 19 state-of-the-art models, simply applying large, general-purpose AI to surgical video didn’t work. Even after fine-tuning, performance improved, but models still struggled to generalize to new procedures, surgeons, and environments. Meanwhile, a much smaller, specialized model outperformed them all. So what’s the takeaway? 👉 Progress in medical AI may be limited as much by access to the right data as by model scale. Surgical environments are highly variable and deeply context-dependent. Without diverse, high-quality, domain-specific data, even the most powerful models fall short. This work is one of the clearest empirical signals we’ve seen on how scaling alone won’t get us to robust surgical AI. Data will. surgicalvideo.io/research #SurgicalAI #MedicalAI #AI #DataScience #HealthcareInnovation

Think of an AI PhD as an artistic pursuit

Agents increasingly make decisions in the same environments as humans. As a consequence, the presentation of choices may be optimized for machines as well as people. We introduce mecha-nudges to formalize and evaluate this phenomenon. Paper: arxiv.org/abs/2603.23433 Thread:

🧵 7/7 Feedback & contributions welcome! arXiv: arxiv.org/abs/2509.20617 GitHub: github.com/Center-for-App…

🧵 6/7 Getting started is simple: point DELM at your data + JSON schema, pick an extraction strategy, and run. Every config is saved for easy reruns and testing. More in the docs: center-for-applied-ai.github.io/delm/getting-s…

Excited to share DELM, a Python toolkit for reliable LLM-based data extraction built with Eric Fithian at @CAAI_Booth 📦pip install delm 🔗github.com/Center-for-App… 🧵 1/7