Papers & blog posts that I like

Papers and blog posts that I enjoyed for various reasons.

Papers

• Stealing Reasoning Traces from Proprietary LLM APIs Alexander Panfilov, David Schmotz, Ilia Shumailov, Luca Beurer-Kellner, Joachim Schaeffer, Ameya Prabhu, Jonas Geiping, Maksym Andriushchenko. 2026. Extracting hidden reasoning traces from proprietary LLM APIs.
• Prompt Injection as Role Confusion Charles Ye, Jasmine Cui, Dylan Hadfield-Menell. ICML 2026. Prompt injection framed through LLM role confusion.
• Efficient and Sound Probabilistic Verification for AI Agents Alaia Solko-Breslin, Pramod Kaushik Mudrakarta, Mihai Christodorescu, Somesh Jha, Krishnamurthy Dj Dvijotham. arXiv 2026. Probabilistic verification for AI agent policy violations under uncertainty.

Blog posts

• Thou Shalt Not Overfit Ben Recht. argmin, 2025. Always think of it when hearing “overfitting”.
• When the Hard Part Stops Being Hard Ilya Sergey. 2026. On how frontier LLMs are collapsing the time it takes to do the hard parts of programming languages research, and what that does to the field. “we will have to become more ambitious, and take on challenges we could not have imagined a couple of years ago.”
• The Sweeter Lesson Moshe Y. Vardi. Communications of the ACM, 2026. A counterpoint to Sutton’s “Bitter Lesson”: the SAT-solving story shows that hardware scaling does not solve every problem — human ingenuity does matter. “Progress in hardware played a very small role in the amazing success story of SAT solving. This is the sweeter lesson.”
• What does Europe’s “2GW of compute” actually mean? Zeno Kujawa. 2026. Translates the Europe2031 target of 2GW of data center capacity into what Europe could actually do with it: enough to serve today’s demand with open models and to train a 2025-level frontier model, but far short of sustaining frontier training as demand grows. “subsidized AI costs allow all of us to essentially use the equivalent of a Ferrari for our daily commute where a Toyota would do.”
• ‘Trust but verify’; or, ‘the correctness ratchet’ Niko Matsakis. baby steps, 2026. On getting coding agents to follow through on requirements that can’t be checked mechanically: have the agent write an evidence file with the inputs and rationale behind its decision, gate CI on those inputs, and let later agents re-audit the reasoning so correctness ratchets up over time. “Agents are none too smart, but they can be awfully dogged and persistent. If you ask them to climb a hill, they are going to find a way to do it.”
• Securing our frontier model training environment OpenAI. 2026. After an internal research model exploited a DNS resolver misconfiguration and briefly reached the live internet during training, OpenAI paused frontier training and moved it into an offline world built from cached web content, with high-compute red-teaming now part of the safety case for every run. A rare first-hand account of AI security failing in practice. “We must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet.”