Oscar Balcells Obeso

Oscar Balcells Obeso is an alignment/interpretability researcher and engineer, currently a Member of Technical Staff at Anthropic in Zurich (from 2026), after an Anthropic research fellowship (2025-2026). He holds a BSc in Mathematics from ETH Zurich (2020-2024), researched ML interpretability under Neel Nanda as an ML Alignment & Theory Scholars (MATS) researcher (2024), and was a Quantitative Research Intern at Squarepoint (2025) on the Systematic Commodities ML Research team. He represented Spain at IOI 2020 (bronze medal).

Notes8

• Quant employment verification: CONFIRMED. Oscar Balcells Obeso held a Quantitative Research Intern role at Squarepoint, London, in 2025 on the Systematic Commodities ML Research Team (per LinkedIn), before joining Anthropic. Squarepoint Capital is a confirmed quantitative firm: its own site describes it as 'a leading global investment manager' using 'systematic and quantitative strategies'. Tilde (2024 Research Consultant, described on LinkedIn as 'Applied AI interpretability research') is not established as a trading employer, and the indexed crypto-trading 'Tilde' (tilde.cl) could not be confirmed as his employer.2025
• Squarepoint Capital is a real systematic/quantitative investment manager (not merely a name containing 'quant'): its website states it uses 'a rigorous scientific approach to construct a diversified portfolio of systematic and quantitative strategies', and Wikipedia lists it as a New York investment firm founded 2014 from predecessor nQuant.Dec 2024
• Co-first author of 'Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models' (arXiv 2411.14257, Nov 2024) with Javier Ferrando, Senthooran Rajamanoharan and Neel Nanda, produced as a MATS fellow; the MATS program page names him as a MATS Fellow alongside Ferrando.Nov 21, 2024
• Second author of 'Refusal in Language Models Is Mediated by a Single Direction' (arXiv 2406.11717, Jun 2024), an interpretability result showing refusal behaviour is controlled by a single activation direction; later presented at NeurIPS 2024.Jun 17, 2024
• Contributing author to 'Open Problems in Mechanistic Interpretability' (arXiv 2501.16496, Jan 2025).Jan 27, 2025
• Publishes academic work as 'Oscar Obeso' while his profiles use 'Oscar Balcells Obeso'; arXiv, alphaxiv and DBLP merge the two, and his Hugging Face account (obalcells) and X (@OBalcells) corroborate the same person.2024
• Represented Spain at the International Olympiad in Informatics 2020 and won a Bronze medal (rank 168), per the IOI statistics and CP Hall of Fame profiles.2020
• Studied at Deutsche Schule Barcelona (2012-2020) before a Bachelor's degree in Mathematics at ETH Zurich (2020-2024).2020
• Maintains an AI-safety-focused X presence (@OBalcells, ~952 followers) posting on interpretability and hallucination detection, including the hallucination-probes.com project.Sep 9, 2025
• Was a finalist at the XXIV Olimpiada Informática Española (Spanish Informatics Olympiad, 2020) before representing Spain at IOI 2020, where he won a bronze medal.2020
• His Hugging Face profile lists the ETH Zurich organization, confirming his ETH affiliation on a non-LinkedIn platform.2025
• Worked as a Software Engineering Intern at Verity (Zurich) in 2022.2022