The Misdirection of Frontier Models: Talk Slides for NYU-KAIST Workshop on Frontier AI Governance

This repository contains the slides for my talk at the NYU-KAIST Summit on Building Governance Infrastructure for Frontier AI, on 2/6/2026.

Abstract

𝘈𝘐 𝘨𝘰𝘷𝘦𝘳𝘯𝘢𝘯𝘤𝘦 𝘤𝘰𝘯𝘷𝘦𝘳𝘴𝘢𝘵𝘪𝘰𝘯𝘴 𝘵𝘦𝘯𝘥 𝘵𝘰 𝘧𝘰𝘤𝘶𝘴 𝘰𝘯 "𝘧𝘳𝘰𝘯𝘵𝘪𝘦𝘳" 𝘮𝘰𝘥𝘦𝘭𝘴—𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘢𝘴𝘴𝘶𝘮𝘦𝘥 𝘵𝘰 𝘣𝘦 𝘧𝘢𝘳 𝘢𝘩𝘦𝘢𝘥 𝘪𝘯 𝘱𝘦𝘳𝘧𝘰𝘳𝘮𝘢𝘯𝘤𝘦 𝘢𝘯𝘥 𝘤𝘢𝘱𝘢𝘣𝘪𝘭𝘪𝘵𝘪𝘦𝘴, 𝘪𝘮𝘱𝘭𝘪𝘤𝘪𝘵𝘭𝘺 𝘵𝘩𝘦 𝘭𝘢𝘳𝘨𝘦𝘴𝘵 𝘢𝘯𝘥 𝘮𝘰𝘴𝘵 𝘦𝘹𝘱𝘦𝘯𝘴𝘪𝘷𝘦. 𝘙𝘦𝘤𝘦𝘯𝘵 𝘥𝘦𝘷𝘦𝘭𝘰𝘱𝘮𝘦𝘯𝘵𝘴 𝘤𝘩𝘢𝘭𝘭𝘦𝘯𝘨𝘦 𝘵𝘩𝘪𝘴 𝘧𝘳𝘢𝘮𝘪𝘯𝘨: 𝘰𝘱𝘦𝘯-𝘸𝘦𝘪𝘨𝘩𝘵 𝘮𝘰𝘥𝘦𝘭𝘴 𝘵𝘳𝘢𝘪𝘯𝘦𝘥 𝘧𝘰𝘳 𝘴𝘪𝘨𝘯𝘪𝘧𝘪𝘤𝘢𝘯𝘵𝘭𝘺 𝘭𝘦𝘴𝘴 𝘯𝘰𝘸 𝘮𝘢𝘵𝘤𝘩 𝘵𝘩𝘦 𝘱𝘳𝘪𝘤𝘪𝘦𝘴𝘵 𝘧𝘭𝘢𝘨𝘴𝘩𝘪𝘱 𝘤𝘰𝘮𝘮𝘦𝘳𝘤𝘪𝘢𝘭 𝘰𝘧𝘧𝘦𝘳𝘪𝘯𝘨𝘴' 𝘱𝘦𝘳𝘧𝘰𝘳𝘮𝘢𝘯𝘤𝘦 𝘢𝘤𝘳𝘰𝘴𝘴 𝘣𝘦𝘯𝘤𝘩𝘮𝘢𝘳𝘬𝘴, 𝘢𝘯𝘥 𝘦𝘷𝘦𝘯 𝘴𝘮𝘢𝘭𝘭𝘦𝘳 𝘢𝘭𝘵𝘦𝘳𝘯𝘢𝘵𝘪𝘷𝘦𝘴 𝘢𝘱𝘱𝘳𝘰𝘢𝘤𝘩 𝘵𝘩𝘦𝘴𝘦 𝘰𝘯 𝘴𝘱𝘦𝘤𝘪𝘧𝘪𝘤 𝘢𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴. 𝘈𝘥𝘮𝘪𝘵𝘵𝘦𝘥𝘭𝘺, 𝘵𝘩𝘦𝘴𝘦 𝘯𝘶𝘮𝘣𝘦𝘳𝘴 𝘥𝘰𝘯'𝘵 𝘤𝘢𝘱𝘵𝘶𝘳𝘦 𝘦𝘷𝘦𝘳𝘺𝘵𝘩𝘪𝘯𝘨: 𝘶𝘴𝘦𝘳𝘴 𝘤𝘰𝘯𝘵𝘪𝘯𝘶𝘦 𝘤𝘩𝘰𝘰𝘴𝘪𝘯𝘨 𝘱𝘳𝘰𝘱𝘳𝘪𝘦𝘵𝘢𝘳𝘺 𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘰𝘷𝘦𝘳 𝘤𝘩𝘦𝘢𝘱𝘦𝘳 𝘰𝘱𝘵𝘪𝘰𝘯𝘴. 𝘛𝘩𝘦 𝘮𝘰𝘴𝘵 𝘭𝘪𝘬𝘦𝘭𝘺 𝘦𝘹𝘱𝘭𝘢𝘯𝘢𝘵𝘪𝘰𝘯 𝘪𝘴 𝘵𝘩𝘦 𝘵𝘳𝘪𝘭𝘭𝘪𝘰𝘯𝘴 𝘰𝘧 𝘵𝘰𝘬𝘦𝘯𝘴 𝘰𝘧 𝘶𝘴𝘦𝘳 𝘥𝘢𝘵𝘢 𝘤𝘰𝘮𝘮𝘦𝘳𝘤𝘪𝘢𝘭 𝘱𝘳𝘰𝘷𝘪𝘥𝘦𝘳𝘴 𝘢𝘮𝘢𝘴𝘴; 𝘢𝘯𝘥 𝘶𝘴𝘦 𝘵𝘰 𝘧𝘶𝘳𝘵𝘩𝘦𝘳 𝘵𝘳𝘢𝘪𝘯 𝘵𝘩𝘦𝘪𝘳 𝘮𝘰𝘥𝘦𝘭𝘴. 𝘊𝘰𝘯𝘴𝘦𝘲𝘶𝘦𝘯𝘵𝘭𝘺, 𝘢𝘥𝘥𝘳𝘦𝘴𝘴𝘪𝘯𝘨 "𝘢𝘥𝘷𝘢𝘯𝘤𝘦𝘥" 𝘈𝘐 𝘳𝘪𝘴𝘬𝘴 𝘸𝘪𝘭𝘭 𝘳𝘦𝘲𝘶𝘪𝘳𝘦 𝘧𝘰𝘤𝘶𝘴𝘪𝘯𝘨 𝘰𝘯 𝘵𝘩𝘦 𝘪𝘯𝘤𝘦𝘯𝘵𝘪𝘷𝘦𝘴 𝘢𝘯𝘥 𝘤𝘰𝘯𝘴𝘦𝘲𝘶𝘦𝘯𝘤𝘦𝘴 𝘰𝘧 𝘵𝘩𝘦𝘴𝘦 𝘥𝘢𝘵𝘢 𝘱𝘳𝘢𝘤𝘵𝘪𝘤𝘦𝘴 𝘳𝘢𝘵𝘩𝘦𝘳 𝘵𝘩𝘢𝘯 𝘫𝘶𝘴𝘵 𝘮𝘰𝘥𝘦𝘭 𝘴𝘤𝘢𝘭𝘦 𝘢𝘯𝘥 𝘦𝘷𝘢𝘭𝘶𝘢𝘵𝘪𝘰𝘯 𝘮𝘦𝘵𝘳𝘪𝘤𝘴.

Cite as:

@misc{jernite2026misdirection,
  author = {Jernite, Yacine},
  title = {The Misdirection of Frontier Models: A Technical and Ecosystem View of Advanced {AI} in 2026},
  howpublished = {Talk slides, NYU-KAIST Summit on Building Governance Infrastructure for Frontier AI},
  year = {2026},
  month = {Feb},
  doi = {10.57967/hf/7789},
  url = {https://huggingface.co/datasets/yjernite/frontier-misdirection-talk/resolve/main/Misdirection_frontier_models_2026.pdf}
}