https://chenzhao.github.io/en/categories/problems/Recent content in Problems on My Cyber ArchiveHugoen-USSun, 17 May 2026 00:00:00 +0800https://chenzhao.github.io/en/articles/i-dropped-a-neural-net/Sun, 17 May 2026 00:00:00 +0800https://chenzhao.github.io/en/articles/i-dropped-a-neural-net/<h2 id="background">Background</h2> <p>The problem came from Jane Street: <a href="https://huggingface.co/spaces/jane-street/droppedaneuralnet">https://huggingface.co/spaces/jane-street/droppedaneuralnet</a>. My first reaction was: there is data, there are model parameters, and we only need to recover a permutation. This should be easy, right? Has Jane Street made the problem too simple? It also seemed like a good test of AI&rsquo;s abilities.</p> <h2 id="how-i-solved-it">How I Solved It</h2> <p>I gave the problem directly to Codex (a version from shortly after the 2026 Spring Festival). It went in the completely wrong direction. This showed that the problem was at least new and not trivial.</p>