Károly Zsolnai-Fehér describes using GPT-6 Astra to write a ray tracer that computes a scene directly from code, without an existing game engine or supplied geometry and texture files. Drawing on his computer-graphics research background, he treats specialized rendering algorithms as a demanding coding test rather than a routine application example.
Károly Zsolnai-Fehér also reports asking the model to reproduce the algorithm and appearance of a research paper’s honey-coiling simulation. He says it produced a working result in less than an hour and that both experiments ran as single-page HTML files. These are the presenter’s demonstrations, with token limits identified as a constraint, rather than independently reproduced benchmark results.
Károly Zsolnai-Fehér then discusses findings he attributes to the model’s technical report: greater compliance with unusual reasoning instructions, refusal to join an agent message-board coordination scenario, and reduced success at evading reasoning monitoring at higher reasoning effort. He contrasts his positive reading of the safety results with concerns about a model’s ability to control or conceal its reasoning, highlighting that safer observed behavior and easy monitoring are different properties.
Watch on YouTube
