The News, Explained
OpenAI launched its science blog on November 15, the date on the announcement. It plans to cover research results, how scientists use AI, ways of checking their work, and where methods fall short. It will also publish researchers’ perspectives. The first guest post describes how a physicist worked with GPT-6 Astra on a mathematical formulation of the Standard Model. Source Source
The Standard Model explains elementary particles and three fundamental forces. Nathaniel Craig, a physicist and visiting scientist at OpenAI, worked on a candidate way to define the theory on a lattice: a grid of points representing space and time. A grid provides a way to calculate the theory on a computer. This proposal comes with conditions on the grid spacing and the strength of the fields. Craig said it would take more work to make the approach useful for realistic calculations. Source Source
Craig’s strategy was to represent the same physical situation on increasingly fine grids and compare those representations. He says Astra proposed a way to transfer the situation onto finer grids and took the lead in developing a detailed proof that the comparisons approach a definite answer. Outside physicists looked over the proof and asked questions. This account of who did what comes from Craig’s guest post. Source
The paper had already been submitted to arXiv on October 6. An arXiv preprint is a manuscript that researchers make public so others can read and scrutinize it. The new blog post explains the research strategy and the AI’s part in reaching the result. Both the paper and the post call it a candidate, and Craig says experts still need to discuss whether it has all the required properties. Source Source
OYOPICK’s Take
An account of AI research is more useful when it shows where the tool entered the work. Here, readers can separate choosing the question and approach, developing the detailed proof, and having other researchers read and question that proof. Someone evaluating tools for a project can then compare examples with the stage where their own work is stuck.
If the blog keeps reporting the assumptions behind a result and what remains to be checked, readers will have a clearer picture of the AI’s contribution. We hope future posts follow this one in linking directly to the manuscript and making the stage of verification clear.