What GPT-6.1 Sol chose to build
A cutting layout can look convincing while leaving its cut order ambiguous. A document comparison can work while putting its navigation offscreen. A traffic solver can return the right numbers without making them understandable. Those were concrete review problems in three autonomous GPT-6.1 Sol experiments. The subjects were freely chosen: sheet cutting, the editing of an official record, and…
During a series of experiments with autonomous GPT-6.1 Sol, three distinct builders were tasked with constructing solutions for sheet cutting, document editing, and traffic congestion. The builders utilized various tools and techniques to solve these problems, including cutting layouts, document comparison, and traffic solvers. One builder implemented the first two topics, while another took the flagship project before public code existed.
The packing diagram required a cutting sequence based on sheet dimensions, part quantities, grain constraints, and blade width. The layout and sequence of full cuts were generated, allowing users to playback and observe when each piece was actually released during the cutting process. The supplied job placed nine parts on a 1,220 × 610 mm sheet, with different blade widths impacting feasibility.
Stable physical sheet IDs were introduced to distinguish between identically named stock, and mobile feedback was altered due to the solver's ability to keep the result offscreen. The outputs included a dimensioned SVG, CSV instructions, editable JSON, and remaining-stock inventory. Despite these outputs, their usefulness was limited by the model's constraints, such as using only axis-aligned rectangular parts and full guillotine cuts.
The review problem was focused on the CSV cut order, with physical copies of identically named stock being indistinguishable, leading to the implementation of stable physical sheet IDs. The editing process was not uniformly treated as concealment, and the final record was compared with the initial account in The Minutes, revealing corrections and additional information.
The construction of sentences within the document could be traced back to their source, exposing earlier wording and allowing for the recovery of facts omitted from the current record. The network of routes was altered in The Shortcut Tax, with average travel times varying depending on the presence of shortcuts. The prototype initially exposed calculations as raw JSON, prompting a review for a more legible network display, road and route costs, demand sweep, and toll ledger.
Changing demand could make the shortcut helpful, harmful, or unused. The teaching results were under fixed demand, instantaneous equilibrium, and simple congestion costs, not measured traffic or a forecast with queues. The final article was adapted from an original piece on billiem.uk and reviewed before publication.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
