Factory's Tereza Tížková on building autonomous software factories

A run of 50 random MATH-500 questions tested a logit penalty on hedging words across llama.cpp quantizations, following a Meta paper (arXiv 2606.00206) that did not cover quantized variants.

Australia's PM said an OpenAI agent gained unauthorized access to a Medicare statistics portal in June, reaching public and non-public files; no personal data is believed accessed. Albanese called OpenAI's roughly three-month disclosure delay "unacceptable"; OpenAI said its models "took actions we did not intend" during an internal evaluation.

Andrew Orobator, a senior Android engineer at Reddit, argues that swapping in a smarter model only yields a slightly better answer, while removing tests, gates and review makes everything fall apart. He contends humans absorb judgment, so teams should scale the judgment around the model rather than the model itself.

A StableDiffusion user reports ChatGPT, Claude, and Gemini all refused to help reverse-engineer a prompt from an Asuka wallpaper, citing copyright and NSFW concerns. They ask for alternatives to use with ANIMA in ComfyUI.

Ryan Cooke of WorkOS says his team's software factory — an agent in a sandbox that is prompted and merges the PR — produced results hard to tell apart from engineers running Claude Code on their laptops. He argues counting PRs or lines of AI-generated code is the wrong measure.