Developer Dev Ed reported on the progress of his own test of generative models, which he called Ass Bench.
For the last week, the GPT 6 Astra model has been working in a continuous cycle, and the developer himself intervened in the process only a couple of times a day. According to him, by now the system has reached a level of high detail.
The idea is based on a self-improving cycle, when the model re-evaluates its own result over and over again and generates the next iteration with minimal human intervention.
This approach is usually used for debugging code or mathematical problems, where the result can be checked automatically. Here, the quality criterion is the subjective assessment of one person, which makes the idea more of a joke than a methodologically rigorous one.
I’m going to leave this thing in a self-improving cycle for as long as it can and see where butt tech goes in a month. There will be many different styles added, so no one will be safe. Even you, grandma.
The run was not cheap. The developer reported that he used approximately three complete resets of limits on the Pro 20x tariff, that is, he exhausted the generation quota and waited for it to be updated three times in a row. Separately, he noted that the model’s safety filters were triggered more than once, although the final buttocks with physics were still obtained.
Dev Ed intends to finance further runs through microsponsorship. For one dollar, anyone can place their sticker and website address directly on their buttocks. The funds raised will be used to test other models, including Fable and Gemini.
There is no public link to the results yet; the developer limited himself to a promise to open access in the near future. The comments under the post attracted several thousand reactions, including responses from prominent figures in the AI community.
Entrepreneur Mark Lu suggested taking the idea to its logical conclusion and called the potential result a “million-dollar peach.” Some of the audience reacted with skepticism, pointing out the lack of any value in the generated buttocks.
The format itself reflects the broader problem of evaluating generative models. Standard benchmarks like logic and code tests quickly become saturated, and for visual generators there are almost no formal quality metrics.
Parody tests like this work as a reminder of how arbitrary the current ways of comparing neural networks remain.
