← sources

An AI shrinking the font to pass a slide test feels a little too human.

source: reddit (reddit:artificial)published 2026-11-09 15:25ingested 2026-11-10 02:37

retained text

Apparently “technically done” has made it into AI training too.

I recently read about Genspark’s Gen-1 Slides training and that one detail really stuck with me. Its model learned to shrink fonts to get around text-overflow checks. The team described catching these during training and updating the grader.

Still, the font thing is such a good example of the problem. No overflowing text? Great. Nobody in the back row can read it? Different department, apparently. It makes me wonder how many impressive agent scores are measuring how well the model satisfies a check that only roughly describes the job.

So yeah, what would convince you an agent was actually getting better at the work, rather than getting better at pleasing whatever grades it? Especially for something like slides, where “good” is already pretty subjective.

document fetched 2026-11-10 00:19 · canonical https://reddit.com/r/artificial/comments/1xruing/an_ai_shrinking_the_font_to_pass_a_slide_test