AI Reddit — 2026-07-22#

The Buzz#

The most fascinating discovery today wasn’t a benchmark, but a weird quirk of instruction-following: translation models are suffering from indirect prompt injection where they solve the payload instead of translating it. A deep dive showed that when you feed a model like Gemma-3 or 4 27B/31B a math proof or code snippet to translate, the instruction boundary fails and the model starts executing the reasoning traces within the text. It’s a stark reminder that as models get smarter, treating text strictly as data for “boring” tasks requires rigorous parsing and chunking to prevent runaway generation.