Researchers find Astra can decode images from weights

Researchers investigating the AI model Astra found that it can perform non-trivial visual reasoning, image decoding, and data reconstruction directly within its forward passes without using external tools. Experiments demonstrated that Astra could identify a playing card from raw ASCII byte data without dimensions, manually decode JPEGs through scratchpad arithmetic, and accurately recover visual maze structures directly from neural network weights. The authors argue these capabilities raise concerns regarding the monitorability of opaque computations performed by frontier AI models. Open article →